---
dataset_info:
  features:
  - name: episode_id
    dtype: string
  - name: scenario_id
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: reward
    dtype: float64
  - name: reward_breakdown
    struct:
    - name: schema_version
      dtype: string
    - name: total
      dtype: float64
    - name: format_gate
      dtype: bool
    - name: components
      struct:
      - name: format_gate
        struct:
        - name: raw
          dtype: float64
        - name: weighted
          dtype: float64
        - name: weight
          dtype: float64
        - name: sub_scores
          struct:
          - name: valid
            dtype: float64
      - name: diagnosis
        struct:
        - name: raw
          dtype: float64
        - name: weighted
          dtype: float64
        - name: weight
          dtype: float64
        - name: sub_scores
          struct:
          - name: matrix_lookup
            dtype: float64
          - name: predicted
            dtype: float64
      - name: action_quality
        struct:
        - name: raw
          dtype: float64
        - name: weighted
          dtype: float64
        - name: weight
          dtype: float64
        - name: sub_scores
          struct:
          - name: no_secondary
            dtype: float64
      - name: cost_efficiency
        struct:
        - name: raw
          dtype: float64
        - name: weighted
          dtype: float64
        - name: weight
          dtype: float64
        - name: sub_scores
          struct:
          - name: total_cost
            dtype: float64
          - name: ratio
            dtype: float64
      - name: investigation
        struct:
        - name: raw
          dtype: float64
        - name: weighted
          dtype: float64
        - name: weight
          dtype: float64
        - name: sub_scores
          struct:
          - name: coverage
            dtype: float64
          - name: ordering
            dtype: float64
          - name: redundancy_penalty
            dtype: float64
          - name: min_ev_multiplier
            dtype: float64
      - name: minimal_evidence
        struct:
        - name: raw
          dtype: float64
        - name: weighted
          dtype: float64
        - name: weight
          dtype: float64
        - name: sub_scores
          struct:
          - name: min_set_used
            dtype: float64
          - name: extras
            dtype: float64
      - name: time
        struct:
        - name: raw
          dtype: float64
        - name: weighted
          dtype: float64
        - name: weight
          dtype: float64
        - name: sub_scores
          struct:
          - name: steps
            dtype: float64
          - name: excess
            dtype: float64
      - name: anti_gaming
        struct:
        - name: raw
          dtype: float64
        - name: weighted
          dtype: float64
        - name: weight
          dtype: float64
        - name: sub_scores
          struct:
          - name: no_info_penalty
            dtype: float64
          - name: quarantine_rate
            dtype: float64
          - name: quarantine_penalty
            dtype: float64
          - name: brier_bonus
            dtype: float64
    - name: counterfactual
      dtype: 'null'
  - name: format_gate_passed
    dtype: bool
  splits:
  - name: train
    num_bytes: 5257828
    num_examples: 718
  download_size: 5258953
  dataset_size: 5257828
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---