Llama-2-70b-chat-hf

Llama-2-70b-chat-hf is Meta's 70-billion-parameter dialogue-tuned large language model, released in July 2023 as part of the Llama 2 family. It is optimized for assistant-style chat through supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), and is distributed in Hugging Face Transformers format under the Llama 2 Community License. The model uses an optimized Transformer architecture with Grouped-Query Attention (GQA) for improved inference scalability, pretrained on 2 trillion tokens with a 4K context length.

On standard academic benchmarks, Llama 2 70B achieves 68.9% on MMLU (5-shot), 51.2% on BBH (3-shot), and 54.2% on AGI Eval. It scores 56.8% on GSM8K and 13.5% on MATH for mathematical reasoning, and 29.9% on HumanEval and 45.0% on MBPP for code generation. In human evaluations, Llama-2-Chat 70B was reported to be competitive with ChatGPT on helpfulness and safety metrics among open-source models at the time of release.

The model accepts text-only input and generates text output, intended primarily for English-language commercial and research applications. Meta applied safety-specific data annotation, red-teaming, and iterative evaluation during fine-tuning, with the chat variant showing substantially improved TruthfulQA scores (64.14%) and near-zero toxic generation rates compared to the base pretrained model.