Elgato_dark/unsloth
1
0
Fork 0
mirror of https://github.com/unslothai/unsloth synced 2026-04-21 13:37:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

Studio (#4237)

Browse source 
* Rebuild Studio branch on top of main

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix security and code quality issues for Studio PR #4237

- Validate models_dir query param against allowed directory roots
  to prevent path traversal in /api/models/local endpoint
- Replace string startswith() with Path.is_relative_to() for
  frontend path traversal check in serve_frontend
- Sanitize SSE error messages to not leak exception details to
  clients (4 locations in inference.py)
- Bind port-discovery socket to 127.0.0.1 instead of all interfaces
  in llama_cpp backend
- Import datasets_root and resolve_output_dir in embedding training
  function to fix NameError and use managed output directory
- Remove stale .gitignore entries for package-lock.json and test
  directories so tests can be tracked in version control
- Add venv-reexecution logic to ui CLI command matching the studio
  command behavior

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Move models_dir path validation before try/except block

The HTTPException(403) was inside the try/except Exception handler,
so it would be caught and re-raised as a 500. Moving the validation
before the try block ensures the 403 is returned directly and also
makes the control flow clearer for static analysis (path is validated
before any filesystem operations).

* Use os.path.realpath + startswith for models_dir validation

CodeQL py/path-injection does not recognize Path.is_relative_to() as
a sanitizer. Switched to os.path.realpath + str.startswith which is
a recognized sanitizer pattern in CodeQL's taint analysis. The
startswith check uses root_str + os.sep to prevent prefix collisions
(e.g. /app/models_evil matching /app/models).

* Never pass user input to Path constructor in models_dir validation

CodeQL traces taint through Path(resolved) even after a startswith
barrier guard. Fix: the user-supplied models_dir is only used as a
string for comparison against allowed roots. The Path object passed
to _scan_models_dir comes from the trusted allowed_roots list, not
from user input. This fully breaks the taint chain.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
This commit is contained in:
Daniel Han 2026-03-12 03:36:19 -07:00 committed by GitHub
parent bced78373f
commit f08aef1804
664 changed files with 103567 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

42
.gitignore vendored
View file

@ -3,6 +3,15 @@ __pycache__/
*.py[cod]
*.class
unsloth_compiled_cache/
# ML artifacts (large files)
feature/
outputs/
exports/
/datasets/
studio/backend/assets/datasets/
unsloth_training_checkpoints/
*.gguf
*.safetensors
# C extensions
*.so
@ -136,6 +145,9 @@ venv/
ENV/
env.bak/
venv.bak/
.venv_overlay/
.venv_t5/
environment.yaml
# Spyder project settings
.spyderproject
@ -172,6 +184,34 @@ cython_debug/
.ruff_cache/
.pre-commit-cache/
# PyPI configuration file
# PyPI configuration file and IDE/Editors
.pypirc
.vscode
.idea/
.claude/
*.swp
*.swo
# oh-my-codex
.omx/
# Firebase
firebase-debug.log
# Other
resources/
tmp/
**/node_modules/
auth.db
# Local working docs
**/CLAUDE.md
**/claude.md
**/AGENT.md
**/agent.md
docs/canvas-lab-architecture.md
log_rtx.txt
log.txt
setup_leo.sh
server.pid
*.log

664
COPYING Normal file
View file

@ -0,0 +1,664 @@
GNU AFFERO GENERAL PUBLIC LICENSE
Version 3, 19 November 2007
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU Affero General Public License is a free, copyleft license for
software and other kinds of works, specifically designed to ensure
cooperation with the community in the case of network server software.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
our General Public Licenses are intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
Developers that use our General Public Licenses protect your rights
with two steps: (1) assert copyright on the software, and (2) offer
you this License which gives you legal permission to copy, distribute
and/or modify the software.
A secondary benefit of defending all users' freedom is that
improvements made in alternate versions of the program, if they
receive widespread use, become available for other developers to
incorporate. Many developers of free software are heartened and
encouraged by the resulting cooperation. However, in the case of
software used on network servers, this result may fail to come about.
The GNU General Public License permits making a modified version and
letting the public access it on a server without ever releasing its
source code to the public.
The GNU Affero General Public License is designed specifically to
ensure that, in such cases, the modified source code becomes available
to the community. It requires the operator of a network server to
provide the source code of the modified version running there to the
users of that server. Therefore, public use of a modified version, on
a publicly accessible server, gives the public access to the source
code of the modified version.
An older license, called the Affero General Public License and
published by Affero, was designed to accomplish similar goals. This is
a different license, not a version of the Affero GPL, but Affero has
released a new version of the Affero GPL which permits relicensing under
this license.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU Affero General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Remote Network Interaction; Use with the GNU General Public License.
Notwithstanding any other provision of this License, if you modify the
Program, your modified version must prominently offer all users
interacting with it remotely through a computer network (if your version
supports such interaction) an opportunity to receive the Corresponding
Source of your version by providing access to the Corresponding Source
from a network server at no charge, through some standard or customary
means of facilitating copying of software. This Corresponding Source
shall include the Corresponding Source for any work covered by version 3
of the GNU General Public License that is incorporated pursuant to the
following paragraph.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the work with which it is combined will remain governed by version
3 of the GNU General Public License.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU Affero General Public License from time to time. Such new versions
will be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU Affero General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU Affero General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU Affero General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Affero General Public License as published
by the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If your software can interact with users remotely through a computer
network, you should also make sure that it provides a way for users to
get its source. For example, if your program is a web application, its
interface could display a "Source" link that leads users to an archive
of the code. There are many ways you could offer source, and different
solutions will be better for different programs; see section 13 for the
specific requirements.
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU AGPL, see
<https://www.gnu.org/licenses/>.
Files under unsloth/*, tests/*, scripts/* are Apache 2.0 licensed.
Files under studio/*, cli/* which is optional to install are AGPLv3 licensed.

4
LICENSE
View file

@ -186,7 +186,9 @@
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [2024-] [Unsloth AI, Daniel Han-Chen & Michael Han-Chen]
Copyright [2024-] [Unsloth AI. Inc team, Daniel Han-Chen & Michael Han-Chen]
Files under unsloth/*, tests/*, scripts/* are Apache 2.0 licensed.
Files under studio/*, cli/* which is optional to install are AGPLv3 licensed.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.

20
build.sh Normal file
View file

@ -0,0 +1,20 @@
#!/usr/bin/env bash
set -euo pipefail
# 1. Build frontend (Vite outputs to dist/)
cd studio/frontend
npm install
npm run build # outputs to studio/frontend/dist/
cd ../..
# 2. Clean old artifacts
rm -rf build dist *.egg-info
# 3. Build wheel
python -m build
# 4. Optionally publish
if [ "${1:-}" = "publish" ]; then
python -m twine upload dist/*
fi

7
cli.py Normal file
View file

@ -0,0 +1,7 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
from cli import app
if __name__ == "__main__":
app()

22
cli/__init__.py Normal file
View file

@ -0,0 +1,22 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import typer
from cli.commands.train import train
from cli.commands.inference import inference
from cli.commands.export import export, list_checkpoints
from cli.commands.ui import ui
from cli.commands.studio import studio_app
app = typer.Typer(
help = "Command-line interface for Unsloth training, inference, and export.",
context_settings = {"help_option_names": ["-h", "--help"]},
)
app.command()(train)
app.command()(inference)
app.command()(export)
app.command("list-checkpoints")(list_checkpoints)
app.command()(ui)
app.add_typer(studio_app, name = "studio", help = "Unsloth Studio commands.")

2
cli/commands/__init__.py Normal file
View file

@ -0,0 +1,2 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

132
cli/commands/export.py Normal file
View file

@ -0,0 +1,132 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
from pathlib import Path
from typing import Optional
import typer
EXPORT_FORMATS = ["merged-16bit", "merged-4bit", "gguf", "lora"]
GGUF_QUANTS = ["q4_k_m", "q5_k_m", "q8_0", "f16"]
def list_checkpoints(
outputs_dir: Path = typer.Option(
Path("./outputs"), "--outputs-dir", help = "Directory that holds training runs."
),
):
"""List checkpoints detected in the outputs directory."""
from studio.backend.core.export import ExportBackend
backend = ExportBackend()
checkpoints = backend.scan_checkpoints(outputs_dir = str(outputs_dir))
if not checkpoints:
typer.echo("No checkpoints found.")
raise typer.Exit()
for model_name, ckpt_list, metadata in checkpoints:
typer.echo(f"\n{model_name}:")
for display, path, loss in ckpt_list:
loss_str = f" (loss: {loss:.4f})" if loss is not None else ""
typer.echo(f" {display}{loss_str}: {path}")
def export(
checkpoint: Path = typer.Argument(..., help = "Path to checkpoint directory."),
output_dir: Path = typer.Argument(..., help = "Directory to save exported model."),
format: str = typer.Option(
"merged-16bit",
"--format",
"-f",
help = f"Export format: {', '.join(EXPORT_FORMATS)}",
),
quantization: str = typer.Option(
"q4_k_m",
"--quantization",
"-q",
help = f"GGUF quantization method: {', '.join(GGUF_QUANTS)}",
),
push_to_hub: bool = typer.Option(
False, "--push-to-hub", help = "Push exported model to HuggingFace Hub."
),
repo_id: Optional[str] = typer.Option(
None, "--repo-id", help = "HuggingFace repo ID (username/model-name)."
),
hf_token: Optional[str] = typer.Option(
None, "--hf-token", envvar = "HF_TOKEN", help = "HuggingFace token."
),
private: bool = typer.Option(
False, "--private", help = "Make the HuggingFace repo private."
),
max_seq_length: int = typer.Option(2048, "--max-seq-length"),
load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
):
"""Export a checkpoint to various formats (merged, GGUF, LoRA adapter)."""
if format not in EXPORT_FORMATS:
typer.echo(
f"Error: Invalid format '{format}'. Choose from: {', '.join(EXPORT_FORMATS)}",
err = True,
)
raise typer.Exit(code = 2)
if push_to_hub and not repo_id:
typer.echo("Error: --repo-id required when using --push-to-hub", err = True)
raise typer.Exit(code = 2)
from studio.backend.core.export import ExportBackend
backend = ExportBackend()
typer.echo(f"Loading checkpoint: {checkpoint}")
success, message = backend.load_checkpoint(
checkpoint_path = str(checkpoint),
max_seq_length = max_seq_length,
load_in_4bit = load_in_4bit,
)
if not success:
typer.echo(f"Error: {message}", err = True)
raise typer.Exit(code = 1)
typer.echo(message)
typer.echo(f"Exporting as {format}...")
if format == "merged-16bit":
success, message = backend.export_merged_model(
save_directory = str(output_dir),
format_type = "16-bit (FP16)",
push_to_hub = push_to_hub,
repo_id = repo_id,
hf_token = hf_token,
private = private,
)
elif format == "merged-4bit":
success, message = backend.export_merged_model(
save_directory = str(output_dir),
format_type = "4-bit (FP4)",
push_to_hub = push_to_hub,
repo_id = repo_id,
hf_token = hf_token,
private = private,
)
elif format == "gguf":
success, message = backend.export_gguf(
save_directory = str(output_dir),
quantization_method = quantization.upper(),
push_to_hub = push_to_hub,
repo_id = repo_id,
hf_token = hf_token,
)
elif format == "lora":
success, message = backend.export_lora_adapter(
save_directory = str(output_dir),
push_to_hub = push_to_hub,
repo_id = repo_id,
hf_token = hf_token,
private = private,
)
if not success:
typer.echo(f"Error: {message}", err = True)
raise typer.Exit(code = 1)
typer.echo(message)

69
cli/commands/inference.py Normal file
View file

@ -0,0 +1,69 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import sys
from typing import Optional
import typer
def inference(
model: str = typer.Argument(..., help = "HF model id or local path."),
prompt: str = typer.Argument(..., help = "Prompt to send to the model."),
hf_token: Optional[str] = typer.Option(
None, "--hf-token", envvar = "HF_TOKEN", help = "Hugging Face token if needed."
),
temperature: float = typer.Option(0.7, "--temperature"),
top_p: float = typer.Option(0.9, "--top-p"),
top_k: int = typer.Option(40, "--top-k"),
max_new_tokens: int = typer.Option(256, "--max-new-tokens"),
repetition_penalty: float = typer.Option(1.1, "--repetition-penalty"),
system_prompt: str = typer.Option(
"",
"--system-prompt",
help = "Optional system prompt to prepend.",
),
max_seq_length: int = typer.Option(2048, "--max-seq-length"),
load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
):
"""Run a single inference using the specified model."""
from studio.backend.core import ModelConfig, get_inference_backend
inference_backend = get_inference_backend()
model_config = ModelConfig.from_ui_selection(
dropdown_value = model, search_value = None, hf_token = hf_token, is_lora = False
)
if not model_config:
typer.echo("Could not resolve model config", err = True)
raise typer.Exit(code = 1)
if not inference_backend.load_model(
config = model_config,
max_seq_length = max_seq_length,
load_in_4bit = load_in_4bit,
hf_token = hf_token,
):
typer.echo("Model load failed", err = True)
raise typer.Exit(code = 1)
messages = [{"role": "user", "content": prompt}]
stream = inference_backend.generate_chat_response(
messages = messages,
system_prompt = system_prompt,
temperature = temperature,
top_p = top_p,
top_k = top_k,
max_new_tokens = max_new_tokens,
repetition_penalty = repetition_penalty,
)
typer.echo("Assistant:", nl = True)
previous = ""
for chunk in stream:
delta = chunk[len(previous) :]
if delta:
sys.stdout.write(delta)
sys.stdout.flush()
previous = chunk
sys.stdout.write("\n")
sys.stdout.flush()

374
cli/commands/studio.py Normal file
View file

@ -0,0 +1,374 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import os
import platform
import subprocess
import sys
import time
from pathlib import Path
from typing import Optional
import typer
studio_app = typer.Typer(help = "Unsloth Studio commands.")
STUDIO_HOME = Path.home() / ".unsloth" / "studio"
# __file__ is cli/commands/studio.py — two parents up is the package root
# (either site-packages or the repo root for editable installs).
_PACKAGE_ROOT = Path(__file__).resolve().parent.parent.parent
def _is_repo_root(path: Path) -> bool:
"""Check if a directory looks like the repo root (actual git clone, not site-packages)."""
return (
(path / ".git").exists()
and (path / "pyproject.toml").is_file()
and (
(path / "studio" / "setup.sh").is_file()
or (path / "studio" / "setup.ps1").is_file()
)
)
def _get_repo_root() -> Optional[Path]:
"""Find the git clone repo root.
Used only by setup() — checks __file__ first (editable install),
then walks CWD parents (wheel install, user is inside the clone).
"""
# Check 1: __file__ is in the repo (editable install)
if _is_repo_root(_PACKAGE_ROOT):
return _PACKAGE_ROOT
# Check 2: CWD or any parent is the repo
cwd = Path.cwd().resolve()
for parent in (cwd, *cwd.parents):
if _is_repo_root(parent):
return parent
return None
def _studio_venv_python() -> Optional[Path]:
"""Return the studio venv Python binary, or None if not set up."""
if platform.system() == "Windows":
p = STUDIO_HOME / ".venv" / "Scripts" / "python.exe"
else:
p = STUDIO_HOME / ".venv" / "bin" / "python"
return p if p.is_file() else None
def _find_run_py() -> Optional[Path]:
"""Find studio/backend/run.py.
No CWD dependency — works from any directory.
Since studio/ is now a proper package (has __init__.py), it lives in
site-packages after pip install, right next to cli/.
"""
# 1. Relative to __file__ (site-packages or editable repo root)
run_py = _PACKAGE_ROOT / "studio" / "backend" / "run.py"
if run_py.is_file():
return run_py
# 2. Studio venv's site-packages (Linux + Windows layouts)
for pattern in (
"lib/python*/site-packages/studio/backend/run.py",
"Lib/site-packages/studio/backend/run.py",
):
for match in (STUDIO_HOME / ".venv").glob(pattern):
return match
return None
def _find_install_script() -> Optional[Path]:
"""Find studio/install_python_stack.py.
No CWD dependency — works from any directory.
"""
# 1. Relative to __file__ (site-packages or editable repo root)
s = _PACKAGE_ROOT / "studio" / "install_python_stack.py"
if s.is_file():
return s
# 2. Studio venv's site-packages
for pattern in (
"lib/python*/site-packages/studio/install_python_stack.py",
"Lib/site-packages/studio/install_python_stack.py",
):
for match in (STUDIO_HOME / ".venv").glob(pattern):
return match
return None
def _find_setup_script() -> Optional[Path]:
"""Find studio/setup.sh or studio/setup.ps1.
No CWD dependency — works from any directory.
"""
name = "setup.ps1" if platform.system() == "Windows" else "setup.sh"
# 1. Relative to __file__ (site-packages or editable repo root)
s = _PACKAGE_ROOT / "studio" / name
if s.is_file():
return s
# 2. Studio venv's site-packages
for pattern in (
f"lib/python*/site-packages/studio/{name}",
f"Lib/site-packages/studio/{name}",
):
for match in (STUDIO_HOME / ".venv").glob(pattern):
return match
return None
# ── unsloth studio (server) ──────────────────────────────────────────
@studio_app.callback(invoke_without_command = True)
def studio_default(
ctx: typer.Context,
port: int = typer.Option(8000, "--port", "-p"),
host: str = typer.Option("0.0.0.0", "--host", "-H"),
frontend: Optional[Path] = typer.Option(None, "--frontend", "-f"),
silent: bool = typer.Option(False, "--silent", "-q"),
):
"""Launch the Unsloth Studio server."""
if ctx.invoked_subcommand is not None:
return
# Always use the studio venv if it exists and we're not already in it
studio_venv_dir = STUDIO_HOME / ".venv"
in_studio_venv = sys.prefix.startswith(str(studio_venv_dir))
if not in_studio_venv:
studio_python = _studio_venv_python()
run_py = _find_run_py()
if studio_python and run_py:
if not silent:
typer.echo("Launching with studio venv...")
args = [
str(studio_python),
str(run_py),
"--host",
host,
"--port",
str(port),
]
if frontend:
args.extend(["--frontend", str(frontend)])
if silent:
args.append("--silent")
os.execvp(str(studio_python), args)
else:
typer.echo("Studio not set up. Run 'unsloth studio setup' first.")
raise typer.Exit(1)
from studio.backend.run import run_server
if not silent:
from studio.backend.run import _resolve_external_ip
display_host = _resolve_external_ip() if host == "0.0.0.0" else host
typer.echo(f"Starting Unsloth Studio on http://{display_host}:{port}")
run_server(
host = host,
port = port,
frontend_path = frontend,
silent = silent,
)
try:
while True:
time.sleep(1)
except KeyboardInterrupt:
typer.echo("\nShutting down...")
# ── unsloth studio setup ─────────────────────────────────────────────
@studio_app.command()
def setup():
"""Run one-time Studio environment setup."""
# If we're inside a git clone, use the full setup script (builds frontend, etc.)
repo = _get_repo_root()
if repo:
_dev_setup(repo)
else:
_pip_setup()
def _dev_setup(repo_root: Path):
"""Git-clone: run setup.sh / setup.ps1."""
studio_dir = repo_root / "studio"
if platform.system() == "Windows":
script = studio_dir / "setup.ps1"
subprocess.run(
["powershell", "-ExecutionPolicy", "Bypass", "-File", str(script)],
check = True,
)
else:
script = studio_dir / "setup.sh"
subprocess.run(["bash", str(script)], check = True)
def _pip_setup():
"""Pip-install: create studio venv, install all deps, build extras."""
import venv as _venv
venv_dir = STUDIO_HOME / ".venv"
venv_t5_dir = STUDIO_HOME / ".venv_t5"
if platform.system() == "Windows":
venv_python = venv_dir / "Scripts" / "python.exe"
venv_pip = venv_dir / "Scripts" / "pip.exe"
else:
venv_python = venv_dir / "bin" / "python"
venv_pip = venv_dir / "bin" / "pip"
typer.echo("Setting up Unsloth Studio...")
# 1. Create venv
if not venv_python.is_file():
typer.echo(f" Creating venv at {venv_dir}...")
STUDIO_HOME.mkdir(parents = True, exist_ok = True)
_venv.create(str(venv_dir), with_pip = True)
# 2. Install all Python deps via install_python_stack.py
install_script = _find_install_script()
if install_script:
typer.echo(" Installing Python dependencies...")
subprocess.run([str(venv_python), str(install_script)], check = True)
else:
typer.echo("Error: Could not find install_python_stack.py")
raise typer.Exit(1)
# 3. Pre-install transformers 5.x overlay
if venv_t5_dir.is_dir() and any(venv_t5_dir.iterdir()):
typer.echo(f" Transformers 5.x overlay already at {venv_t5_dir}")
else:
typer.echo(" Installing transformers 5.x overlay...")
venv_t5_dir.mkdir(parents = True, exist_ok = True)
subprocess.run(
[
str(venv_pip),
"install",
"--target",
str(venv_t5_dir),
"--no-deps",
"transformers==5.2.0",
],
check = True,
)
subprocess.run(
[
str(venv_pip),
"install",
"--target",
str(venv_t5_dir),
"--no-deps",
"huggingface_hub==1.3.0",
],
check = True,
)
typer.echo(f" Installed to {venv_t5_dir}")
# 4. Build llama.cpp
_build_llama_cpp()
typer.echo("")
typer.echo("Setup complete! Run 'unsloth studio' to start.")
def _build_llama_cpp():
"""Build llama.cpp at ~/.unsloth/llama.cpp/."""
import shutil
unsloth_home = Path.home() / ".unsloth"
llama_dir = unsloth_home / "llama.cpp"
if not shutil.which("cmake"):
typer.echo(" cmake not found — skipping llama.cpp build")
return
if not shutil.which("git"):
typer.echo(" git not found — skipping llama.cpp build")
return
typer.echo(" Building llama.cpp for GGUF inference...")
if llama_dir.exists():
shutil.rmtree(llama_dir)
unsloth_home.mkdir(parents = True, exist_ok = True)
result = subprocess.run(
[
"git",
"clone",
"--depth",
"1",
"https://github.com/ggml-org/llama.cpp.git",
str(llama_dir),
],
stdout = subprocess.PIPE,
stderr = subprocess.STDOUT,
)
if result.returncode != 0:
typer.echo(" Failed to clone llama.cpp")
return
cmake_args = []
nvcc_path = shutil.which("nvcc")
if not nvcc_path and Path("/usr/local/cuda/bin/nvcc").is_file():
nvcc_path = "/usr/local/cuda/bin/nvcc"
if nvcc_path:
typer.echo(f" Building with CUDA (nvcc: {nvcc_path})...")
cmake_args.append("-DGGML_CUDA=ON")
else:
typer.echo(" Building CPU-only...")
build_dir = llama_dir / "build"
result = subprocess.run(
["cmake", "-S", str(llama_dir), "-B", str(build_dir)] + cmake_args,
stdout = subprocess.PIPE,
stderr = subprocess.STDOUT,
)
if result.returncode != 0:
typer.echo(" cmake configure failed")
return
ncpu = str(os.cpu_count() or 4)
result = subprocess.run(
[
"cmake",
"--build",
str(build_dir),
"--config",
"Release",
"--target",
"llama-server",
f"-j{ncpu}",
],
stdout = subprocess.PIPE,
stderr = subprocess.STDOUT,
)
if result.returncode != 0:
typer.echo(" llama-server build failed")
return
subprocess.run(
[
"cmake",
"--build",
str(build_dir),
"--config",
"Release",
"--target",
"llama-quantize",
f"-j{ncpu}",
],
stdout = subprocess.PIPE,
stderr = subprocess.STDOUT,
)
server_bin = build_dir / "bin" / "llama-server"
if server_bin.is_file():
typer.echo(f" llama-server built at {server_bin}")
else:
typer.echo(" llama-server binary not found after build")

144
cli/commands/train.py Normal file
View file

@ -0,0 +1,144 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import time
from pathlib import Path
from typing import Optional
import typer
from cli.config import Config, load_config
from cli.options import add_options_from_config
@add_options_from_config(Config)
def train(
config: Optional[Path] = typer.Option(
None,
"--config",
"-c",
help = "Path to YAML/JSON config file. CLI flags override config values.",
),
hf_token: Optional[str] = typer.Option(
None, "--hf-token", envvar = "HF_TOKEN", help = "Hugging Face token if needed."
),
wandb_token: Optional[str] = typer.Option(
None, "--wandb-token", envvar = "WANDB_API_KEY", help = "Weights & Biases API key."
),
dry_run: bool = typer.Option(
False,
"--dry-run",
help = "Show resolved config and exit without training.",
),
config_overrides: dict = None,
):
"""Launch training using the existing Unsloth training backend."""
try:
cfg = load_config(config)
except FileNotFoundError as e:
typer.echo(f"Error: {e}", err = True)
raise typer.Exit(code = 2)
cfg.apply_overrides(**config_overrides)
# CLI/env tokens take precedence over config
# Handle case where typer.Option isn't resolved (decorator interaction)
from typer.models import OptionInfo
if isinstance(hf_token, OptionInfo):
hf_token = None
if isinstance(wandb_token, OptionInfo):
wandb_token = None
hf_token = hf_token or cfg.logging.hf_token
wandb_token = wandb_token or cfg.logging.wandb_token
if dry_run:
import yaml
data = cfg.model_dump()
data["training"]["output_dir"] = str(data["training"]["output_dir"])
typer.echo(yaml.dump(data, default_flow_style = False, sort_keys = False))
raise typer.Exit(code = 0)
if not cfg.model:
typer.echo("Error: provide --model or set model in --config", err = True)
raise typer.Exit(code = 2)
if not cfg.data.dataset and not cfg.data.local_dataset:
typer.echo(
"Error: provide --dataset or --local-dataset (or via --config)", err = True
)
raise typer.Exit(code = 2)
# Check if the model path is a LoRA adapter (has adapter_config.json)
model_path = Path(cfg.model) if cfg.model else None
model_is_lora = (
model_path
and model_path.is_dir()
and (model_path / "adapter_config.json").exists()
)
use_lora = cfg.training.training_type.lower() == "lora"
if model_is_lora and not use_lora:
typer.echo(
"Error: Cannot do full finetuning on a LoRA adapter. "
"Use --training-type lora or provide a base model.",
err = True,
)
raise typer.Exit(code = 2)
from studio.backend.core.training.trainer import UnslothTrainer
trainer = UnslothTrainer()
# Load model (trainer.is_vlm is set after this)
if not trainer.load_model(
model_name = cfg.model,
max_seq_length = cfg.training.max_seq_length,
load_in_4bit = cfg.training.load_in_4bit if use_lora else False,
hf_token = hf_token,
):
typer.echo("Model load failed", err = True)
raise typer.Exit(code = 1)
is_vision = trainer.is_vlm
if not trainer.prepare_model_for_training(**cfg.model_kwargs(use_lora, is_vision)):
typer.echo("Model preparation failed", err = True)
raise typer.Exit(code = 1)
result = trainer.load_and_format_dataset(
dataset_source = cfg.data.dataset or "",
format_type = cfg.data.format_type,
local_datasets = cfg.data.local_dataset,
)
if result is None:
typer.echo("Dataset load failed", err = True)
raise typer.Exit(code = 1)
ds, eval_ds = result
training_kwargs = cfg.training_kwargs()
training_kwargs["wandb_token"] = wandb_token # CLI/env takes precedence
started = trainer.start_training(
dataset = ds, eval_dataset = eval_ds, **training_kwargs
)
if not started:
typer.echo("Training failed to start", err = True)
raise typer.Exit(code = 1)
try:
while trainer.training_thread and trainer.training_thread.is_alive():
time.sleep(1)
except KeyboardInterrupt:
typer.echo("Stopping training (Ctrl+C detected)...")
trainer.stop_training()
finally:
if trainer.training_thread:
trainer.training_thread.join()
final = trainer.get_training_progress()
if getattr(final, "error", None):
typer.echo(f"Training error: {final.error}", err = True)
raise typer.Exit(code = 1)

76
cli/commands/ui.py Normal file
View file

@ -0,0 +1,76 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import os
import sys
import time
from pathlib import Path
from typing import Optional
import typer
def ui(
port: int = typer.Option(
8000, "--port", "-p", help = "Port to run the UI server on."
),
host: str = typer.Option(
"0.0.0.0", "--host", "-H", help = "Host address to bind to."
),
frontend: Optional[Path] = typer.Option(
None, "--frontend", "-f", help = "Path to frontend build directory."
),
silent: bool = typer.Option(
False, "--silent", "-q", help = "Suppress startup messages."
),
):
"""Launch the Unsloth web UI backend server (alias for 'unsloth studio')."""
from cli.commands.studio import _studio_venv_python, _find_run_py, STUDIO_HOME
# Re-execute in studio venv if available and not already inside it
studio_venv_dir = STUDIO_HOME / ".venv"
in_studio_venv = sys.prefix.startswith(str(studio_venv_dir))
if not in_studio_venv:
studio_python = _studio_venv_python()
run_py = _find_run_py()
if studio_python and run_py:
if not silent:
typer.echo("Launching with studio venv...")
args = [
str(studio_python),
str(run_py),
"--host",
host,
"--port",
str(port),
]
if frontend:
args.extend(["--frontend", str(frontend)])
if silent:
args.append("--silent")
os.execvp(str(studio_python), args)
else:
typer.echo("Studio not set up. Run 'unsloth studio setup' first.")
raise typer.Exit(1)
from studio.backend.run import run_server
if not silent:
from studio.backend.run import _resolve_external_ip
display_host = _resolve_external_ip() if host == "0.0.0.0" else host
typer.echo(f"Starting Unsloth Studio on http://{display_host}:{port}")
run_server(
host = host,
port = port,
frontend_path = frontend,
silent = silent,
)
try:
while True:
time.sleep(1)
except KeyboardInterrupt:
typer.echo("\nShutting down...")

149
cli/config.py Normal file
View file

@ -0,0 +1,149 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
from pathlib import Path
from typing import Literal, Optional, List
import yaml
from pydantic import BaseModel, Field
class DataConfig(BaseModel):
dataset: Optional[str] = None
local_dataset: Optional[List[str]] = None
format_type: Literal["auto", "alpaca", "chatml", "sharegpt"] = "auto"
class TrainingConfig(BaseModel):
training_type: Literal["lora", "full"] = "lora"
max_seq_length: int = 2048
load_in_4bit: bool = True
output_dir: Path = Path("./outputs")
num_epochs: int = 3
learning_rate: float = 2e-4
batch_size: int = 2
gradient_accumulation_steps: int = 4
warmup_steps: int = 5
max_steps: int = 0
save_steps: int = 0
weight_decay: float = 0.01
random_seed: int = 3407
packing: bool = False
train_on_completions: bool = False
gradient_checkpointing: Literal["unsloth", "true", "none"] = "unsloth"
class LoraConfig(BaseModel):
lora_r: int = 64
lora_alpha: int = 16
lora_dropout: float = 0.0
target_modules: str = "q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj"
vision_all_linear: bool = False
use_rslora: bool = False
use_loftq: bool = False
finetune_vision_layers: bool = True
finetune_language_layers: bool = True
finetune_attention_modules: bool = True
finetune_mlp_modules: bool = True
class LoggingConfig(BaseModel):
enable_wandb: bool = False
wandb_project: str = "unsloth-training"
wandb_token: Optional[str] = None
enable_tensorboard: bool = False
tensorboard_dir: str = "runs"
hf_token: Optional[str] = None
class Config(BaseModel):
model: Optional[str] = None
data: DataConfig = Field(default_factory = DataConfig)
training: TrainingConfig = Field(default_factory = TrainingConfig)
lora: LoraConfig = Field(default_factory = LoraConfig)
logging: LoggingConfig = Field(default_factory = LoggingConfig)
def apply_overrides(self, **kwargs):
"""Apply CLI overrides by matching arg names to config fields."""
for key, value in kwargs.items():
if value is None:
continue
if hasattr(self, key):
setattr(self, key, value)
else:
for section in (self.data, self.training, self.lora, self.logging):
if hasattr(section, key):
setattr(section, key, value)
break
def model_kwargs(self, use_lora: bool, is_vision: bool) -> dict:
"""Return kwargs for trainer.prepare_model_for_training()."""
# Determine target modules based on model type
if use_lora and is_vision:
# Vision models expect a string (e.g., "all-linear"); fall back to None to use trainer defaults
target_modules = "all-linear" if self.lora.vision_all_linear else None
else:
parsed = [
m.strip()
for m in str(self.lora.target_modules).split(",")
if m and m.strip()
]
target_modules = parsed or None
return {
"use_lora": use_lora,
"finetune_vision_layers": self.lora.finetune_vision_layers,
"finetune_language_layers": self.lora.finetune_language_layers,
"finetune_attention_modules": self.lora.finetune_attention_modules,
"finetune_mlp_modules": self.lora.finetune_mlp_modules,
"target_modules": target_modules,
"lora_r": self.lora.lora_r,
"lora_alpha": self.lora.lora_alpha,
"lora_dropout": self.lora.lora_dropout,
"use_gradient_checkpointing": self.training.gradient_checkpointing,
"use_rslora": self.lora.use_rslora,
"use_loftq": self.lora.use_loftq,
}
def training_kwargs(self) -> dict:
"""Return kwargs for trainer.start_training()."""
return {
"output_dir": str(self.training.output_dir),
"num_epochs": self.training.num_epochs,
"learning_rate": self.training.learning_rate,
"batch_size": self.training.batch_size,
"gradient_accumulation_steps": self.training.gradient_accumulation_steps,
"warmup_steps": self.training.warmup_steps,
"max_steps": self.training.max_steps,
"save_steps": self.training.save_steps,
"weight_decay": self.training.weight_decay,
"random_seed": self.training.random_seed,
"packing": self.training.packing,
"train_on_completions": self.training.train_on_completions,
"max_seq_length": self.training.max_seq_length,
"enable_wandb": self.logging.enable_wandb,
"wandb_project": self.logging.wandb_project,
"wandb_token": self.logging.wandb_token,
"enable_tensorboard": self.logging.enable_tensorboard,
"tensorboard_dir": self.logging.tensorboard_dir,
}
def load_config(path: Optional[Path]) -> Config:
"""Load config from YAML/JSON file, or return defaults if no path given."""
if not path:
return Config()
path = Path(path)
if not path.exists():
raise FileNotFoundError(f"Config file not found: {path}")
text = path.read_text(encoding = "utf-8")
if path.suffix.lower() in {".yaml", ".yml"}:
data = yaml.safe_load(text) or {}
else:
import json
data = json.loads(text or "{}")
return Config(**data)

153
cli/options.py Normal file
View file

@ -0,0 +1,153 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""Generate Typer CLI options from Pydantic models."""
import functools
import inspect
from pathlib import Path
from typing import Any, Callable, Optional, get_args, get_origin
import typer
from pydantic import BaseModel
def _python_name_to_cli_flag(name: str) -> str:
"""Convert python_name to --cli-flag."""
return "--" + name.replace("_", "-")
def _unwrap_optional(annotation: Any) -> Any:
"""Unwrap Optional[X] to X."""
origin = get_origin(annotation)
if origin is not None:
args = get_args(annotation)
if type(None) in args:
non_none = [a for a in args if a is not type(None)]
if non_none:
return non_none[0]
return annotation
def _is_bool_field(annotation: Any) -> bool:
"""Check if field is a boolean (including Optional[bool])."""
return _unwrap_optional(annotation) is bool
def _is_list_type(annotation: Any) -> bool:
"""Check if type is a List."""
return get_origin(annotation) is list
def _get_python_type(annotation: Any) -> type:
"""Get the Python type for annotation."""
unwrapped = _unwrap_optional(annotation)
if unwrapped in (str, int, float, bool, Path):
return unwrapped
return str
def _collect_config_fields(config_class: type[BaseModel]) -> list[tuple[str, Any]]:
"""
Collect all fields from a config class, flattening nested models. Returns list of
(name, field_info) tuples. Raises ValueError on duplicate field names.
"""
fields = []
seen_names: set[str] = set()
for name, field_info in config_class.model_fields.items():
annotation = field_info.annotation
# Skip nested models - recurse into them
if isinstance(annotation, type) and issubclass(annotation, BaseModel):
for nested_name, nested_field in annotation.model_fields.items():
if nested_name in seen_names:
raise ValueError(f"Duplicate field name '{nested_name}' in config")
seen_names.add(nested_name)
fields.append((nested_name, nested_field))
else:
if name in seen_names:
raise ValueError(f"Duplicate field name '{name}' in config")
seen_names.add(name)
fields.append((name, field_info))
return fields
def add_options_from_config(config_class: type[BaseModel]) -> Callable:
"""
Decorator that adds CLI options for all fields in a Pydantic config model.
The decorated function should declare a `config_overrides: dict = None` parameter
which will receive a dict of all CLI-provided config values.
"""
fields = _collect_config_fields(config_class)
field_names = {
name for name, field_info in fields if not _is_list_type(field_info.annotation)
}
def decorator(func: Callable) -> Callable:
sig = inspect.signature(func)
original_params = list(sig.parameters.values())
original_param_names = {p.name for p in original_params}
# Build new parameters: config fields first, then original params
new_params = []
for field_name, field_info in fields:
# Skip fields already defined in function signature (e.g., with envvar)
if field_name in original_param_names:
continue
annotation = field_info.annotation
if _is_list_type(annotation):
continue
flag_name = _python_name_to_cli_flag(field_name)
help_text = field_info.description or ""
if _is_bool_field(annotation):
default = typer.Option(
None,
f"{flag_name}/--no-{field_name.replace('_', '-')}",
help = help_text,
)
param = inspect.Parameter(
field_name,
inspect.Parameter.POSITIONAL_OR_KEYWORD,
default = default,
annotation = Optional[bool],
)
else:
py_type = _get_python_type(annotation)
default = typer.Option(None, flag_name, help = help_text)
param = inspect.Parameter(
field_name,
inspect.Parameter.POSITIONAL_OR_KEYWORD,
default = default,
annotation = Optional[py_type],
)
new_params.append(param)
# Add original params, excluding config_overrides (will be injected)
for param in original_params:
if param.name != "config_overrides":
new_params.append(param)
new_sig = sig.replace(parameters = new_params)
@functools.wraps(func)
def wrapper(*args, **kwargs):
config_overrides = {}
for key in list(kwargs.keys()):
if key in field_names:
if kwargs[key] is not None:
config_overrides[key] = kwargs[key]
# Only delete if not an explicitly declared parameter
if key not in original_param_names:
del kwargs[key]
kwargs["config_overrides"] = config_overrides
return func(*args, **kwargs)
wrapper.__signature__ = new_sig
return wrapper
return decorator

16
pyproject.toml
View file

@ -15,7 +15,7 @@ authors = [
{name = "Unsloth AI team"},
]
maintainers = [
{name = "Daniel Han", email = "danielhanchen@gmail.com"},
{name = "Daniel Han", email = "daniel@unsloth.ai"},
{name = "Michael Han", email = "info@unsloth.ai"},
]
classifiers = [
@ -24,12 +24,24 @@ classifiers = [
"Environment :: GPU :: NVIDIA CUDA",
"Topic :: Scientific/Engineering :: Artificial Intelligence",
]
dependencies = [
"typer",
"pydantic",
"pyyaml",
"nest-asyncio",
]
[project.scripts]
unsloth = "cli:app"
[tool.setuptools.dynamic]
version = {attr = "unsloth.models._utils.__version__"}
[tool.setuptools]
include-package-data = false
include-package-data = true
[tool.setuptools.package-data]
studio = ["frontend/dist/**/*"]
[tool.setuptools.packages.find]
exclude = ["images*", "tests*", "kernels/moe*"]

661
studio/LICENSE.AGPL-3.0 Normal file
View file

@ -0,0 +1,661 @@
GNU AFFERO GENERAL PUBLIC LICENSE
Version 3, 19 November 2007
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU Affero General Public License is a free, copyleft license for
software and other kinds of works, specifically designed to ensure
cooperation with the community in the case of network server software.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
our General Public Licenses are intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
Developers that use our General Public Licenses protect your rights
with two steps: (1) assert copyright on the software, and (2) offer
you this License which gives you legal permission to copy, distribute
and/or modify the software.
A secondary benefit of defending all users' freedom is that
improvements made in alternate versions of the program, if they
receive widespread use, become available for other developers to
incorporate. Many developers of free software are heartened and
encouraged by the resulting cooperation. However, in the case of
software used on network servers, this result may fail to come about.
The GNU General Public License permits making a modified version and
letting the public access it on a server without ever releasing its
source code to the public.
The GNU Affero General Public License is designed specifically to
ensure that, in such cases, the modified source code becomes available
to the community. It requires the operator of a network server to
provide the source code of the modified version running there to the
users of that server. Therefore, public use of a modified version, on
a publicly accessible server, gives the public access to the source
code of the modified version.
An older license, called the Affero General Public License and
published by Affero, was designed to accomplish similar goals. This is
a different license, not a version of the Affero GPL, but Affero has
released a new version of the Affero GPL which permits relicensing under
this license.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU Affero General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Remote Network Interaction; Use with the GNU General Public License.
Notwithstanding any other provision of this License, if you modify the
Program, your modified version must prominently offer all users
interacting with it remotely through a computer network (if your version
supports such interaction) an opportunity to receive the Corresponding
Source of your version by providing access to the Corresponding Source
from a network server at no charge, through some standard or customary
means of facilitating copying of software. This Corresponding Source
shall include the Corresponding Source for any work covered by version 3
of the GNU General Public License that is incorporated pursuant to the
following paragraph.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the work with which it is combined will remain governed by version
3 of the GNU General Public License.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU Affero General Public License from time to time. Such new versions
will be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU Affero General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU Affero General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU Affero General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Affero General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If your software can interact with users remotely through a computer
network, you should also make sure that it provides a way for users to
get its source. For example, if your program is a web application, its
interface could display a "Source" link that leads users to an archive
of the code. There are many ways you could offer source, and different
solutions will be better for different programs; see section 13 for the
specific requirements.
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU AGPL, see
<https://www.gnu.org/licenses/>.

111
studio/Unsloth_Studio_Colab.ipynb Normal file
View file

@ -0,0 +1,111 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "f2b0c6a1",
"metadata": {},
"source": [
"**License Notice**\n",
"\n",
"SPDX-License-Identifier: AGPL-3.0-only\n",
"\n",
"Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "447c1156",
"metadata": {},
"outputs": [],
"source": [
"# ===========================================\n",
"# ⚠️ GPU Check - Run This First!\n",
"# ===========================================\n",
"import torch\n",
"\n",
"print(\"🔍 Checking for GPU...\")\n",
"if not torch.cuda.is_available():\n",
" print(\"❌ ERROR: No GPU detected!\")\n",
" print(\"\\n📋 To enable GPU:\")\n",
" print(\" 1. Go to: Runtime → Change runtime type\")\n",
" print(\" 2. Select: Hardware accelerator → GPU (T4 is free)\")\n",
" print(\" 3. Click: Save\")\n",
" print(\" 4. Restart and re-run all cells\")\n",
" raise RuntimeError(\"⛔ GPU required for Unsloth Studio\")\n",
"else:\n",
" gpu_name = torch.cuda.get_device_name(0)\n",
" print(f\"✅ GPU detected: {gpu_name}\")\n",
" print(\" Ready to proceed!\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "f04a9b46",
"metadata": {},
"outputs": [],
"source": [
"# ===========================================\n",
"# GitHub Authentication (Private Repo)\n",
"# ===========================================\n",
"from getpass import getpass\n",
"import os\n",
"\n",
"print(\"🔐 GitHub Token Required\")\n",
"print(\"Get token: https://github.com/settings/tokens\")\n",
"print(\"Scope needed: 'repo'\")\n",
"print(\"-\" * 50)\n",
"\n",
"github_token = getpass(\"Enter GitHub Token: \")\n",
"os.environ['GITHUB_TOKEN'] = github_token\n",
"print(\"✅ Token stored\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "27e68f91",
"metadata": {},
"outputs": [],
"source": [
"# ===========================================\n",
"# Setup: Clone repo and run setup\n",
"# ===========================================\n",
"\n",
"import os\n",
"github_token = os.environ['GITHUB_TOKEN']\n",
"!git clone https://{github_token}@github.com/unslothai/new-ui-prototype.git\n",
"%cd /content/new-ui-prototype\n",
"\n",
"# Run setup script\n",
"!chmod +x setup.sh\n",
"!./setup.sh"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "277e431e",
"metadata": {},
"outputs": [],
"source": [
"# ===========================================\n",
"# Start Unsloth Studio\n",
"# ===========================================\n",
"import sys\n",
"sys.path.insert(0, '/content/new-ui-prototype/studio/backend')\n",
"\n",
"from colab import start\n",
"start()"
]
}
],
"metadata": {
"language_info": {
"name": "python"
}
},
"nbformat": 4,
"nbformat_minor": 5
}

2
studio/__init__.py Normal file
View file

@ -0,0 +1,2 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

2
studio/backend/__init__.py Normal file
View file

@ -0,0 +1,2 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

2
studio/backend/assets/__init__.py Normal file
View file

@ -0,0 +1,2 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

2
studio/backend/assets/configs/__init__.py Normal file
View file

@ -0,0 +1,2 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

42
studio/backend/assets/configs/full_finetune.yaml Normal file
View file

@ -0,0 +1,42 @@
model: unsloth/Qwen2.5-0.5B
data:
dataset: tatsu-lab/alpaca
format_type: auto
training:
training_type: full
max_seq_length: 2048
load_in_4bit: false
output_dir: outputs
num_epochs: 1
learning_rate: 0.0002
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 0
save_steps: 0
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: false
gradient_checkpointing: "unsloth"
lora:
lora_r: 64
lora_alpha: 16
lora_dropout: 0.0
target_modules: ""
vision_all_linear: false
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: unsloth-training
enable_tensorboard: false
tensorboard_dir: runs

42
studio/backend/assets/configs/lora_text.yaml Normal file
View file

@ -0,0 +1,42 @@
model: unsloth/Qwen2.5-0.5B
data:
dataset: tatsu-lab/alpaca
format_type: auto
training:
training_type: lora
max_seq_length: 2048
load_in_4bit: true
output_dir: outputs
num_epochs: 1
learning_rate: 0.0002
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 0
save_steps: 0
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: false
gradient_checkpointing: "unsloth"
lora:
lora_r: 64
lora_alpha: 16
lora_dropout: 0.0
target_modules: "q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj"
vision_all_linear: false
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: unsloth-training
enable_tensorboard: false
tensorboard_dir: runs

56
studio/backend/assets/configs/model_defaults/default.yaml Normal file
View file

@ -0,0 +1,56 @@
# Default model training parameters
# Used for models without specific configurations
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 5e-5
batch_size: 2
gradient_accumulation_steps: 4
warmup_ratio: 0.1
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.7
top_p: 0.95
top_k: -1
min_p: 0.01

43
studio/backend/assets/configs/model_defaults/embedding/unsloth_Qwen3-Embedding-0.6B.yaml Normal file
View file

@ -0,0 +1,43 @@
# Model defaults for unsloth/Qwen3-Embedding-0.6B
# Based on Qwen3_Embedding_(0_6B).py embedding notebook
# Also applies to: unsloth/Qwen3-Embedding-4B
training:
max_seq_length: 512
# num_epochs: 2
num_epochs: 0
learning_rate: 3e-5
batch_size: 256
gradient_accumulation_steps: 1
warmup_ratio: 0.03
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: false
gradient_checkpointing: false
optim: "adamw_8bit"
lr_scheduler_type: "constant_with_warmup"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "embedding-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 50

39
studio/backend/assets/configs/model_defaults/embedding/unsloth_all-MiniLM-L6-v2.yaml Normal file
View file

@ -0,0 +1,39 @@
# Model defaults for unsloth/all-MiniLM-L6-v2
# Based on All_MiniLM_L6_v2.py embedding notebook
training:
max_seq_length: 512
# num_epochs: 2
num_epochs: 0
learning_rate: 2e-4
batch_size: 256
gradient_accumulation_steps: 1
warmup_ratio: 0.03
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: false
gradient_checkpointing: false
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 64
lora_alpha: 128
lora_dropout: 0.0
target_modules:
- "value"
- "key"
- "dense"
- "query"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "embedding-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 50

39
studio/backend/assets/configs/model_defaults/embedding/unsloth_bge-m3.yaml Normal file
View file

@ -0,0 +1,39 @@
# Model defaults for unsloth/bge-m3
# Based on BGE_M3.py embedding notebook
training:
max_seq_length: 512
# num_epochs: 2
num_epochs: 0
learning_rate: 3e-5
batch_size: 256
gradient_accumulation_steps: 1
warmup_ratio: 0.03
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: false
gradient_checkpointing: false
optim: "adamw_8bit"
lr_scheduler_type: "constant_with_warmup"
lora:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.0
target_modules:
- "key"
- "query"
- "dense"
- "value"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "embedding-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 50

42
studio/backend/assets/configs/model_defaults/embedding/unsloth_embeddinggemma-300m.yaml Normal file
View file

@ -0,0 +1,42 @@
# Model defaults for unsloth/embeddinggemma-300m
# Based on EmbeddingGemma_(300M).py embedding notebook
training:
max_seq_length: 1024
# num_epochs: 1
num_epochs: 0
learning_rate: 2e-5
batch_size: 64
gradient_accumulation_steps: 2
warmup_ratio: 0.03
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: false
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "embedding-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 5

38
studio/backend/assets/configs/model_defaults/embedding/unsloth_gte-modernbert-base.yaml Normal file
View file

@ -0,0 +1,38 @@
# Model defaults for unsloth/gte-modernbert-base
# Based on ModernBert.py embedding notebook
training:
max_seq_length: 512
# num_epochs: 2
num_epochs: 0
learning_rate: 3e-5
batch_size: 256
gradient_accumulation_steps: 1
warmup_ratio: 0.03
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: false
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "constant_with_warmup"
lora:
lora_r: 64
lora_alpha: 128
lora_dropout: 0.0
target_modules:
- "Wi"
- "Wo"
- "Wqkv"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "embedding-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 50

47
studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-21B-A3B-PT.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/ERNIE-4.5-21B-A3B-PT
# Based on ERNIE_4_5_21B_A3B_PT-Conversational.ipynb
# Also applies to: unsloth/ERNIE-4.5-21B-A3B-PT
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 4
gradient_accumulation_steps: 2
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

55
studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml Normal file
View file

@ -0,0 +1,55 @@
# Model defaults for unsloth/ERNIE-4.5-VL-28B-A3B-PT
# Based on ERNIE_4_5_VL_28B_A3B_PT_Vision.ipynb
# Also applies to: unsloth/ERNIE-4.5-VL-28B-A3B-PT
# added inference parameters from unsloth notebook
training:
trust_remote_code: true
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 2
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: true
temperature: 1.5
min_p: 0.1

47
studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for tiiuae/Falcon-H1-0.5B-Instruct
# Based on Falcon_H1_(0.5B)-Alpaca.ipynb
# Also applies to: tiiuae/Falcon-H1-0.5B-Instruct, unsloth/Falcon-H1-0.5B-Instruct
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 8
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: false
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.1
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

50
studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,50 @@
# Model defaults for unsloth/codegemma-7b-bnb-4bit
# Based on CodeGemma_(7B)-Conversational.ipynb
# Also applies to: unsloth/codegemma-7b, google/codegemma-7b
# added inference parameters from Ollama
training:
trust_remote_code: false
max_seq_length: 4096
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0
top_p: 0.9

53
studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for unsloth/functiongemma-270m-it
# Based on FunctionGemma_(270M).ipynb
# Also applies to: unsloth/functiongemma-270m-it-unsloth-bnb-4bit, google/functiongemma-270m-it, unsloth/functiongemma-270m-it-unsloth-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 4096
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 4
gradient_accumulation_steps: 2
warmup_steps: 10
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 128
lora_alpha: 256
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.0
top_k: 64
top_p: 0.95
min_p: 0.0

46
studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,46 @@
# Model defaults for unsloth/gemma-2-27b-bnb-4bit
# Based on Gemma2_(9B)-Alpaca.ipynb (same defaults for larger models)
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/gemma-2-2b
# Based on Gemma2_(2B)-Alpaca.ipynb
# Also applies to: unsloth/gemma-2-2b-bnb-4bit, google/gemma-2-2b
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

53
studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for unsloth/gemma-3-270m-it
# Based on Gemma3_(270M).ipynb
# Also applies to: unsloth/gemma-3-270m-it-unsloth-bnb-4bit, google/gemma-3-270m-it, unsloth/gemma-3-270m-it-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 5e-5
batch_size: 4
gradient_accumulation_steps: 1
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 128
lora_alpha: 128
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.0
top_k: 64
top_p: 0.95
min_p: 0.0

51
studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for unsloth/gemma-3-27b-it
# Based on Gemma3_(27B)_A100-Conversational.ipynb
# Also applies to: unsloth/gemma-3-27b-it-unsloth-bnb-4bit, google/gemma-3-27b-it, unsloth/gemma-3-27b-it-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 8
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.0
top_k: 64
top_p: 0.95
min_p: 0.0

51
studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for unsloth/gemma-3-4b-it
# Based on Gemma3_(4B).ipynb
# Also applies to: unsloth/gemma-3-4b-it-unsloth-bnb-4bit, google/gemma-3-4b-it, unsloth/gemma-3-4b-it-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 8
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.0
top_k: 64
top_p: 0.95
min_p: 0.0

51
studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for unsloth/gemma-3-4b-pt
# Based on Gemma3_(4B)-Vision.ipynb
# Also applies to: unsloth/gemma-3-4b-pt-unsloth-bnb-4bit, google/gemma-3-4b-pt, unsloth/gemma-3-4b-pt-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 2
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_ratio: 0.03
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: true
optim: "adamw_torch_fused"
lr_scheduler_type: "cosine"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.0
top_k: 64
top_p: 0.95
min_p: 0.0

53
studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for unsloth/gemma-3n-E4B-it
# Based on Gemma3N_(4B)-Conversational.ipynb
# Also applies to: unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit, google/gemma-3n-E4B-it, unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 1024
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 8
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
audio_input: true
inference:
trust_remote_code: false
temperature: 1.0
top_k: 64
top_p: 0.95
min_p: 0.0

53
studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for unsloth/gemma-3n-E4B
# Based on Gemma3N_(4B)-Vision.ipynb
# Also applies to: unsloth/gemma-3n-E4B-unsloth-bnb-4bit, google/gemma-3n-E4B
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 2
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_ratio: 0.03
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: true
optim: "adamw_torch_fused"
lr_scheduler_type: "cosine"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
audio_input: true
inference:
trust_remote_code: false
temperature: 1.0
top_k: 64
top_p: 0.95
min_p: 0.0

52
studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml Normal file
View file

@ -0,0 +1,52 @@
# Model defaults for unsloth/gpt-oss-120b
# Based on gpt-oss-(120B)_A100-Fine-tuning.ipynb
# Also applies to: openai/gpt-oss-120b, unsloth/gpt-oss-120b-unsloth-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 4096
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 4
gradient_accumulation_steps: 1
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.0
top_p: 1.0
top_k: 0

52
studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml Normal file
View file

@ -0,0 +1,52 @@
# Model defaults for unsloth/gpt-oss-20b
# Based on gpt-oss-(20B)-Fine-tuning.ipynb
# Also applies to: openai/gpt-oss-20b, unsloth/gpt-oss-20b-unsloth-bnb-4bit, unsloth/gpt-oss-20b-BF16
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 1024
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.0
top_p: 1.0
top_k: 0

54
studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,54 @@
# Model defaults for unsloth/granite-4.0-350m
# Based on Granite4.0_350M.ipynb
# Also applies to: ibm-granite/granite-4.0-350m, unsloth/granite-4.0-350m-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
- "shared_mlp.input_linear"
- "shared_mlp.output_linear"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.0
top_p: 1.0
top_k: 0

54
studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml Normal file
View file

@ -0,0 +1,54 @@
# Model defaults for unsloth/granite-4.0-h-micro
# Based on Granite4.0.ipynb
# Also applies to: ibm-granite/granite-4.0-h-micro, unsloth/granite-4.0-h-micro-bnb-4bit, unsloth/granite-4.0-h-micro-unsloth-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
- "shared_mlp.input_linear"
- "shared_mlp.output_linear"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.0
top_p: 1.0
top_k: 0

49
studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml Normal file
View file

@ -0,0 +1,49 @@
# Model defaults for unsloth/Llama-3.2-11B-Vision-Instruct
# Based on Llama3.2_(11B)-Vision.ipynb
# Also applies to: unsloth/Llama-3.2-11B-Vision-Instruct-unsloth-bnb-4bit, meta-llama/Llama-3.2-11B-Vision-Instruct, unsloth/Llama-3.2-11B-Vision-Instruct-bnb-4bit
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.5
min_p: 0.1

47
studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Llama-3.2-1B-Instruct
# Based on Llama3.2_(1B)-RAFT.ipynb
# Also applies to: unsloth/Llama-3.2-1B-Instruct-unsloth-bnb-4bit, meta-llama/Llama-3.2-1B-Instruct, unsloth/Llama-3.2-1B-Instruct-bnb-4bit, RedHatAI/Llama-3.2-1B-Instruct-FP8, unsloth/Llama-3.2-1B-Instruct-FP8-Block, unsloth/Llama-3.2-1B-Instruct-FP8-Dynamic
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 5
num_epochs: 0
learning_rate: 2e-5
batch_size: 1
gradient_accumulation_steps: 8
warmup_steps: 0
max_steps: 30
save_steps: 30
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: true
optim: "adamw_torch"
lr_scheduler_type: "cosine"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

51
studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for unsloth/Llama-3.2-3B-Instruct
# Based on Llama3.2_(1B_and_3B)-Conversational.ipynb
# Also applies to: unsloth/Llama-3.2-3B-Instruct-unsloth-bnb-4bit, meta-llama/Llama-3.2-3B-Instruct, unsloth/Llama-3.2-3B-Instruct-bnb-4bit, RedHatAI/Llama-3.2-3B-Instruct-FP8, unsloth/Llama-3.2-3B-Instruct-FP8-Block, unsloth/Llama-3.2-3B-Instruct-FP8-Dynamic
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.5
min_p: 0.1

51
studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for unsloth/Llama-3.3-70B-Instruct
# Based on Llama3.3_(70B)_A100-Conversational.ipynb
# Also applies to: unsloth/Llama-3.3-70B-Instruct-unsloth-bnb-4bit, meta-llama/Llama-3.3-70B-Instruct, unsloth/Llama-3.3-70B-Instruct-bnb-4bit, RedHatAI/Llama-3.3-70B-Instruct-FP8, unsloth/Llama-3.3-70B-Instruct-FP8-Block, unsloth/Llama-3.3-70B-Instruct-FP8-Dynamic
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.5
min_p: 0.1

47
studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Meta-Llama-3.1-70B-bnb-4bit
# Based on Llama3.1_(8B)-Alpaca.ipynb
# Also applies to: unsloth/Meta-Llama-3.1-8B-bnb-4bit, unsloth/Meta-Llama-3.1-8B-unsloth-bnb-4bit, meta-llama/Meta-Llama-3.1-8B, unsloth/Meta-Llama-3.1-8B, unsloth/Meta-Llama-3.1-70B, meta-llama/Meta-Llama-3.1-70B, unsloth/Meta-Llama-3.1-405B-bnb-4bit, meta-llama/Meta-Llama-3.1-405B
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit
# Based on Llama3.1_(8B)-Inference.ipynb
# Also applies to: "unsloth/Meta-Llama-3.1-8B-Instruct-unsloth-bnb-4bit", "meta-llama/Meta-Llama-3.1-8B-Instruct", "unsloth/Meta-Llama-3.1-8B-Instruct","RedHatAI/Llama-3.1-8B-Instruct-FP8","unsloth/Llama-3.1-8B-Instruct-FP8-Block","unsloth/Llama-3.1-8B-Instruct-FP8-Dynamic"
training:
trust_remote_code: false
max_seq_length: 8192
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/llama-3-8b-Instruct-bnb-4bit
# Based on Llama3_(8B)-Conversational.ipynb
# Also applies to: unsloth/llama-3-8b-Instruct, meta-llama/Meta-Llama-3-8B-Instruct
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/llama-3-8b-bnb-4bit
# Based on Llama3_(8B)-Alpaca.ipynb
# Also applies to: unsloth/llama-3-8b, meta-llama/Meta-Llama-3-8B
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

46
studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml Normal file
View file

@ -0,0 +1,46 @@
# Model defaults for unsloth/Llasa-3B
# Based on Llasa_TTS_(3B).ipynb and Llasa_TTS_(1B).ipynb
# Also applies to: HKUSTAudio/Llasa-1B
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 5e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 128
lora_alpha: 128
lora_dropout: 0.0
target_modules:
- "q_proj"
- "v_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.2
top_p: 1.2

56
studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,56 @@
# Model defaults for unsloth/Magistral-Small-2509
# Based on Magistral_(24B)-Reasoning-Conversational.ipynb
# Also applies to: mistralai/Magistral-Small-2509, unsloth/Magistral-Small-2509-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 2
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.7
min_p: 0.01
top_p: 0.95

55
studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml Normal file
View file

@ -0,0 +1,55 @@
# Model defaults for unsloth/Ministral-3-3B-Instruct-2512
# Based on Ministral_3_VL_(3B)_Vision.ipynb
# Also applies to: unsloth/Ministral-3-3B-Instruct-2512
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 4
gradient_accumulation_steps: 2
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.15
top_p: default

47
studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Mistral-Nemo-Base-2407-bnb-4bit
# Based on Mistral_Nemo_(12B)-Alpaca.ipynb
# Also applies to: "unsloth/Mistral-Nemo-Base-2407", "mistralai/Mistral-Nemo-Base-2407", "unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit", "unsloth/Mistral-Nemo-Instruct-2407", "mistralai/Mistral-Nemo-Instruct-2407",
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Mistral-Small-Instruct-2409
# Based on Mistral_Small_(22B)-Alpaca.ipynb
# Also applies to: unsloth/Mistral-Small-Instruct-2409-bnb-4bit, mistralai/Mistral-Small-Instruct-2409
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

49
studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml Normal file
View file

@ -0,0 +1,49 @@
# Model defaults for unsloth/Pixtral-12B-2409
# Based on Pixtral_(12B)-Vision.ipynb
# Also applies to: unsloth/Pixtral-12B-2409-unsloth-bnb-4bit, mistralai/Pixtral-12B-2409, unsloth/Pixtral-12B-2409-bnb-4bit
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "paged_adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 8
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: false
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.5
min_p: 0.1

47
studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/mistral-7b-instruct-v0.3-bnb-4bit
# Based on Mistral_v0.3_(7B)-Conversational.ipynb
# Also applies to: unsloth/mistral-7b-instruct-v0.3, mistralai/Mistral-7B-Instruct-v0.3
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

46
studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,46 @@
# Model defaults for unsloth/mistral-7b-v0.3-bnb-4bit
# Based on Mistral_v0.3_(7B)-Alpaca.ipynb
# Also applies to: "unsloth/mistral-7b-v0.3", "mistralai/Mistral-7B-v0.3",
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

51
studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for OuteAI/Llama-OuteTTS-1.0-1B
# Based on Oute_TTS_(1B).ipynb
# Also applies to: OuteAI/Llama-OuteTTS-1.0-1B
# added inference parameters from unsloth notebook
audio_type: dac
training:
trust_remote_code: false
eval_steps: 0
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 128
lora_alpha: 128
lora_dropout: 0.0
target_modules:
- "q_proj"
- "v_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.4
top_k: 40
top_p: 0.9
min_p: 0.05

55
studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml Normal file
View file

@ -0,0 +1,55 @@
# Model defaults for Spark-TTS-0.5B/LLM
# Based on Spark_TTS_(0_5B).ipynb
# Also applies to: Spark-TTS-0.5B/LLM
# added inference parameters from unsloth notebook
audio_type: bicodec
training:
trust_remote_code: false
eval_steps: 0
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 128
lora_alpha: 128
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.8
top_k: 50
top_p: 1.0

50
studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml Normal file
View file

@ -0,0 +1,50 @@
# Model defaults for sesame/csm-1b
# Based on Sesame_CSM_(1B)-TTS.ipynb
# Also applies to: sesame/csm-1b
audio_type: csm
training:
trust_remote_code: false
eval_steps: 0
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

52
studio/backend/assets/configs/model_defaults/other/unsloth_GLM-4.7-Flash.yaml Normal file
View file

@ -0,0 +1,52 @@
# Model defaults for unsloth/GLM-4.7-Flash
# Based on GLM_Flash_A100(80GB).py
# Also applies to: unsloth/GLM-4.7-Flash-unsloth-bnb-4bit, unsloth/GLM-4.7-Flash-bnb-4bit, THUDM/GLM-4.7-Flash
training:
trust_remote_code: true
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 4
gradient_accumulation_steps: 2
warmup_steps: 5
max_steps: 60
save_steps: 60
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
- "out_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: true
temperature: 0.7
top_p: 0.8
top_k: 20

45
studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml Normal file
View file

@ -0,0 +1,45 @@
# Model defaults for unsloth/LFM2-1.2B
# Based on Liquid_LFM2_(1.2B)-Conversational.ipynb
# Also applies to: unsloth/LFM2-1.2B
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.3
min_p: 0.15

53
studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for unsloth/Nemotron-3-Nano-30B-A3B
# Based on Nemotron-3-Nano-30B-A3B_A100.ipynb
# Also applies to: unsloth/Nemotron-3-Nano-30B-A3B
# added inference parameters from unsloth guides
training:
trust_remote_code: true
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 4
gradient_accumulation_steps: 2
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 8
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
- "in_proj"
- "out_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: true
temperature: 1.0
top_p: 1.0

55
studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml Normal file
View file

@ -0,0 +1,55 @@
# Model defaults for unsloth/PaddleOCR-VL
# Based on Paddle_OCR_(1B)_Vision.ipynb
# Also applies to: unsloth/PaddleOCR-VL
# added inference parameters from unsloth notebook
training:
trust_remote_code: true
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 5e-5
batch_size: 4
gradient_accumulation_steps: 2
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 64
lora_alpha: 64
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: true
temperature: 1.5
min_p: 0.1

46
studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml Normal file
View file

@ -0,0 +1,46 @@
# Model defaults for answerdotai/ModernBERT-large
# Based on bert_classification.ipynb
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 1
num_epochs: 0
learning_rate: 5e-5
batch_size: 32
gradient_accumulation_steps: 1
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

54
studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml Normal file
View file

@ -0,0 +1,54 @@
# Model defaults for unsloth/orpheus-3b-0.1-ft
# Based on Orpheus_(3B)-TTS.ipynb
# Also applies to: unsloth/orpheus-3b-0.1-ft-unsloth-bnb-4bit, canopylabs/orpheus-3b-0.1-ft, unsloth/orpheus-3b-0.1-ft-bnb-4bit
# added inference parameters from unsloth notebook
audio_type: snac
training:
trust_remote_code: false
eval_steps: 0
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 64
lora_alpha: 64
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.6
top_p: 0.95

47
studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/tinyllama
# Based on TinyLlama_(1.1B)-Alpaca.ipynb
# Also applies to: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
training:
trust_remote_code: false
max_seq_length: 4096
# num_epochs: 1
num_epochs: 0
learning_rate: 2e-5
batch_size: 2
gradient_accumulation_steps: 4
warmup_ratio: 0.1
max_steps: 30
save_steps: 30
weight_decay: 0.1
random_seed: 3407
packing: true
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

46
studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml Normal file
View file

@ -0,0 +1,46 @@
# Model defaults for unsloth/whisper-large-v3
# Based on Whisper.ipynb
# Also applies to: unsloth/whisper-large-v3, openai/whisper-large-v3
audio_type: whisper
audio_input: true
training:
trust_remote_code: false
eval_steps: 5
max_seq_length: 448
# num_epochs: 4
num_epochs: 0
learning_rate: 1e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 64
lora_alpha: 64
lora_dropout: 0.0
target_modules:
- "q_proj"
- "v_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Phi-3-medium-4k-instruct
# Based on Phi_3_Medium-Conversational.ipynb
# Also applies to: "unsloth/Phi-3-medium-4k-instruct-bnb-4bit", "microsoft/Phi-3-medium-4k-instruct",
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Phi-3.5-mini-instruct
# Based on Phi_3.5_Mini-Conversational.ipynb
# Also applies to: "unsloth/Phi-3.5-mini-instruct-bnb-4bit", "microsoft/Phi-3.5-mini-instruct"
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

51
studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for unsloth/Phi-4
# Based on Phi_4-Conversational.ipynb
# Also applies to: unsloth/phi-4-unsloth-bnb-4bit, microsoft/phi-4, unsloth/phi-4-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.8
top_p: 0.95

53
studio/backend/assets/configs/model_defaults/qwen/imdatta0_tiny_qwen3_moe_2.8B_0.7B.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for imdatta0/tiny_qwen3_moe_2.8B_0.7B
# Based on TinyQwen3_MoE.py
# Dummy model of qwen3moe architecture created to fit in T4
# MoE model - includes gate_up_proj for MoE layers
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 1
warmup_steps: 5
max_steps: 50
save_steps: 50
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
- "gate_up_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.6
top_k: 20
top_p: 0.95

47
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Qwen2-7B
# Based on Qwen2_(7B)-Alpaca.ipynb
# Also applies to: unsloth/Qwen2-7B-bnb-4bit, Qwen/Qwen2-7B
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

49
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml Normal file
View file

@ -0,0 +1,49 @@
# Model defaults for unsloth/Qwen2-VL-7B-Instruct
# Based on Qwen2_VL_(7B)-Vision.ipynb
# Also applies to: unsloth/Qwen2-VL-7B-Instruct-unsloth-bnb-4bit, Qwen/Qwen2-VL-7B-Instruct, unsloth/Qwen2-VL-7B-Instruct-bnb-4bit
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.5
min_p: 0.1

47
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Qwen2.5-1.5B-Instruct
# Based on nemo_gym_sudoku.ipynb
# Also applies to: unsloth/Qwen2.5-1.5B-Instruct-unsloth-bnb-4bit, Qwen/Qwen2.5-1.5B-Instruct, unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit
training:
trust_remote_code: false
max_seq_length: 4096
# num_epochs: 4
num_epochs: 0
learning_rate: 1e-5
batch_size: 1
gradient_accumulation_steps: 64
warmup_ratio: 0.1
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 42
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 4
lora_alpha: 8
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Qwen2.5-7B
# Based on Qwen2.5_(7B)-Alpaca.ipynb
# Also applies to: unsloth/Qwen2.5-7B-unsloth-bnb-4bit, Qwen/Qwen2.5-7B, unsloth/Qwen2.5-7B-bnb-4bit
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

47
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Qwen2.5-Coder-1.5B-Instruct
# Based on Qwen2.5_Coder_(1.5B)-Tool_Calling.ipynb
# Also applies to: unsloth/Qwen2.5-Coder-1.5B-Instruct-bnb-4bit, Qwen/Qwen2.5-Coder-1.5B-Instruct
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

51
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for unsloth/Qwen2.5-Coder-14B-Instruct
# Based on Qwen2.5_Coder_(14B)-Conversational.ipynb
# Also applies to: unsloth/Qwen2.5-Coder-14B-Instruct-bnb-4bit, Qwen/Qwen2.5-Coder-14B-Instruct
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "paged_adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.5
min_p: 0.1

47
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,47 @@
# Model defaults for unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit
# Based on CodeForces-cot-Finetune_for_Reasoning_on_CodeForces.ipynb
# Also applies to: unsloth/Qwen2.5-Coder-7B-Instruct, Qwen/Qwen2.5-Coder-7B-Instruct
training:
trust_remote_code: false
max_seq_length: 32768
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false

49
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,49 @@
# Model defaults for unsloth/Qwen2.5-VL-7B-Instruct-bnb-4bit
# Based on Qwen2.5_VL_(7B)-Vision.ipynb
# Also applies to: unsloth/Qwen2.5-VL-7B-Instruct, Qwen/Qwen2.5-VL-7B-Instruct, unsloth/Qwen2.5-VL-7B-Instruct-unsloth-bnb-4bit
# added inference parameters from unsloth notebook
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 1.5
min_p: 0.1

52
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml Normal file
View file

@ -0,0 +1,52 @@
# Model defaults for unsloth/Qwen3-0.6B
# Based on Qwen3_(0_6B)-Phone_Deployment.ipynb
# Also applies to: unsloth/Qwen3-0.6B-unsloth-bnb-4bit, Qwen/Qwen3-0.6B, unsloth/Qwen3-0.6B-bnb-4bit, Qwen/Qwen3-0.6B-FP8, unsloth/Qwen3-0.6B-FP8
# added inference parameters from Ollama
training:
trust_remote_code: false
max_seq_length: 1024
# num_epochs: 4
num_epochs: 0
learning_rate: 5e-5
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.6
top_k: 20
top_p: 0.95

51
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,51 @@
# Model defaults for unsloth/Qwen3-14B-Base
# Based on Qwen3_(14B)-Alpaca.ipynb
# Also applies to: unsloth/Qwen3-14B-Base, Qwen/Qwen3-14B-Base, unsloth/Qwen3-14B-Base-bnb-4bit
# added inference parameters from Ollama
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.6
top_k: 20
top_p: 0.95

52
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml Normal file
View file

@ -0,0 +1,52 @@
# Model defaults for unsloth/Qwen3-14B
# Based on Qwen3_(14B).ipynb
# Also applies to: unsloth/Qwen3-14B-unsloth-bnb-4bit, Qwen/Qwen3-14B, unsloth/Qwen3-14B-bnb-4bit, Qwen/Qwen3-14B-FP8, unsloth/Qwen3-14B-FP8
# added inference parameters from Ollama
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.6
top_k: 20
top_p: 0.95

53
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-30B-A3B-Instruct-2507.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for unsloth/Qwen3-30B-A3B-Instruct-2507
# Based on Qwen3_MoE.py
# Also applies to: Qwen/Qwen3-30B-A3B-Instruct-2507, unsloth/Qwen3-30B-A3B-Instruct-2507-bnb-4bit
# MoE model - includes gate_up_proj for MoE layers
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 1
gradient_accumulation_steps: 1
warmup_steps: 5
max_steps: 50
save_steps: 50
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
- "gate_up_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.6
top_k: 20
top_p: 0.95

52
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml Normal file
View file

@ -0,0 +1,52 @@
# Model defaults for unsloth/Qwen3-32B
# Based on Qwen3_(32B)_A100-Reasoning-Conversational.ipynb
# Also applies to: unsloth/Qwen3-32B-unsloth-bnb-4bit, Qwen/Qwen3-32B, unsloth/Qwen3-32B-bnb-4bit, Qwen/Qwen3-32B-FP8, unsloth/Qwen3-32B-FP8
# added inference parameters from Ollama
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.6
top_k: 20
top_p: 0.95

53
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for unsloth/Qwen3-4B-Instruct-2507
# Based on Qwen3_(4B)-Instruct.ipynb
# Also applies to: unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit, Qwen/Qwen3-4B-Instruct-2507, unsloth/Qwen3-4B-Instruct-2507-bnb-4bit, Qwen/Qwen3-4B-Instruct-2507-FP8, unsloth/Qwen3-4B-Instruct-2507-FP8
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.7
top_p: 0.80
top_k: 20
min_p: 0.00

53
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml Normal file
View file

@ -0,0 +1,53 @@
# Model defaults for unsloth/Qwen3-4B-Thinking-2507
# Based on Qwen3_(4B)-Thinking.ipynb
# Also applies to: unsloth/Qwen3-4B-Thinking-2507-unsloth-bnb-4bit, Qwen/Qwen3-4B-Thinking-2507, unsloth/Qwen3-4B-Thinking-2507-bnb-4bit, Qwen/Qwen3-4B-Thinking-2507-FP8, unsloth/Qwen3-4B-Thinking-2507-FP8
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
target_modules:
- "q_proj"
- "k_proj"
- "v_proj"
- "o_proj"
- "gate_proj"
- "up_proj"
- "down_proj"
use_rslora: false
use_loftq: false
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.6
top_p: 0.95
top_k: 20
min_p: 0.00

50
studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml Normal file
View file

@ -0,0 +1,50 @@
# Model defaults for unsloth/Qwen3-VL-8B-Instruct
# Based on Qwen3_VL_(8B)-Vision.ipynb
# Also applies to: Qwen/Qwen3-VL-8B-Instruct-FP8, unsloth/Qwen3-VL-8B-Instruct-FP8, unsloth/Qwen3-VL-8B-Instruct, Qwen/Qwen3-VL-8B-Instruct, unsloth/Qwen3-VL-8B-Instruct-bnb-4bit
# added inference parameters from unsloth guides
training:
trust_remote_code: false
max_seq_length: 2048
# num_epochs: 4
num_epochs: 0
learning_rate: 2e-4
batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 30
save_steps: 30
weight_decay: 0.001
random_seed: 3407
packing: false
train_on_completions: true
gradient_checkpointing: "unsloth"
optim: "adamw_8bit"
lr_scheduler_type: "linear"
lora:
lora_r: 16
lora_alpha: 16
lora_dropout: 0.0
target_modules:
- "all-linear"
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: "llm-finetuning"
enable_tensorboard: false
tensorboard_dir: "runs"
log_frequency: 10
inference:
trust_remote_code: false
temperature: 0.7
top_p: 0.8
top_k: 20

42
studio/backend/assets/configs/vision_lora.yaml Normal file
View file

@ -0,0 +1,42 @@
model: unsloth/Qwen2-VL-2B-Instruct-bnb-4bit
data:
dataset: philschmid/amazon-product-descriptions-vlm
format_type: auto
training:
training_type: lora
max_seq_length: 2048
load_in_4bit: true
output_dir: outputs
num_epochs: 1
learning_rate: 0.0002
batch_size: 1
gradient_accumulation_steps: 4
warmup_steps: 5
max_steps: 0
save_steps: 0
weight_decay: 0.01
random_seed: 3407
packing: false
train_on_completions: false
gradient_checkpointing: "unsloth"
lora:
lora_r: 64
lora_alpha: 16
lora_dropout: 0.0
target_modules: "" # vision uses vision_all_linear by default
vision_all_linear: true
use_rslora: false
use_loftq: false
finetune_vision_layers: true
finetune_language_layers: true
finetune_attention_modules: true
finetune_mlp_modules: true
logging:
enable_wandb: false
wandb_project: unsloth-training
enable_tensorboard: false
tensorboard_dir: runs

1288
studio/backend/assets/datasets/alpaca_unsloth.json Normal file
View file

File diff suppressed because it is too large Load diff

0
studio/backend/auth/.gitkeep Normal file
View file

47
studio/backend/auth/__init__.py Normal file
View file

@ -0,0 +1,47 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""
Authentication module for JWT-based auth with SQLite storage.
"""
from .authentication import (
create_access_token,
create_refresh_token,
refresh_access_token,
get_current_subject,
reload_secret,
)
from .storage import (
is_initialized,
create_initial_user,
get_user_and_secret,
load_jwt_secret,
save_setup_token,
consume_setup_token,
has_pending_setup_token,
save_refresh_token,
verify_refresh_token,
revoke_user_refresh_tokens,
)
from .hashing import hash_password, verify_password
__all__ = [
"create_access_token",
"create_refresh_token",
"refresh_access_token",
"get_current_subject",
"reload_secret",
"is_initialized",
"create_initial_user",
"get_user_and_secret",
"load_jwt_secret",
"save_setup_token",
"consume_setup_token",
"has_pending_setup_token",
"save_refresh_token",
"verify_refresh_token",
"revoke_user_refresh_tokens",
"hash_password",
"verify_password",
]

108
studio/backend/auth/authentication.py Normal file
View file

@ -0,0 +1,108 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import secrets
from datetime import datetime, timedelta, timezone
from typing import Optional
from fastapi import Depends, HTTPException, status
from fastapi.security import HTTPAuthorizationCredentials, HTTPBearer
import jwt
from .storage import load_jwt_secret, save_refresh_token, verify_refresh_token
ALGORITHM = "HS256"
ACCESS_TOKEN_EXPIRE_MINUTES = 60
REFRESH_TOKEN_EXPIRE_DAYS = 7
# Load stable secret from SQLite (set during first-time setup)
# This will raise RuntimeError if auth hasn't been initialized yet
try:
SECRET_KEY = load_jwt_secret()
except RuntimeError:
# Fallback: use a temporary secret until setup is complete
# This allows the app to start, but protected routes will fail until setup
SECRET_KEY = secrets.token_urlsafe(64)
security = HTTPBearer() # Reads Authorization: Bearer <token>
def create_access_token(
subject: str,
expires_delta: Optional[timedelta] = None,
) -> str:
"""
Create a signed JWT for the given subject (e.g. username).
Tokens are valid across restarts because SECRET_KEY is stored in SQLite.
"""
to_encode = {"sub": subject}
expire = datetime.now(timezone.utc) + (
expires_delta or timedelta(minutes = ACCESS_TOKEN_EXPIRE_MINUTES)
)
to_encode.update({"exp": expire})
return jwt.encode(to_encode, SECRET_KEY, algorithm = ALGORITHM)
def create_refresh_token(subject: str) -> str:
"""
Create a random refresh token, store its hash in SQLite, and return it.
Refresh tokens are opaque (not JWTs) and expire after REFRESH_TOKEN_EXPIRE_DAYS.
"""
token = secrets.token_urlsafe(48)
expires_at = datetime.now(timezone.utc) + timedelta(days = REFRESH_TOKEN_EXPIRE_DAYS)
save_refresh_token(token, subject, expires_at.isoformat())
return token
def refresh_access_token(refresh_token: str) -> Optional[str]:
"""
Validate a refresh token and issue a new access token.
The refresh token itself is NOT consumed — it stays valid until expiry.
Returns a new access_token or None if the refresh token is invalid/expired.
"""
username = verify_refresh_token(refresh_token)
if username is None:
return None
return create_access_token(subject = username)
def reload_secret() -> None:
"""
Reload the JWT secret from SQLite.
Call this after setup to ensure new tokens use the persistent secret.
"""
global SECRET_KEY
SECRET_KEY = load_jwt_secret()
async def get_current_subject(
credentials: HTTPAuthorizationCredentials = Depends(security),
) -> str:
"""
FastAPI dependency to validate the JWT and return the subject.
Use this as a dependency on routes that should be protected, e.g.:
@router.get("/secure")
async def secure_endpoint(current_subject: str = Depends(get_current_subject)):
...
"""
token = credentials.credentials
try:
payload = jwt.decode(token, SECRET_KEY, algorithms = [ALGORITHM])
subject: Optional[str] = payload.get("sub")
if subject is None:
raise HTTPException(
status_code = status.HTTP_401_UNAUTHORIZED,
detail = "Invalid token payload",
)
return subject
except jwt.InvalidTokenError:
raise HTTPException(
status_code = status.HTTP_401_UNAUTHORIZED,
detail = "Invalid or expired token",
)

43
studio/backend/auth/hashing.py Normal file
View file

@ -0,0 +1,43 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""
Password hashing utilities using PBKDF2.
"""
import hashlib
import hmac
import secrets
from typing import Tuple
def hash_password(password: str, salt: str | None = None) -> Tuple[str, str]:
"""
Hash a password using PBKDF2-HMAC-SHA256.
Returns (salt, hex_hash) tuple.
"""
if salt is None:
salt = secrets.token_hex(16)
dk = hashlib.pbkdf2_hmac(
"sha256",
password.encode("utf-8"),
salt.encode("utf-8"),
100_000, # 100k iterations
)
return salt, dk.hex()
def verify_password(password: str, salt: str, hashed: str) -> bool:
"""
Verify a password against a stored salt and hash.
Uses constant-time comparison to prevent timing attacks.
"""
dk = hashlib.pbkdf2_hmac(
"sha256",
password.encode("utf-8"),
salt.encode("utf-8"),
100_000,
)
return hmac.compare_digest(dk.hex(), hashed)

263
studio/backend/auth/storage.py Normal file
View file

@ -0,0 +1,263 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""
SQLite storage for authentication data (user credentials + JWT secret).
"""
import hashlib
import sqlite3
from datetime import datetime, timezone
from typing import Optional, Tuple
from utils.paths import auth_db_path, ensure_dir
DB_PATH = auth_db_path()
def _hash_token(token: str) -> str:
"""SHA-256 hash a setup token for safe storage."""
return hashlib.sha256(token.encode("utf-8")).hexdigest()
def get_connection() -> sqlite3.Connection:
"""Get a connection to the auth database, creating tables if needed."""
ensure_dir(DB_PATH.parent)
conn = sqlite3.connect(DB_PATH)
conn.row_factory = sqlite3.Row
conn.execute(
"""
CREATE TABLE IF NOT EXISTS auth_user (
id INTEGER PRIMARY KEY,
username TEXT UNIQUE NOT NULL,
password_salt TEXT NOT NULL,
password_hash TEXT NOT NULL,
jwt_secret TEXT NOT NULL
);
"""
)
conn.execute(
"""
CREATE TABLE IF NOT EXISTS setup_tokens (
id INTEGER PRIMARY KEY,
token_hash TEXT NOT NULL
);
"""
)
conn.execute(
"""
CREATE TABLE IF NOT EXISTS refresh_tokens (
id INTEGER PRIMARY KEY,
token_hash TEXT NOT NULL,
username TEXT NOT NULL,
expires_at TEXT NOT NULL
);
"""
)
conn.commit()
return conn
def is_initialized() -> bool:
"""Check if auth has been set up (user exists in DB)."""
conn = get_connection()
cur = conn.execute("SELECT COUNT(*) AS c FROM auth_user")
row = cur.fetchone()
conn.close()
return bool(row["c"])
def create_initial_user(username: str, password: str, jwt_secret: str) -> None:
"""
Create the initial admin user in the database.
Raises sqlite3.IntegrityError if username already exists.
"""
from .hashing import hash_password
salt, pwd_hash = hash_password(password)
conn = get_connection()
try:
conn.execute(
"""
INSERT INTO auth_user (username, password_salt, password_hash, jwt_secret)
VALUES (?, ?, ?, ?)
""",
(username, salt, pwd_hash, jwt_secret),
)
conn.commit()
finally:
conn.close()
def delete_user(username: str) -> None:
"""
Delete a user from the database.
Used for rollback when setup fails after user creation.
"""
conn = get_connection()
try:
conn.execute("DELETE FROM auth_user WHERE username = ?", (username,))
conn.commit()
finally:
conn.close()
def get_user_and_secret(username: str) -> Optional[Tuple[str, str, str]]:
"""
Get user's password salt, hash, and JWT secret.
Returns (password_salt, password_hash, jwt_secret) or None if user not found.
"""
conn = get_connection()
try:
cur = conn.execute(
"""
SELECT password_salt, password_hash, jwt_secret
FROM auth_user
WHERE username = ?
""",
(username,),
)
row = cur.fetchone()
if not row:
return None
return row["password_salt"], row["password_hash"], row["jwt_secret"]
finally:
conn.close()
def load_jwt_secret() -> str:
"""
Load the JWT secret from the database.
Raises RuntimeError if auth is not initialized.
"""
conn = get_connection()
try:
cur = conn.execute("SELECT jwt_secret FROM auth_user LIMIT 1")
row = cur.fetchone()
if not row:
raise RuntimeError(
"Auth is not initialized. Please set up a password first."
)
return row["jwt_secret"]
finally:
conn.close()
def save_setup_token(token: str) -> None:
"""
Store a hashed setup token, replacing any existing one.
"""
token_hash = _hash_token(token)
conn = get_connection()
try:
conn.execute("DELETE FROM setup_tokens")
conn.execute("INSERT INTO setup_tokens (token_hash) VALUES (?)", (token_hash,))
conn.commit()
finally:
conn.close()
def consume_setup_token(token: str) -> bool:
"""
Verify a setup token and delete it if valid.
Returns True if the token was valid (and is now consumed), False otherwise.
"""
token_hash = _hash_token(token)
conn = get_connection()
try:
cur = conn.execute(
"SELECT id FROM setup_tokens WHERE token_hash = ?", (token_hash,)
)
row = cur.fetchone()
if row is None:
return False
conn.execute("DELETE FROM setup_tokens WHERE id = ?", (row["id"],))
conn.commit()
return True
finally:
conn.close()
def has_pending_setup_token() -> bool:
"""Check if a setup token is waiting to be consumed."""
conn = get_connection()
try:
cur = conn.execute("SELECT COUNT(*) AS c FROM setup_tokens")
row = cur.fetchone()
return bool(row["c"])
finally:
conn.close()
def save_refresh_token(token: str, username: str, expires_at: str) -> None:
"""
Store a hashed refresh token with its associated username and expiry.
"""
token_hash = _hash_token(token)
conn = get_connection()
try:
conn.execute(
"""
INSERT INTO refresh_tokens (token_hash, username, expires_at)
VALUES (?, ?, ?)
""",
(token_hash, username, expires_at),
)
conn.commit()
finally:
conn.close()
def verify_refresh_token(token: str) -> Optional[str]:
"""
Verify a refresh token and return the username.
Returns the username if valid and not expired, None otherwise.
The token is NOT consumed — it stays valid until it expires.
"""
token_hash = _hash_token(token)
conn = get_connection()
try:
# Clean up any expired tokens while we're here
conn.execute(
"DELETE FROM refresh_tokens WHERE expires_at < ?",
(datetime.now(timezone.utc).isoformat(),),
)
conn.commit()
cur = conn.execute(
"""
SELECT id, username, expires_at FROM refresh_tokens
WHERE token_hash = ?
""",
(token_hash,),
)
row = cur.fetchone()
if row is None:
return None
# Check expiry
expires_at = datetime.fromisoformat(row["expires_at"])
if datetime.now(timezone.utc) > expires_at:
conn.execute("DELETE FROM refresh_tokens WHERE id = ?", (row["id"],))
conn.commit()
return None
return row["username"]
finally:
conn.close()
def revoke_user_refresh_tokens(username: str) -> None:
"""Revoke all refresh tokens for a user (e.g. on logout)."""
conn = get_connection()
try:
conn.execute("DELETE FROM refresh_tokens WHERE username = ?", (username,))
conn.commit()
finally:
conn.close()

Some files were not shown because too many files have changed in this diff Show more