A notebook buffet for fine-tuning Llama, Falcon, and friends

What it does

This repository is a curated cookbook of Jupyter notebooks for parameter-efficient fine-tuning. It bundles around twenty ready-to-run recipes—each backed by a Colab badge—showing how to adapt large language models with LoRA, QLoRA, and the Hugging Face PEFT library. Targets range from small BLOOM and Phi variants up to 30B-parameter instruction models and GPT-NeoX-20B.

The interesting bit

Instead of shipping a new framework, the project acts as a broad-spectrum reference. It covers not just standard fine-tuning for Llama 2, Falcon, and Mistral, but also RLHF, RAG pipelines with LangChain, knowledge-graph construction with Neo4j, and 1-bit quantization experiments. For developers who want to compare techniques across architectures without writing boilerplate, it is a pragmatic—if uneven—shortcut.

Key highlights

• Every notebook carries a one-click Colab badge, so you can experiment without a local GPU.
• Covers diverse architectures: LLaMA, Llama 2, Falcon-7b, Mistral-7b, GPT-NeoX-20B, MPT-30B, Microsoft Phi-1.5, and Gemma2-9b.
• Techniques range from standard LoRA and QLoRA to RLHF, 8-bit/4-bit quantization, and 1-bit training via the HQQ framework.
• Includes adjacent LLM workflows: LangChain RAG tutorials, Neo4j knowledge-graph extraction, and OpenAI GPT-3.5 Turbo fine-tuning.
• Recent additions use Unsloth for faster Gemma2 fine-tuning.

Caveats

• README descriptions are often generic (“Details and code for…”), so you may need to open a notebook to gauge its depth.
• This is a tutorial anthology, not a maintained library or unified API; expect to manage dependencies and environment drift yourself.

Verdict

Worth bookmarking if you want a broad survey of fine-tuning patterns across many models. Skip it if you need a production framework or deep theoretical explanations.