source

Llama 2 is an open-source large language model developed by Meta. It is available for free for both research and commercial use. This next-generation model includes pretrained and fine-tuned language models that range from 7B to 70B parameters. The pretrained models were trained on 2 trillion tokens and offer double the context length compared to Llama 1. The fine-tuned models have been trained on over 1 million human annotations.

Llama 2 outperforms other open-source language models in various external benchmarks, including reasoning, coding, proficiency, and knowledge tests. The model leverages publicly available online data sources for pretraining and publicly available instruction datasets, along with over 1 million human annotations, for fine-tuning.

Meta has established partnerships with cloud providers, companies, and researchers who support their open approach. They aim to create a responsible and collaborative AI innovation ecosystem. They provide resources such as a Responsible Use Guide for developers, safety red-teaming to enhance performance and safety, an Open Innovation AI Research Community for academic researchers, and the Llama Impact Challenge to encourage using Llama 2 to address environmental and educational challenges.

Meta also maintains a Generative AI Community Forum, in consultation with Stanford Deliberative Democracy Lab and the Behavioral Insights Team, to involve the community in decision-making around generative AI technologies.

To access Llama 2, users can complete a download form, agreeing to Meta’s privacy policy. Additional resources, including a technical overview, research paper, and blog post, provide more information on Llama 2 and its responsible use.