Code Llama from Meta is close to GPT-4 in code generation accuracy

Dima Solomin/Unsplash

Meta has updated its large language model for code generation Code Llama 70B. According to the developer, its accuracy is now close to OpenAI's GPT-4.

Here's What We Know

According to Meta, the updated Code Llama 70B can handle more queries and produce more accurate results. It scored 53% on the accuracy metric in the HumanEval test. In comparison, GPT-3.5 has a score of 48.1 per cent and GPT-4 has a score of 67 per cent.

Meta's Code Llama helps programmers generate code snippets based on hints and debug existing programmes. The 70B model was trained on 1 TB of programming-related data. It is available for free for research and business.

Meta previously introduced language-specific tools Code Llama Python and Code Llama Instruct. Now the company says its AI models are showing "the best results in helping developers."

Source: The Verge