metadata/closurebench_dynamic_dialogue_metadata.json

{
  "source_dataset": "data/base/full.jsonl",
  "dataset": "data/dynamic_dialogue/full.jsonl",
  "n_dialogues": 100,
  "n_turns": 300,
  "turns_per_dialogue": 3,
  "seed": 11,
  "gold_patterns": {
    "unknown|false|false": 50,
    "unknown|false|unknown": 50
  },
  "dialogue_types": {
    "persist_after_narrowing": 50,
    "reopen": 50
  },
  "domains": {
    "cloud_deployment": 8,
    "finance_controls": 15,
    "hospital_access": 13,
    "library_services": 11,
    "manufacturing_quality": 13,
    "procurement_review": 12,
    "robotics_operations": 15,
    "university_registration": 13
  },
  "families": {
    "closed_derived_missing_antecedent": 17,
    "closed_missing_direct": 15,
    "closed_missing_with_open_distractor": 18,
    "open_derived_missing_antecedent": 16,
    "open_missing_direct": 18,
    "open_missing_with_closed_distractor": 16
  },
  "turn_update_sequence": [
    "SET_COMPLETE_PREDICATES([])",
    "ADD_COMPLETE_PREDICATES(vocabulary)",
    "REPLACE_COMPLETE_PREDICATES(local_subset)"
  ],
  "primary_metrics": [
    "dialogue_all_answer_accuracy",
    "revision_accuracy",
    "persistence_accuracy",
    "dialogue_all_turn_accuracy",
    "completeness_addition_accuracy",
    "reopening_accuracy",
    "narrowing_persistence_accuracy",
    "closure_inertia_rate",
    "closure_retention_failure_rate",
    "changed_flag_accuracy",
    "applied_update_accuracy",
    "anchoring_error_rate"
  ]
}