Facebook's Code Llama is built on the basis of Llama 2 and has been fine-tuned specifically for generating and discussing code. It is free for both research and commercial use. :)
Don't look at the ads, look at the results (the higher the score, the better the effect):
Code Llama was previously introduced as the code-specific version of Llama 2, supporting all mainstream programming languages, including Python, C++, Java, PHP, Typescript (Javascript), C#, Bash, etc.
It was trained on 500B tokens of code and code-related data, with three sub-models and two branches:
The 7B model can run on a single GPU. The 13B model is faster and more suitable for tasks that require low latency, such as real-time code completion. The 34B model returns the best results and allows for better coding assistance. Code Llama - Python has been further fine-tuned on 100 billion Python code tokens. Code Llama - Instruct is a variant that has undergone instruction fine-tuning and alignment. Instruction tuning continues the training process but with a different objective. This model receives natural language instruction inputs and expected outputs. This makes it better at understanding what people expect from their prompts to generate useful and safe answers in natural language.
