We present to your attention Stable Code Instruct 3B

Stable Code Instruct 3B is Stability AI's latest customizable-by-instruction large language model built on top of Stable Code 3B. This model improves code completion and supports natural language interaction, aiming to improve the efficiency and intuitiveness of programming and software development tasks. Our analysis shows that Stable Code Instruct 3B outperforms comparable models such as Codellama 7B Instruct and DeepSeek-Coder Instruct 1.3B in a variety of coding-related tasks. You can read more about these results in our technical report. The model is now commercially available with Stability AI Membership. Weights and code for Stable Code Instruct 3B are now available on Hugging Face.

Opportunities

Based on the results of the Stack Overflow 2023 developer survey, Stable Code Instruct 3B focuses on languages such as Python, Javascript, Java, Java, C, C++, and Go, which are some of the most popular and meaningful languages for a variety of developers. While these languages have been chosen as the primary focus for training, the model also includes training in additional widely used languages such as SQL, PHP, and Rust, as detailed in our technical report.

Stable Code Instruct 3B demonstrates high test scores even in languages not originally included in the training set, such as Lua. This proficiency may be due to an understanding of coding principles and the ability to adapt these concepts in different programming environments, capitalizing on the inherent predictability of coding tasks.

Stable Code Instructor 3B can not only generate code, but also perform FIM (Fill in the Middle) tasks, database queries, translation, explanation, and code creation. Instructional customization allows him/her to understand and act on nuanced instructions, facilitating a wide range of coding tasks beyond simple code completion, including significant abilities in mathematical understanding, logical reasoning, and processing complex technical narratives associated with software development.

Performance

When directly compared to leading models such as the Codellama 7B Instruct and DeepSeek-Coder Instruct 1.3B, the Stable Code Instruct 3B demonstrates superior performance on a range of coding tasks. Our testing, described in the technical report, shows that Stable Code Instruct 3B not only matches but often surpasses these models in code completion accuracy, understanding of natural language instructions, and versatility across programming languages.

In the graphs above, we compare all three models using the Multi-PL benchmark to evaluate the strength of the results obtained in different languages.As a result,we find that Stable Code Instruct 3B significantly outperforms CodeLlama in all languages despite having fewer parameters.

Technical details

Key technical aspects of Stable Code Instruct 3B include:

Language Support: While the focus is on the major programming languages such as Python, C++, and JavaScript, the model's training dataset covers a wide range of languages. A full list of languages in which the model has been trained is provided in our technical report.

Instruction following: This model's instruction customization allows it to interpret and execute complex and nuanced queries. This capability allows Stable Code Instruct 3B to provide more relevant and context-sensitive code suggestions, increasing developer productivity and allowing them to iterate and improve results.

Hardware Accessibility: Stable Code Instruct 3B's compact size and low hardware requirements make it accessible to a wide audience. It supports code completion, FIM tasks and understands subtle program requests, allowing developers to work more efficiently and effectively with less effort.