It is reported that the Llama 2 model currently supports multiple parameter scale versions such as 7 billion, 13 billion, and 70 billion. Compared to the first generation of Llama, Llama 2 has been trained with 2 trillion tokens, and its context training is twice as long as Llama’s, reaching 4,096. In addition, the Llama-2-chat model has been trained with more than 1 million manual annotations. Due to its strong parameter scale and performance, it is considered the most powerful model to challenge the “GPT-4”.
According to Meta, Llama 2 has 40% more training material than Llama, including publicly available online sources, and outperforms large language models such as Falcon and MPT in reasoning, editing programs, and knowledge testing.
In a press release, Meta explained that its decision to open source Llama 2 was made in the hopes of giving companies, startup teams, and researchers access to more AI tools for experimentation and trial.
“We believe an open approach is the right decision for today’s AI model development, especially in the field of generative AI where the technology is rapidly evolving,” Meta noted, adding that it means developers and researchers will be able to test and quickly identify and solve development problems.
While opening up Llama 2, Meta also wanted to increase security and transparency…
