LLaMA2

Llama2 is the next generation open-source language model developed by Meta AI based on the previous Llama model. It has been pretrained on 2 trillion tokens which is double that of Llama 1 and comes in multiple sizes – 7 billion, 13 billion and 70 billion parameters.

The 70B parameter Llama2 model shows state-of-the-art performance exceeding the previous best models like GPT-3 and Jurassic-1 Jumbo on benchmarks.

On specific tasks like reading comprehension, math and reasoning, Llama2 outperforms even commercial models.

• There are also fine-tuned versions focused on chat and coding tasks like Llama Chat and Code Llama.
• These leverage additional training with over 1 million human annotations to align better with human preferences.