Meta Releases Llama 70B, the AI Coding Model, Described as “Largest” and “Best-Performing” in the Llama Family

Code Llama 70B outperformed GPT 3.5 in both accuracy and benchmark scores, scoring 53% on the HumanEval benchmark and 62.4 percent on the MBPP benchmark

The most recent version of Meta’s open-source artificial intelligence (AI) coding model, Code Llama 70B, was released recently. The California-based software company described it as “the largest and best-performing model in the Code Llama family” while announcing the release. According to the company’s research, Code Llama 70B demonstrated capabilities that were close to OpenAI’s GPT 4, which received a score of 67 percent, with an accuracy score of 53 percent on the HumanEval benchmark. The company’s current coding models, Code Llama 7B, Code Llama 13B, and Code Llama 34B, are joined by the newest AI assistant.

Code Llama 70B was unveiled by Meta CEO Mark Zuckerberg on Facebook with the following message: “We’re open sourcing a new and improved Code Llama, including a larger 70B parameter model.” Currently, one of the most significant applications of AI models is the writing and editing of code. […] I’m happy with the development made here and am eager to include it into Llama 3 and next models.

According to Meta’s blog post, Code Llama 70B is offered in three versions: the basic model, Code Llama – Python, and Code Llama – Instruct. Because Instruct contains natural language processing (LNP) capabilities and Python is the designated programming language, you can use it even if you are not proficient in coding.

Read also:-IMF Updates Global Forecast for 2024; India Is Still a Bright Spot

In addition to producing codes, the Meta AI coding assistant can also provide natural language responses, which is useful for elucidating codes and responding to inquiries about them. One trillion tokens, or around 750 million words, of coding and code-related data were used to train the 70B model. Code Llama 70B is available for free for both business and research use, just like all other LLama AI models. Hugging Face, a coding repository, serves as its host.

Regarding benchmarks, the business has shared its accuracy ratings and contrasted them with those of all other competing AI models that prioritize code. 70B achieved a score of 53 percent on the HumanEval benchmark and 62.4 percent on the Mostly Basic Python Programming (MBPP) benchmark. It has outperformed OpenAI’s GPT 3.5 in both benchmarks, which scored 48.1 percent and 52.2 percent, respectively. GPT 4 only provided its HumanEval accuracy results publicly, scoring 67 percent and narrowly surpassing the Llama Code 70B criterion.

Based on the Llama 2 core model and trained exclusively on its coding-based datasets, Meta published Code Llama in August 2023. It can generate responses in both plain language and codes, and it accepts both for requests. Code Llama is able to create, modify, examine, and troubleshoot code. The Instruct version of it can also assist users in comprehending the codes in plain language.