Hugging Face Overview
OpenCodeInterpreter-CL-13B: Enhanced Code Generation with Execution Feedback
OpenCodeInterpreter-CL-13B is a 13 billion parameter model from the OpenCodeInterpreter family, developed by m-a-p. It is based on CodeLlama-13b-Python-hf and is specifically designed to improve code generation by incorporating execution and iterative refinement capabilities, similar to advanced proprietary systems like GPT-4 Code Interpreter.
Key Capabilities & Features
- Execution Feedback Integration: The model leverages execution feedback to iteratively refine generated code, leading to higher accuracy and robustness.
- Iterative Refinement: It can improve its code outputs through multiple iterations, addressing errors and optimizing solutions based on execution results.
- Strong Benchmark Performance: On the HumanEval benchmark, OpenCodeInterpreter-CL-13B achieves 77.4% (73.8% extended) without feedback, improving to 81.1% (76.8% extended) with one iteration of execution feedback. For MBPP, it scores 70.7% (59.2% extended) without feedback, rising to 78.2% (67.2% extended) with feedback.
- CodeLlama Base: Built on the robust CodeLlama-13b-Python-hf architecture, providing a strong foundation for Python code generation.
Good For
- Automated Code Generation: Generating functional and correct code snippets or programs.
- Code Debugging and Refinement: Assisting in identifying and correcting errors in generated code through an iterative process.
- Benchmarking Code LLMs: Serving as a strong baseline or comparison point for models focusing on code interpretation and execution tasks.
- Developing Code Interpreter Systems: Providing a foundation for building more sophisticated code generation and execution environments.