250+ notebooks that turn Colab into a fine-tuning factory

What it does This is a massive index of Google Colab notebooks maintained by the Unsloth team. Each notebook bundles data preparation, training, and inference for a specific model—no local GPU setup required. The collection spans text, vision, audio, embeddings, and text-to-speech, with a heavy emphasis on recent open-weight models like Gemma 4, Qwen3.5, and gpt-oss 20B.

The interesting bit The README is auto-generated by a script (update_all_notebooks.py), which suggests the team treats this as a living distribution channel rather than a static gallery. The GRPO and reinforcement-learning notebooks are the less obvious gems—Sudoku solvers, 2048 game training, auto kernel creation—tasks that go well beyond standard instruction tuning.

Key highlights

• One-click Colab badges for every notebook; runs locally or in cloud
• Covers modalities often ignored by fine-tuning guides: audio (Gemma3N, Gemma4), TTS (Orpheus), embeddings (Qwen3-Embedding, EmbeddingGemma)
• GRPO/RL section includes game-playing and math reasoning environments, not just chat tuning
• Includes Unsloth Studio, a chat UI notebook for quick inference
• Docs link out to a broader fine-tuning guide for context

Caveats

• The repo is essentially a link farm; all compute happens in Colab or your own hardware
• README warns against manual editing of the notebook table—automation is tight, human curation less visible
• No stated compatibility matrix for Colab free tier vs. paid; some 20B+ models likely need Pro

Verdict Great if you want to experiment with modern fine-tuning and RL without wrestling with dependencies. Skip it if you need a library or framework—this is documentation and recipes, not code you import.