unslothai notebooks: a generated catalogue of Colab fine-tuning recipes
250+ Fine-tuning & RL Notebooks for text, vision, audio, embedding, TTS models.
At a glance
- What is it?
- unslothai/notebooks holds hundreds of fine-tuning and RL notebooks organised by model, and the root file's tables are produced by update_all_notebooks.py rather than by hand. The repository also carries nb/ and kaggle/ directories alongside helper scripts such as replace_text.py and update_max_seq_length.py.
- Who is it for?
- Use this repository when you want a runnable starting point for a specific model and task, and expect the recipe to live in a notebook rather than in a library you can call. It is not a place to look for a versioned release, because the repository has no GitHub releases, and it is not a training framework: the dependency list is notebook runtime only.
- Can I use it commercially?
- Yes, with conditions. LGPL-3.0 is a weak copyleft licence: you can use it inside commercial and closed-source software, but if you distribute changes to its own files, you must publish those changes under the same licence.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Jupyter Notebook, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- unslothai/notebooks5,810 ★ · +120
| Repository | 2026-09-23 | 2026-11-15 |
|---|---|---|
| unslothai/notebooks | 5.7k | 5.8k |
120 Stars in the last 53 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The tables in the root file are written by a script, not by an author
The model specific tables underneath the Main Notebooks table are generated. The root file says so in a comment that also functions as a fence: DO NOT EDIT MANUALLY THIS SECTION UNTIL `end of notebook links`, and THIS SECTION IS GENERATED BY `update_all_notebooks.py` AUTOMATICALLY, followed by a START OF EDITING marker that opens the generated block. That single arrangement explains most of the repository's shape. A table row is a record of a notebook that exists, not a reviewed description of it, so a row can advertise a recipe whose contents have drifted. It also means the row's Type column is the only classification you get: Vision, Fine-tuning, GRPO, Conversational, Embeddings, TTS, Audio, Alpaca, Chat UI. There is no difficulty, hardware requirement, or runtime column, because nothing generates one.
The same recipe is materialised into nb/, molab/ and kaggle/
The repository listing shows several parallel surfaces for one body of work: nb/, kaggle/, plus Template_Notebook.ipynb, an original_template/ directory, replace_text.py, update_max_seq_length.py, and a scripts/ directory.
Python 3.13 is the floor, and the notebook dependencies are runtime only
The project metadata is deliberately thin.
requires-python = ">=3.13"
dependencies = [
"ipykernel>=6.30.1",
"ipywidgets>=8.1.7",
"nbconvert>=7.16.6",
]Three dependencies, all of them notebook plumbing: a kernel, widgets, and a converter. The fine-tuning library itself is not in that list, and neither is anything for text, vision, audio, or TTS work. So the declared environment gets a notebook running, and what the notebook then installs is defined inside the notebook, which is the opposite of a pinned dependency graph. Two more files at the root tell you the intended runtime: a .python-version file and a uv.lock, so the project expects a managed virtual environment rather than a hand built one. The unsloth_windows.ps1 script and Dockerfile_DGX_Spark cover a Windows path and a DGX Spark image.
No importable package, and setuptools was told to stay out of the way
The pyproject.toml ends with a comment that explains an unusual packaging decision. With the flat layout and top level scripts/ and tests/ directories, setuptools would try to auto-discover them all as packages, which fails the pip install -e. invocation. The chosen fix is explicit discovery: packages = [] and py-modules = [], with generator helpers living as flat modules under scripts/ and imported via sys.path.insert at runtime by the tests. The consequence is a hard boundary for anyone expecting a library. Nothing here can be imported as a dependency, the version field is the placeholder 0.1.0, and the description field still reads "Add your description here". Tests run with testpaths = ["tests"] and pythonpath = ["."], which tells you the test suite is run from a checkout with the repository root on the path, not from an installed artifact.
GRPO in this catalogue means Sudoku and 2048
The RL side of the catalogue is where the type column earns its keep, and it is also where expectations need setting. GRPO appears against gpt-oss (20B) and against a Qwen3 advanced GRPO entry, and the Gemma 4 table adds three reinforcement learning notebooks for the E2B model that are named after their tasks: Sudoku, auto kernel creation, and a 2048 game. If you came to this repository for a production reinforcement learning pipeline, the root file gives you a scaffold and a vocabulary, not an algorithm implementation. The description on the repository covers text, vision, audio, embedding, and TTS models, and the notebook types in the tables are the concrete form of that list.
Every link points at Colab, so the canonical copy is a rendering
Every notebook link in the root file resolves to colab.research.google.com with the file path inside the repository, which sets the expectation clearly: the recipe is meant to be opened and run there, with the repository as the storage layer. The root file adds that the notebooks run locally and feature data prep, training and inference, and it points to a separate documentation site at unsloth.ai/docs for the full catalogue and a fine-tuning guide. The links point at files on main, so the version a reader executes is whatever main held when the link resolved.
LGPL-3.0 covers the notebooks, not the models you point them at
The repository is licensed under LGPL-3.0, and that licence attaches to the notebooks and the scripts in this tree. The models the notebooks fine-tune are a separate matter, and the root file's tables name families from Gemma, Qwen, Llama, Phi, Mistral, gpt-oss, and Orpheus. For a team shipping a fine-tuned artefact, that gap is the one to close before anything else, because the output licence is set by the base model's licence and by your data, not by the notebook you started from. There is also no GitHub release to cite: the repository has none, so a notebook is identified only by its path on main, whose last recorded push is 2026-10-31. That date, rather than a version number, is the closest thing the repository offers to a release marker.
Editorial conclusion
Use this repository when you want a runnable starting point for a specific model and task, and expect the recipe to live in a notebook rather than in a library you can call. It is not a place to look for a versioned release, because the repository has no GitHub releases, and it is not a training framework: the dependency list is notebook runtime only. Before you commit to a recipe, check that the notebook's task matches yours and that your Python is 3.13 or newer.
Frequently asked questions
What kind of notebooks does the unslothai/notebooks repository hold?
Fine-tuning and reinforcement learning notebooks for text, vision, audio, embedding, and TTS models, organised by model in tables that cover training types such as Vision, Conversational, GRPO, Embeddings, TTS, and Alpaca.
Do the unslothai notebooks run in Colab or locally?
Both. Every notebook link in the root file opens in Colab, and the file states that the notebooks run locally and cover data prep, training and inference.
What does the unslothai/notebooks repository require in terms of Python and dependencies?
Python 3.13 or newer, with ipykernel, ipywidgets, and nbconvert as the declared dependencies. The repository also carries a .python-version file and a uv.lock.
Can I edit the notebook tables in the unslothai/notebooks README by hand?
Not inside the generated block. The root file marks that section as generated by update_all_notebooks.py, with hand editing allowed again after the `end of notebook links` marker.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/unslothai-notebooks)Star history chart for your README
[](https://hysenlabs.com/projects/unslothai-notebooks)Was this useful?