Meta’s LLaMA is one of the most popular families of large language models. Its non-commercial license and easy-to-obtain weights made it one of the most used open-source models in academia and beyond. A look at the ecosystem that’s developed:
Fine-tuned offshoots
- Replicated, but under a permissive license — RedPajama, OpenLLaMA, OpenAlpaca,
- Instruction-following model — Alpaca
- Chatbot Model — Vicuna, Koala, Orca
- More training data for different languages — Chinese-LLaMA-Alpaca
- Quantization of LLaMA — GPTQ-for-LLaMA
- Fine-tune on consumer hardware (with LoRA) — Alpaca-lora
- Training data from other LLMs — WizardLM
- “Uncensored” training data — WizardLM-Uncensored
Tools to run LLaMA
- llama.cpp — port of LLaMA in pure C/C++
- dalai — a command line tool that makes it easy to run llama locally
- chat.matt-rickard.com — WebGPU accelerated Vicuna in the browser