llama 2

LLaMA 2 improved on LLaMA with better training, safety alignment, and open commercial licensing. **Release**: July 2023, partnership with Microsoft. **Sizes**: 7B, 13B, 70B parameters (dropped 33B). **Key improvements**: 40% more training data (2T tokens), doubled context length (4K), grouped query attention (GQA) for 70B efficiency. **Chat models**: LLaMA 2-Chat versions fine-tuned for dialogue with RLHF, safety training. **Safety work**: Red teaming, safety evaluations, responsible use guide. Most aligned open model at release. **Commercial license**: Unlike LLaMA 1, freely available for commercial use (with restrictions above 700M monthly users). **Performance**: Competitive with GPT-3.5, approaching GPT-4 at 70B on some tasks. **Ecosystem**: Foundation for countless fine-tunes, merges, and applications. Code LLaMA for programming. **Training details**: Published extensive technical report on training process and safety methodology. **Impact**: Set standard for responsible open model release, enabled commercial open-source AI applications.