codedebugger / openenv.yaml
psdhanushkumar's picture
Upload folder using huggingface_hub
f361447 verified
Raw
History Blame Contribute Delete
2.48 kB
# CodeDebugger — OpenEnv Registry Manifest
# Meta + Scalar OpenEnv Hackathon 2026
env:
name: CodeDebugger-v1
version: "1.0.0"
description: >
RL environment where an LLM agent sees buggy Python code and error messages,
proposes fixes step by step, and gets rewarded based on test-case pass rates.
author: CodeDebugger Team
license: MIT
interface:
type: openenv
action_space: text # agent submits fixed Python code as a string
observation_space: text # dict with buggy_code, error_message, history
reward_range: [0.0, 120.0]
max_episode_steps: 5
entry_point: codedebugger.env.codedebugger_env:CodeDebuggerEnv
api:
type: fastapi
host: "0.0.0.0"
port: 8000
endpoints:
reset: POST /reset
step: POST /step
render: GET /render
metadata: GET /metadata
problems: GET /problems
reward:
components:
- name: test_score
range: "0-50"
description: Proportional score based on test cases passed, +10 bonus for 100%
- name: fix_quality
range: "0-20"
description: Rewards minimal, targeted changes to the original code
- name: format_score
range: "0-15"
description: Valid Python syntax, preserved function names, consistent style
- name: speed_score
range: "0-10"
description: Rewards solving in fewer iterations (10/7/4 for iter 1/2/3)
- name: safety_score
range: "0-10"
description: No forbidden imports or dangerous patterns detected
- name: improvement_score
range: "0-15"
description: Rewards passing more tests than previous iteration
- name: anti_hack_penalty
range: "negative"
description: Penalty for hardcoded returns, code gutting, function deletion
- name: process_bonus
range: "0-3"
description: Bonus for fix addressing the root cause (keyword heuristic)
anti_hacking:
checks:
- hardcoded_returns
- function_deletion
- trivial_pass
- code_gutting
penalty_factor: 0.1
dataset:
total_problems: 30
difficulties: [easy, medium, hard]
problems_per_difficulty: 10
test_cases_per_problem: 4
source: codedebugger.data.bug_dataset
training:
method: GRPO
framework: trl
optimizer: unsloth
default_model: unsloth/Llama-3.2-1B-Instruct
deployment:
platform: huggingface_spaces
docker: true
streamlit_port: 7860