BurnyCoder/qwen2.5-coder-0.5b-swe-rl
BurnyCoder/qwen2.5-coder-0.5b-swe-rl is a 0.5 billion parameter Qwen2.5-Coder-Instruct model fine-tuned by BurnyCoder using Reinforcement Learning (RL) with GRPO. This model is specifically optimized for single-function Python code repair, focusing on fixing bugs rather than repository-scale software engineering. It features a 32768 token context length and demonstrates improved greedy pass@1 performance on internal and external code repair benchmarks.
Loading preview...
Model Overview
BurnyCoder/qwen2.5-coder-0.5b-swe-rl is a 0.5 billion parameter language model derived from Qwen/Qwen2.5-Coder-0.5B-Instruct. It has been fine-tuned by BurnyCoder using a Reinforcement Learning approach, specifically GRPO (Generalized Reinforcement Learning with Policy Optimization) with LoRA, to specialize in code repair tasks.
Key Capabilities
- Single-Function Python Repair: The model is designed to identify and fix bugs within individual Python functions, making it suitable for targeted code correction.
- RL-Enhanced Performance: Through GRPO training, the model shows an improvement in its ability to generate correct code repairs, as evidenced by an increase in greedy pass@1 metrics on both internal validation splits and the HumanEvalFix benchmark.
- Context Length: It supports a substantial context window of 32768 tokens, allowing it to process larger code snippets for repair.
Performance Highlights
On an internal 12-task validation split, the model achieved a greedy pass@1 of 0.5000 compared to the base model's 0.4167. For external evaluation on 164 Python repairs from HumanEvalFix, it showed a greedy pass@1 of 0.2317, slightly outperforming the base model's 0.2256. These results indicate its specialized effectiveness in code repair, though the benchmarks are not directly comparable to standard leaderboard scores due to specific sandbox environments and time limits.
Good For
- Automated Bug Fixing: Developers looking for a compact model to assist with automated bug detection and correction in Python functions.
- Code Repair Research: As a specialized RL-tuned model, it serves as a valuable resource for research into reinforcement learning applications for software engineering tasks.
- Resource-Constrained Environments: Its 0.5B parameter size makes it efficient for deployment in environments with limited computational resources, while still offering specialized code repair capabilities.