debug-env / openenv.yaml
hugsbun's picture
modified: inference.py
5bbf4f7
Raw
History Blame Contribute Delete
4.11 kB
spec_version: 1
name: debug_env
type: space
runtime: fastapi
app: debug_env.server.app:app
port: 7860
description: >
A real-world debugging environment where LLM agents fix broken Python code.
Tasks range from simple syntax errors to complex logic bugs. Agents use
read_file, edit_file, list_files, and run_tests tools to identify and fix issues.
Rewards are based on test pass rate (0.0-1.0).
author: Meta Platforms, Inc.
version: 0.1.0
tasks:
- id: task1
title: "Fix Simple Syntax Error"
description: >
A Python module with a syntax error in the function definition.
Agent must identify the missing colon and fix the function signature.
difficulty: easy
estimated_steps: 3
- id: task2
title: "Fix Logic Bug in Conditional"
description: >
A function with incorrect conditional logic that fails test assertions.
Agent must read tests, understand the expected behavior, and fix the logic.
difficulty: medium
estimated_steps: 5
- id: task3
title: "Fix Import and Type Issues"
description: >
Multiple functions with missing imports, type errors, and undefined variables.
Agent must diagnose and fix all issues to pass all tests.
difficulty: hard
estimated_steps: 8
- id: task4
title: "Type Error Detection"
description: >
Identify and fix type annotation errors detected by static type checking.
difficulty: medium
estimated_steps: 5
- id: task5
title: "Unused Imports and Dead Code"
description: >
Find and remove unused imports and dead code using code analysis tools.
difficulty: medium
estimated_steps: 5
- id: task6
title: "Complex Multi-file Refactoring"
description: >
Refactor code across multiple files while maintaining test coverage and fixing architectural issues.
difficulty: hard
estimated_steps: 8
- id: task7
title: "Code Review Workflow"
description: >
Multi-user collaborative task: Junior developer submits fix that needs code review and approval.
difficulty: medium
estimated_steps: 5
- id: task8
title: "Cross-team Bug Investigation"
description: >
Multi-user collaborative task: QA reported bug, dev team investigates and implements fix.
difficulty: hard
estimated_steps: 8
- id: task9
title: "Collaborative Architecture Refactoring"
description: >
Multi-user collaborative task: Large refactoring with multiple developers working in parallel.
difficulty: hard
estimated_steps: 10
action_schema:
type: object
properties:
tool:
type: string
enum: ["run_tests", "read_file", "edit_file", "list_files"]
description: Name of the tool to execute
args:
type: object
description: Tool-specific arguments
required: ["tool"]
observation_schema:
type: object
properties:
pass_rate:
type: number
minimum: 0.0
maximum: 1.0
description: Fraction of tests passing (0.0-1.0)
logs:
type: string
description: Test output, file content, or execution logs
reward:
type: number
minimum: 0.0
maximum: 1.0
description: Reward signal based on pass_rate
done:
type: boolean
description: Whether all tests pass (episode complete)
required: ["pass_rate", "logs"]
reward_schema:
type: object
properties:
reward:
type: number
minimum: 0.0
maximum: 1.0
description: Episode reward based on final pass_rate
done:
type: boolean
description: Whether episode is complete
required: ["reward", "done"]
performance_metrics:
- pass_rate: Fraction of tests passing
- success_rate: Percentage of runs with full pass (across multiple attempts)
- pass@k: Probability that at least k runs succeed
- avg_steps: Average number of tool calls per successful run
- avg_time_ms: Average execution time per run
deployment:
huggingface_space: true
dockerfile: ./Dockerfile
docker_compose: ./docker-compose.yml
health_check: /health
max_runtime_minutes: 20
min_vcpu: 2
min_memory_gb: 8