open_llama

Why we included this project

OpenLLaMA is a permissive open source reimplementation of Meta's LLaMA, trained on the public RedPajama corpus and later mixtures. The project ships 3B, 7B, and 13B checkpoints trained on roughly a trillion tokens, with weights in both PyTorch and JAX formats. Because the weights are meant to drop into existing LLaMA pipelines, teams can fine-tune or serve these models without Meta's original license terms. You also get Hugging Face transformers and EasyLM checkpoints, plus evaluation notes for comparing quality to the original. It's a solid base model to build on, not a turnkey application.