unslothai/notebooks

250+ Fine-tuning & RL Notebooks for text, vision, audio, embedding, TTS models.

unslothai/notebooks – Ready‑to‑run Colab notebooks for LLM fine‑tuning, inference and experiments

What it is – A curated collection of Jupyter notebooks (hosted on Google Colab) that demonstrate how to use the Unsloth stack with a wide variety of open‑source large language models (LLMs). The notebooks cover:

  • Fine‑tuning (e.g., GPT‑OSS 20 B, Llama 3.1, Gemma 4) – data preparation, training loops and saving the tuned checkpoint.
  • Inference / chat UIs – the Unsloth Studio notebook gives a simple web‑chat front‑end.
  • Multimodal – vision (Gemma 4, Qwen 3‑VL, Mistral Ministral 3), audio (Gemma 3N, Orpheus‑TTS) and embeddings (EmbeddingGemma, Qwen3‑Embedding).
  • Reinforcement‑learning / GRPO / ORPO – notebooks that turn LLMs into game‑playing agents (Sudoku, 2048, Minesweeper) or run reward‑model training (GRPO, DPO, ORPO) on tasks like GSM8K math.
  • Hardware‑specific demos – BF16, FP8, A100‑specific kernels, DGX‑Spark examples.

All notebooks are self‑contained: they install the required unsloth Python package, download the model weights, run the training or inference code, and display results – all inside a free Colab session.

How to use

  1. Click the Open in Colab badge for the notebook you need (e.g., gpt‑oss (20B) – Fine‑tuning). The notebook will open in a new Colab tab.
  2. Run the first cells to install unsloth and any optional dependencies (torch, transformers, etc.).
  3. Follow the step‑by‑step sections:
    • Data prep – load a dataset (often from HuggingFace) or upload your own CSV/JSON.
    • Training – a single trainer.fit() call that leverages Unsloth’s fast‑kernel optimisations.
    • Evaluation / inference – generate text, run a vision model on images, or launch the chat UI.
  4. When finished, you can download the fine‑tuned checkpoint or export a Gradio‑style demo.

What makes it useful

Feature Why it matters
Broad model coverage Shows Unsloth working with the latest open‑source LLM families (Gemma, Qwen, Llama 3.x, Mistral, Phi‑4, etc.).
Multimodal examples Vision, audio and embedding notebooks let users see how to plug non‑text modalities into the same training pipeline.
RL / GRPO notebooks Demonstrates using LLMs as agents for games and reward‑model training, a cutting‑edge research direction.
Hardware‑specific kernels BF16, FP8 and A100‑optimised notebooks illustrate how to squeeze extra performance out of the same code base.
Zero‑setup on Colab No local GPU required; anyone with a Google account can try the full workflow instantly.

Who should look at this

  • Researchers wanting quick reproducible scripts for fine‑tuning or RL‑based LLM training.
  • Practitioners who need a starter kit to adapt a public model to their own data without writing boiler‑plate code.
  • Students learning about modern LLM training tricks (PEFT, quantisation, GRPO/DPO) and who prefer an interactive notebook environment.

Where to find more

TL;DR – unslothai/notebooks is a genuine, actively maintained repository that supplies ready‑to‑run Colab notebooks for fine‑tuning, inference and RL experiments across dozens of open‑source LLMs, using the high‑performance Unsloth library. It’s a practical learning and prototyping resource rather than a code library itself.

Related

  • Project
    unslothai/unsloth

    A desktop app and framework for running and training AI models locally, offering significantly faster fine-tuning with reduced VRAM usage across multiple hardware platforms.

  • Dispatch
    Unsloth and Hugging Face TRL Integration for Faster LLM Fine-tuning

    Unsloth is a lightweight library that accelerates LLM fine-tuning by up to 2.7x and reduces memory usage by up to 74% with 0% accuracy degradation compared to QLoRA.

  • Project
    unslothai/unsloth-zoo

    A utility library for Unsloth that enables faster and more memory-efficient fine-tuning of LLMs, vision, and TTS models with significantly reduced VRAM requirements.

  • Project
    hollowstrawberry/kohya-colab

    A collection of Google Colab notebooks for preparing datasets and training Stable Diffusion LoRA models, including support for SDXL.

  • Dispatch
    Unsloth Dynamic 3.0 GGUF Release

    Unsloth has released Dynamic 3.0 GGUFs, featuring improved quantization methodology that delivers over 10% better top-1% accuracy for Qwen3.8-27B compared to other providers at the same model size.