Llama 2 Chat 13B is a generative text model developed by Meta, specifically optimized for dialogue and conversational use cases. As the mid-sized variant in the Llama 2 family, it provides a balance between computational efficiency and reasoning performance. The model utilizes an auto-regressive transformer architecture and supports a context window of 4,096 tokens.
The model underwent a rigorous training process, beginning with pre-training on 2 trillion tokens from publicly available online sources. To refine its performance for interactive applications, Meta applied Supervised Fine-Tuning (SFT) followed by Reinforcement Learning from Human Feedback (RLHF). These alignment stages were designed to improve both the helpfulness and safety of the model's responses during multi-turn conversations.
Meta released the model under a custom permissive license that allows for research and commercial use, subject to certain scale-based restrictions. The chat-optimized versions were tested across multiple benchmarks and human evaluations to ensure they met specific safety and utility standards, making them suitable for virtual assistants and complex text generation tasks.