CoolFace
Apppublic

AMD21/codefixerenv

sourceHugging Faceupdated 6mo agoView on Hugging Face
0likes
App README

CodeFixerEnv 🔧

An OpenEnv reinforcement learning environment for Python code-debugging tasks.

Built for the Meta × PyTorch × Hugging Face OpenEnv Hackathon — Round 1.

What is CodeFixerEnv?

CodeFixerEnv is a reinforcement learning environment that teaches an AI agent how to debug Python code. The agent receives a buggy function, reads a description of what the function should do, and submits corrected code. The environment runs the code against a deterministic test suite, grades the submission on multiple weighted criteria, and returns a dense reward signal that guides the agent toward a correct and efficient solution.

Architecture

codefixer-final/
├── env/
│   ├── models.py         ← Pydantic: Observation, Action, Reward, StepResult
│   └── environment.py    ← CodeFixerEnv class (reset / step / state)
├── tasks/
│   └── tasks.py          ← 3 tasks + deterministic graders
├── tests/
│   └── test_env.py       ← 20 unit tests
├── server.py             ← FastAPI REST server (port 7860)
├── inference.py          ← Baseline agent using OpenAI client
├── openenv.yaml          ← OpenEnv spec metadata
├── Dockerfile
└── requirements.txt

Observation Space

FieldTypeDescription
inputstrThe buggy Python code the agent must fix
contextstrPlain-English description of what the correct code should do
historylist[dict]Previous action/feedback pairs this episode
step_numberintSteps elapsed since reset
max_stepsintEpisode length cap (default 5)

Action Space

`type`Effect
"fix"Submit corrected code — graded immediately
"explain"Request a hint — costs −0.05 reward
"give_up"End episode early — costs −0.10 reward

Reward Space

Dense rewards — meaningful signal every step, not just at the end.

fix action:
  reward = grader_score − previous_best   (improvement delta)
  +0.20 efficiency bonus if solved in ≤ 2 steps
  −0.05 if no improvement (redundancy penalty)

explain: −0.05
give_up: −0.10

Tasks

🟢 Easy — sum_range(n)

Fix an off-by-one error. range(n) → range(1, n+1). Grading: syntax 20% + correctness 70% + no hardcoding 10%

🟡 Medium — flatten(lst)

Fix two bugs in a nested list flattener. Grading: syntax 15% + correctness 75% + isinstance usage 10%

🔴 Hard — longest_unique(s)

Rewrite O(n²) brute-force as O(n) sliding window. Grading: syntax 10% + correctness 70% + linear complexity 20%

API Endpoints

MethodPathDescription
GET/healthLiveness check
POST/resetStart new episode {"difficulty": "easy"}
POST/stepSubmit action {"action_type": "fix", "action_content": "..."}
GET/stateInternal env state
GET/tasksList all tasks
GET/docsSwagger UI

Setup

bash
pip install -r requirements.txt
python server.py
# Server at http://localhost:7860

Docker

bash
docker build -t codefixerenv .
docker run -p 7860:7860 codefixerenv

Running the Baseline Agent

bash
export API_BASE_URL=http://localhost:7860
export MODEL_NAME=Qwen/Qwen2.5-72B-Instruct
export HF_TOKEN=hf_...
python inference.py

Running Tests

bash
python -m pytest tests/test_env.py -v
# Expected: 20 passed