Spaces:
Paused
Paused
| title: DebugArena | |
| emoji: 🐛 | |
| colorFrom: blue | |
| colorTo: green | |
| sdk: docker | |
| pinned: false | |
| license: apache-2.0 | |
| tags: | |
| - reinforcement-learning | |
| - openenv | |
| - code-debugging | |
| - grpo | |
| # DebugArena 🐛 | |
| *Meta PyTorch OpenEnv Hackathon × Scaler School of Technology 2026 | Theme 4: Self-Improving Agents* | |
| DebugArena is an RL environment where an agent learns to fix buggy Python code through trial and error. Fine-tuned on **Qwen2.5-Coder-0.5B** using GRPO, the model improved its solve rate from 35% to 45% entirely through test-failure reward signals. | |
| ### 📖 Read the Full Story | |
| The organizers requested that the main project write-up be kept separate from the README. | |
| **➡️ [Click here to read the full project breakdown in Blog.MD](./Blog.MD)** *(The blog covers the problem statement, the GRPO training pipeline, the custom reward functions, and how we overcame 'Reward Hacking' during training).* | |
| ### 🛠️ Quick Links | |
| * **Training Notebook:** [View the code on Kaggle](https://www.kaggle.com/code/bharathvikas/debugarena) | |
| * **Environment Code:** Check the `server/environment.py` file in this repository. |