Introducing Code Llama: Meta’s new AI Coding Tool

Meta has joined the AI race by unveiling “Code Llama”, a state-of-the-art large language model (LLM) that is ready to change the way code is written, understood and debugged. This is a direct challenge to ChatGPT, OpenAI’s star-AI model.

Today we’re releasing Code Llama, a large language model built on top of Llama 2, fine-tuned for coding & state-of-the-art for publicly available coding tools.

Keeping with our open approach, Code Llama is publicly-available now for both research & commercial use.

More ⬇️

— Meta AI (@MetaAI) August 24, 2023

How does Code Llama work?

As Meta explained in their blog post, Code Llama is a code-specialized version of Llama 2 that features enhanced coding capabilities can generate code, and natural language about code, from both code and natural language prompts. (eg: Write me a function that outputs the fibonacci sequence). This will undoubtedly be very helpful in a variety of tasks, ranging from writing new software to debugging existing code.

Meta’s Code Llama offers its users three distinct models

• The foundational code model.
• Codel Llama, a version explicitly fine-tuned for Python programming.
• Code Llama – Instruct, designed to handle natural language instructions more effectively.

Code Llama is available in three different sizes with 7B, 13B, and 34B parameters, catering to various serving and latency requirements of its users.

The LLM model stands out for its coding capabilities and has been specifically trained to utilize popular programming languages like Python, C++, Java, PHP, etc.

Evaluation:

Meta sought to evaluate their new AI tool’s performance and used two popular benchmarks for the same: HumanEval and Mostly Basic Python Programming (MBPP).

Their results were positive and promising. Their benchmark showed that Code Llama performed better than open-source, code-specific LLMs and outperformed Llama 2.

Meta also ran a quantitative evaluation of Code Llama’s risk of generating malicious code, and their results our results found that Code Llama answered with safer responses compared to ChatGPT.

Responsibility:

Meta has updated their Responsible Use Guide regarding the same, which includes guidance on developing models, including:

• Defining content policies and mitigations.
• Preparing data.
• Fine-tuning the model.
• Evaluating and improving performance.
• Addressing input- and output-level risks.
• Building transparency and reporting mechanisms in user interactions.

Meta has announced that Code Llama is currently available for free.

More details are available on Meta’s helpful blog post and on their research paper.