m-a-p/OpenCodeInterpreter-CL-34B
OpenCodeInterpreter-CL-34B is a 34 billion parameter code generation system developed by m-a-p, based on CodeLlama-34b-Python-hf, with a 32768 token context length. It integrates execution and iterative refinement functionalities to enhance code generation. This model is specifically designed to bridge the gap between large language models and advanced proprietary systems like GPT-4 Code Interpreter, excelling in code interpretation and execution tasks through execution feedback.
Loading preview...
OpenCodeInterpreter-CL-34B: Enhanced Code Generation with Execution Feedback
OpenCodeInterpreter-CL-34B is a 34 billion parameter model built upon CodeLlama-34b-Python-hf, designed to significantly advance open-source code generation capabilities. Its core innovation lies in integrating execution and iterative refinement functionalities, allowing the model to generate and then improve code based on execution feedback.
Key Capabilities & Features
- Execution Feedback Integration: The model leverages execution results to iteratively refine generated code, a key differentiator from traditional code generation models.
- High Performance on Code Benchmarks: This model demonstrates strong performance on coding benchmarks like HumanEval and MBPP, with scores significantly improving after incorporating execution feedback. For instance, on HumanEval, it achieves 81.7% with execution feedback (up from 78.0%), and on MBPP, it reaches 80.2% (up from 73.4%).
- Large Context Window: Supports a context length of 32768 tokens, enabling it to handle complex and extensive codebases.
Good for
- Advanced Code Generation: Ideal for tasks requiring not just code generation but also self-correction and refinement based on execution outcomes.
- Bridging the Gap with Proprietary Models: Aims to provide open-source alternatives to advanced proprietary systems like GPT-4 Code Interpreter.
- Research and Development: Useful for researchers exploring iterative code generation and execution feedback mechanisms.