Meta announced that Code Llama will simplify the process of completing code.
Meta debuts Code Llama, a tool developed using its Llama 2 large language model, designed to generate fresh code and assist in debugging code written by humans, according to the company’s announcement. Utilizing the identical community license as Llama 2, Code Llama is accessible for both research and commercial applications at no cost.
According to Meta‘s announcement, Code Llama has the capability to generate lines of code based on prompts, as well as to finalize and troubleshoot code when directed towards a specific code segment. Meta has also unveiled variations of the base Code Llama model, including Code Llama-Python, tailored for Python programming, and Code Llama-Instrct, designed to comprehend instructions given in natural language. Notably, Meta emphasizes that these distinct versions of Code Llama are not interchangeable, and the company advises against utilizing the base Code Llama or Code Llama-Python for interpreting natural language instructions.
Meta recently published a blog post where they emphasized that programmers are presently utilizing LLMs to support a wide array of tasks, including the creation of new software and the troubleshooting of existing code. According to Meta, the overarching goal is to amplify the productivity of developer workflows, allowing them to focus on the aspects of their work that are most centered around human interaction and creativity.
Meta asserts that Code Llama exhibited superior performance compared to publicly accessible LLMs, as confirmed through benchmark evaluations. However, the specific models against which it was evaluated were not explicitly disclosed by the company. Meta further elaborated that Code Llama achieved a noteworthy score of 53.7 percent on the HumanEval code benchmark, successfully translating textual descriptions into accurate code.
As part of its release, Meta will offer three distinct sizes of the Code Llama model. Notably, the smallest iteration is designed to operate on a single GPU, ensuring suitability for projects that require low-latency processing.
For a considerable duration, code generation tools have been assisting developers in their tasks. A case in point is GitHub’s introduction of Copilot in March, which utilizes OpenAI’s GPT-4 to expedite the process of code composition and verification. Copilot not only aids in the creation of new code but also in the revision of existing code to ensure its modernity. Likewise, Amazon’s AWS offers CodeWhisperer, a tool that serves a similar purpose by generating, validating, and updating code. Meanwhile, Google is also developing its own code-writing tool, named AlphaCode, although its release is still pending.
It’s worth noting that Microsoft, the parent company of GitHub, and OpenAI have encountered legal challenges regarding Copilot, as accusations have been made that the tool may infringe upon copyright laws due to its potential ability to replicate licensed code.
