Code Llama (neural network for writing program code)

2024: Product Announcement

At the end of January 2024, Meta (recognized as an extremist organization; activities in the Russian Federation are prohibited) presented a model of artificial intelligence Code Llama 70B for generating program code. It is claimed that at the time of the announcement it is the largest and most effective neural network of this type.

Code Llama tools are free for both research and commercial use. The neural network not only helps developers create programs by generating code based on prompts, but also debugs the work of blocks written by a person. The model is trained on 1 TB of code and associated data.

Pexels

According to Meta, Code Llama 70B can handle more requests than previous versions, meaning developers can enter more detailed instructions and get more accurate and effective results. In the HumanEval test, the Code Llama 70B model showed 53% accuracy. For comparison: in the case of OpenAI GPT-3.5 this figure is 48.1%, and in GPT-4 it is 67%.

Developers also have access to specialized Code Llama tools: Code Llama - Python and Code Llama - Instruct, which are focused on specific programming languages. In particular, the CodeLlama-70B-Python model was trained on an additional 100 billion Python code tokens. In turn, the CodeLlama-70B-Instruct can solve various auxiliary problems, such as sorting, searching, filtering and manipulating data, as well as implementing algorithms. The Code Llama 70B model is located in the Hugging Face code repository. Companies can use a code generation system in their private environment. It gives them control and confidence to protect intellectual property.^[1][2]

Notes