openenv.yaml

name: debugml
description: >
  An environment where an agent improves a machine learning pipeline
  by applying transformations like scaling, feature adjustment, and split correction.

entry_point: env.environments:DebugMLEnv

observation_space:
  type: object
  properties:
    accuracy:
      type: float
    precision:
      type: float
    recall:
      type: float
    scaling:
      type: boolean
    feature_count:
      type: integer
    test_split:
      type: float
    model_type:
      type: string

action_space:
  type: object
  properties:
    type:
      type: string
      enum:
        - add_scaling
        - fix_split
        - add_feature
        - remove_feature

reward_range: [0.01, 0.99]

tasks:
  - name: fix_basics
    description: Fix a pipeline with missing scaling and a bad train/test split
    difficulty: easy
    grader: grade_task

  - name: optimize_features
    description: Tune feature count when scaling is already applied
    difficulty: medium
    grader: grade_task

  - name: full_pipeline_optimization
    description: Fix everything from a random starting state
    difficulty: hard
    grader: grade_task

  - name: stability_optimization
    description: Maintain high accuracy with minimal unnecessary steps
    difficulty: hard
    grader: grade_task