stabilityai/stable-code-instruct-3b
The Stable Code Instruct 3B model by Stability AI is a 2.7 billion parameter decoder-only language model, fine-tuned from Stable Code 3B. It is optimized for general-purpose code and software engineering conversations, including SQL generation. This model demonstrates strong performance on MultiPL-E metrics across multiple programming languages and on the code portions of MT Bench, making it suitable for various coding assistance tasks.
Loading preview...
Stable Code Instruct 3B: Code-Focused Instruction Model
Stable Code Instruct 3B, developed by Stability AI, is a 2.7 billion parameter decoder-only language model. It is an instruction-tuned variant of the Stable Code 3B base model, trained using a mix of publicly available and synthetic datasets, incorporating Direct Preference Optimization (DPO).
Key Capabilities
- Code Generation & Software Engineering: Excels in general-purpose code and software engineering conversations.
- SQL Generation: Specifically fine-tuned for SQL-related generation and conversational tasks.
- Multi-language Performance: Demonstrates strong performance on MultiPL-E metrics across various programming languages, including Python, C++, JavaScript, Java, PHP, and Rust.
- Instruction Following: Achieves a competitive score of 5.8 on the MT Bench coding benchmark for its size class.
Performance Highlights
Compared to models of similar size, Stable Code Instruct 3B (DPO) shows leading performance:
- MultiPL-E Average: Achieves an average score of 0.47, outperforming Codellama Instruct 7B and Deepseek Instruct 1.3B.
- MT-Bench Coding: Scores 5.8, surpassing DeepSeek Coder 1.3B and CodeLlama Instruct 7B.
Good For
- Developers seeking a compact yet capable model for code generation.
- Applications requiring conversational assistance for software engineering tasks.
- Generating and interacting with SQL queries.