LLaMA-Adapter

Why we included this project

For anyone working with LLaMA-family checkpoints who wants instruction-following or multimodal behavior without the cost of full fine-tuning, this is the reference implementation of the LLaMA-Adapter method. It trains only about 1.2M parameters via zero-init attention adapters, which brings the job down to roughly an hour on commodity hardware while staying competitive with fully fine-tuned baselines such as Stanford Alpaca. The repo also ships fine-tuned checkpoints, a V2 variant for image-conditioned dialog, and notes on hooking it into LangChain, so it works as both a practical starting point and a clear example of how parameter-efficient tuning compares with other PEFT approaches.