// the find
ashishpatel26/LLM-Finetuning
LLM Finetuning with peft
A collection of 23 Jupyter notebooks covering LLM fine-tuning techniques — LoRA, QLoRA, RLHF, and quantization — across models like LLaMA 2, Falcon, Mistral, and Gemma. Aimed at ML practitioners who want working Colab examples they can run immediately without setting up local GPU infrastructure. It has drifted beyond fine-tuning into RAG and knowledge graphs, which is either a feature or a scope problem depending on what you came for.
Colab badges on every notebook mean zero setup friction — you can be running a fine-tuning job in minutes on a free GPU. The coverage is genuinely broad: QLoRA for memory-constrained setups, RLHF with TRL, 1-bit quantization via HQQ, and Unsloth's 2x speedup are all represented. Recent additions (Gemma 2, CAG, MLflow evaluation) show the repo is still being updated as the landscape shifts. Useful as a quick-start reference when you need to see how a specific technique wires together before reading the paper.
Every notebook is a standalone script with no shared utilities — if you want to adapt multiple techniques, you're copy-pasting boilerplate between files. Many notebooks are pinned to model versions (LLaMA 2, Falcon-7b) that are now two generations behind; the fine-tuning code will likely still work but the models themselves aren't worth training on anymore. There's no explanation of when to choose one technique over another — QLoRA vs full LoRA vs RLHF is left as an exercise. The RAG and knowledge graph notebooks feel bolted on and would be better in their own repo.