Meta’s upgraded version of Code Llama AI programming tool surpasses GPT-3.5 and leads in accuracy

News from this site on January 30, Zuckerberg announced that Meta recently updated its code generation AI model Code Llama 70B, claiming that this is the best-performing model to date.

The Code Llama tool was released last August and has always been free for everyone to use. The new version of Code Llama 70B now launched is even better than before in terms of query processing. Its accuracy in the HumanEval benchmark reached 53%, which is even better than GPT-3.5's 48.1%. However, there is still a certain gap compared to the GPT-4 accuracy (67%) mentioned in the OpenAI paper.

The Meta model released last year not only launched Code Llama, but also two tools optimized for specific programming languages, namely Code Llama - Python and Code Llama - Instruct. Just like their literal meaning, they are used in the development and optimization of specific programming languages. Additionally, Code Llama 70B also comes in three versions and is still free for research and commercial use.

This large model was trained on 1TB of code and code-related data, hosted on the code repository Hugging Face. According to Meta, the larger models 34B and 70B provide the best results and provide better coding assistance.

The above is the detailed content of Meta’s upgraded version of Code Llama AI programming tool surpasses GPT-3.5 and leads in accuracy. For more information, please follow other related articles on the PHP Chinese website!