Llama-2-7b-hf

Llama-2-7B is a foundational large language model developed by Meta as part of the Llama 2 family, designed for general-purpose text generation tasks. It is a 7 billion parameter auto-regressive transformer trained on 2 trillion tokens from publicly available sources, using an optimized architecture without Grouped-Query Attention (GQA). This model is the pretrained version, intended for research and commercial use in English, and can be adapted for downstream applications such as summarization, question answering, and code generation. Llama-2-7B outputs text only, with input also restricted to text. Fine-tuned variants of this model, such as Llama-2-Chat, demonstrate competitive performance in dialogue tasks and safety benchmarks, rivaling proprietary models. The model was trained using Meta’s infrastructure, including 184,320 GPU hours and emissions fully offset by Meta’s sustainability program.