Meta announced its new version of large language model (LLM), Llama 2, on July 18, 2023, which is fully open source for both research and commercial use. Different from its previous version (Llama 1) only for research use, Llama 2 provides access for companies and people in tech, academia, and policy fields. As compared to Llama 1, more than 40% more data (2 trillion tokens) with double length of context (4096) were used for the training of Llama 2. Llama 2 provides three different sizes of models, including 7B, 13B, and 70B parameter models. Besides the foundation model, Llama 2 provides a fine-tuned model for chat use cases, with supervised fine-tuning and human preferences being considered for helpfulness and safety during the data collection. Llama 2 outperforms different LLM models in different benchmarks in reasoning, coding, proficiency, and knowledge tests. Meta especially partnered with Microsoft for enabling developers to use Llama 2 on Microsoft Azure. Llama 2 is also available through Amazon Web Services (AWS), Hugging Face, and other providers. Llama 2 provides everyone with accessibility to LLM, which everyone could dig deeper and fine tune the model for our own use.
Resources:
Llama 2 Website: https://ai.meta.com/llama/
Llama 2 Paper: https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/
Caisky Consulting 