{
  "missing_values": "Retained from frozen inputs; missing does not mean negative or zero.",
  "configs": {
    "pairs": {
      "rows": 456,
      "split": "test",
      "columns": {
        "source_dataset": {
          "type": "large_string",
          "description": "Original corpus and split identifier."
        },
        "family": {
          "type": "large_string",
          "description": "Source corpus family."
        },
        "item_id": {
          "type": "large_string",
          "description": "Encounter identifier within source_dataset."
        },
        "distractor_type": {
          "type": "large_string",
          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
        },
        "clean_transcript": {
          "type": "large_string",
          "description": "Unmodified source transcript (Notes) or saved ASR of clean recording (Audio)."
        },
        "distracted_transcript": {
          "type": "large_string",
          "description": "Transcript with inserted conversation (Notes) or saved ASR of mixed recording (Audio)."
        },
        "reference_note": {
          "type": "large_string",
          "description": "Reference clinical note supplied by original source corpus."
        },
        "distractor_conversation": {
          "type": "large_string",
          "description": "Inserted exchange or human-transcribed donor speech."
        },
        "distractor_summary": {
          "type": "large_string",
          "description": "Target content supplied to the contamination judge."
        },
        "distractor_topic": {
          "type": "large_string",
          "description": "Generated aside topic or audio donor-content label."
        },
        "donor": {
          "type": "int64",
          "description": "Seeded donor ordinal for a foreground consultation."
        },
        "donor_id": {
          "type": "large_string",
          "description": "Identifier of the background consultation."
        },
        "level_db": {
          "type": "double",
          "description": "Background RMS relative to foreground RMS, in dB."
        },
        "onset": {
          "type": "double",
          "description": "Overlay onset in foreground audio, seconds."
        },
        "segment_start": {
          "type": "double",
          "description": "Donor segment start, seconds in original donor recording."
        },
        "segment_end": {
          "type": "double",
          "description": "Donor segment end, seconds in original donor recording."
        },
        "segment_seconds": {
          "type": "double",
          "description": "Duration of donor segment, seconds."
        },
        "segment_score": {
          "type": "int64",
          "description": "Medical-keyword score used for donor-window selection."
        },
        "wav": {
          "type": "large_string",
          "description": "Historical relative WAV path; use recordings configuration for embedded audio."
        },
        "human_transcript": {
          "type": "large_string",
          "description": "Original human reference transcript for foreground consultation."
        }
      }
    },
    "transcripts": {
      "rows": 513,
      "split": "test",
      "columns": {
        "item_id": {
          "type": "large_string",
          "description": "Encounter identifier within source_dataset."
        },
        "distractor_type": {
          "type": "large_string",
          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
        },
        "condition": {
          "type": "large_string",
          "description": "clean or distracted."
        },
        "wav": {
          "type": "large_string",
          "description": "Historical relative WAV path; use recordings configuration for embedded audio."
        },
        "text": {
          "type": "large_string",
          "description": "Saved ASR transcript."
        },
        "segments": {
          "type": "large_string",
          "description": "Serialized ASR chunk texts and time spans."
        },
        "asr_model": {
          "type": "large_string",
          "description": "ASR model identifier."
        },
        "timestamps": {
          "type": "large_string",
          "description": "Serialized chunk timestamp values."
        },
        "seconds": {
          "type": "double",
          "description": "Historical ASR processing time in seconds, not audio duration."
        }
      }
    },
    "notes": {
      "rows": 3648,
      "split": "test",
      "columns": {
        "source_dataset": {
          "type": "large_string",
          "description": "Original corpus and split identifier."
        },
        "family": {
          "type": "large_string",
          "description": "Source corpus family."
        },
        "item_id": {
          "type": "large_string",
          "description": "Encounter identifier within source_dataset."
        },
        "distractor_type": {
          "type": "large_string",
          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
        },
        "condition": {
          "type": "large_string",
          "description": "clean or distracted."
        },
        "transcript": {
          "type": "large_string",
          "description": "Input transcript supplied to note generator."
        },
        "reference_note": {
          "type": "large_string",
          "description": "Reference clinical note supplied by original source corpus."
        },
        "distractor_summary": {
          "type": "large_string",
          "description": "Target content supplied to the contamination judge."
        },
        "note": {
          "type": "large_string",
          "description": "Generated clinical note; empty saved records are retained."
        },
        "model": {
          "type": "large_string",
          "description": "Note-generation model identifier used in saved experiment."
        }
      }
    },
    "judgments": {
      "rows": 3648,
      "split": "test",
      "columns": {
        "source_dataset": {
          "type": "large_string",
          "description": "Original corpus and split identifier."
        },
        "item_id": {
          "type": "large_string",
          "description": "Encounter identifier within source_dataset."
        },
        "family": {
          "type": "large_string",
          "description": "Source corpus family."
        },
        "model": {
          "type": "large_string",
          "description": "Note-generation model identifier used in saved experiment."
        },
        "distractor_type": {
          "type": "large_string",
          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
        },
        "condition": {
          "type": "large_string",
          "description": "clean or distracted."
        },
        "contamination": {
          "type": "double",
          "description": "Binary target-contamination flag."
        },
        "severity": {
          "type": "double",
          "description": "Target-contamination severity, 0–3."
        },
        "judge_reasoning": {
          "type": "large_string",
          "description": "Model-generated scoring explanation."
        },
        "parse_error": {
          "type": "bool",
          "description": "True if automated response could not be parsed; do not convert missing values to zero."
        },
        "judge_model": {
          "type": "large_string",
          "description": "Judge model identifier."
        },
        "judge_protocol": {
          "type": "large_string",
          "description": "Scoring protocol identifier."
        },
        "clinical_correctness": {
          "type": "double",
          "description": "Automated quality score, 1–5, higher is better."
        },
        "completeness": {
          "type": "double",
          "description": "Automated completeness score, 1–5, higher is better."
        },
        "succinctness": {
          "type": "double",
          "description": "Automated concision score, 1–5, higher is better."
        },
        "hallucination": {
          "type": "double",
          "description": "Automated binary flag for findings unsupported by the transcript."
        },
        "overall_quality": {
          "type": "double",
          "description": "Automated overall quality score, 1–5, higher is better."
        }
      }
    },
    "leakage": {
      "rows": 456,
      "split": "test",
      "columns": {
        "item_id": {
          "type": "large_string",
          "description": "Encounter identifier within source_dataset."
        },
        "distractor_type": {
          "type": "large_string",
          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
        },
        "donor": {
          "type": "int64",
          "description": "Seeded donor ordinal for a foreground consultation."
        },
        "donor_id": {
          "type": "large_string",
          "description": "Identifier of the background consultation."
        },
        "level_db": {
          "type": "double",
          "description": "Background RMS relative to foreground RMS, in dB."
        },
        "onset": {
          "type": "double",
          "description": "Overlay onset in foreground audio, seconds."
        },
        "segment_start": {
          "type": "double",
          "description": "Donor segment start, seconds in original donor recording."
        },
        "segment_end": {
          "type": "double",
          "description": "Donor segment end, seconds in original donor recording."
        },
        "segment_seconds": {
          "type": "double",
          "description": "Duration of donor segment, seconds."
        },
        "segment_score": {
          "type": "int64",
          "description": "Medical-keyword score used for donor-window selection."
        },
        "n_content": {
          "type": "int64",
          "description": "Number of donor segment content words considered."
        },
        "n_leaked": {
          "type": "int64",
          "description": "Number of donor content words in mixed ASR but absent from clean ASR."
        },
        "leakage": {
          "type": "double",
          "description": "Fraction n_leaked/n_content."
        },
        "overlap_clean": {
          "type": "double",
          "description": "Fraction of donor content words already present in clean ASR."
        },
        "leaked_words": {
          "type": "large_string",
          "description": "Donor content words newly present in mixed ASR."
        },
        "wer_vs_clean": {
          "type": "double",
          "description": "Mixed-ASR word error rate relative to clean-ASR transcript."
        },
        "wer_clean_vs_human": {
          "type": "double",
          "description": "Clean-ASR word error rate relative to human reference."
        }
      }
    },
    "recordings": {
      "rows": 513,
      "split": "test",
      "columns": {
        "recording_id": {
          "type": "string",
          "description": "Unique recording key: item_id plus distractor_type."
        },
        "item_id": {
          "type": "string",
          "description": "Encounter identifier within source_dataset."
        },
        "distractor_type": {
          "type": "string",
          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
        },
        "condition": {
          "type": "string",
          "description": "clean or distracted."
        },
        "donor_id": {
          "type": "string",
          "description": "Identifier of the background consultation."
        },
        "level_db": {
          "type": "double",
          "description": "Background RMS relative to foreground RMS, in dB."
        },
        "onset": {
          "type": "double",
          "description": "Overlay onset in foreground audio, seconds."
        },
        "segment_start": {
          "type": "double",
          "description": "Donor segment start, seconds in original donor recording."
        },
        "segment_end": {
          "type": "double",
          "description": "Donor segment end, seconds in original donor recording."
        },
        "duration_seconds": {
          "type": "double",
          "description": "Waveform duration in seconds."
        },
        "sampling_rate": {
          "type": "int32",
          "description": "Samples per second; 16000."
        },
        "pcm_sha256": {
          "type": "string",
          "description": "SHA256 of little-endian signed 16-bit reconstructed PCM samples."
        },
        "audio": {
          "type": "struct<bytes: binary, path: string>",
          "description": "Embedded FLAC bytes and filename with Hugging Face Audio metadata."
        }
      }
    }
  }
}