On Thursday, Meta unveiled Code Llama, a new artificial intelligence-powered code-writing tool based on its Llama 2 big language model.
According to Meta, Code Llama can employ text prompts to generate new code or debug code written by humans.
“It has the potential to make workflows faster and more efficient for developers while also lowering the barrier to entry for people who are learning to code,” the business stated in a statement.
Llama 2 was trained on code-specific datasets to develop Code Llama. Code Llama can create natural language about code from both code and natural language prompts, in addition to better coding capabilities.
The new system is capable of supporting “many of today’s most popular programming languages, including Python, C++, Java, PHP, Typescript (Javascript), C#, Bash, and others.”
According to Meta, Code Llama – Python is a language-specialized variation trained on Python, “the most benchmarked language for code generation.”
Code Llama – Instruct is “an instruction fine-tuned and aligned variation of Code Llama,” according to Meta.
“Instruction tuning is a continuation of the training process, but with a different goal.” A natural language instruction input and the predicted outcome are provided into the model. This improves its grasp of what individuals expect from their prompts,” Meta explained.
“We recommend using Code Llama – Instruct variants whenever using Code Llama for code generation,” Meta added.
According to Meta, Code Llama will be offered in three sizes, the smallest of which can fit on a single GPU.
