Shulou(Shulou.com)11/24 Report--

Thanks to CTOnews.com netizens and soft media users for 1520111 clue delivery! CTOnews.com August 24, Meta has released an AI coding tool called Code Llama, which is called "the most advanced large coding language model".

This model is based on the large language model of Llama 2, which can be understood as "the special version of Llama 2 for writing code". It can be used to "generate new code and debug the work written by human beings". It is now available on Github.

It is reported that Code Llama will use the same community license as Llama 2 and will be free for research and business.

Figure source PexelsMeta points out that Code Llama can generate the code you want and the natural language of the code according to the code core natural language hints / requirements, or refine and debug when pointing to specific code.

In addition to the basic version of Code Llama, Meta also offers a version designed specifically for Python, Code Llama-Python, and a version that understands natural language instructions, Code Llama-Instrct.

In the Meta benchmark, Code Llama outperformed the state-of-the-art publicly available LLM model on programming tasks.

According to Meta, each particular version of Code Llama is not switchable / replaceable, and the company does not recommend using basic Code Llama or Code Llama-Python for natural language instructions.

Meta has released three sizes of Code Llama with 7B, 13B and 34B parameters, respectively. Each model is trained with 500B code tokens and code-related data, and the 7B and 13B base models and instruction models are also trained with the FIM function, allowing them to insert code into existing code, which means you can use it now for the Code completion Plan.

CTOnews.com note that these three models have their own advantages and disadvantages, such as the 7B model can run on a single GPU, the 34B model provides the best results and better coding assistance, and the 7B and 13B models are faster than the 34B model and more suitable for tasks requiring low latency.

"programmers are already using LLMs to help with a variety of tasks, from writing new software to debugging existing code," Meta said in his blog. "the goal is to make developers' workflows more efficient so that they can focus on the most people-oriented aspects of their work."

Meta claims that Code Llama outperforms the existing publicly available LLMs in benchmarking, but does not specify which models it is testing against.

The company said that Code Llama scored 53.7% in the HumanEval code benchmark and was able to write code accurately according to the text description, and 56.2% on MBPP, on a par with the most advanced ChatGPT.

It's worth mentioning that GitHub launched GPT-4-based Copliot in March to help users write and review code quickly, rewrite old code to update it, but face lawsuits for alleged copyright violations.

In addition, Amazon AWS also has a CodeWhisperer that can write, review and update code, while Google also has a coding tool in AlphaCode that has not yet been released to the public.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.