{
  "CS2W": {
    "source": "https://github.com/guozishan/CS2W",
    "revision": "be6e0519b433cfa9d8388b23308f1380716254fd",
    "license_declared": "CC-BY-SA-4.0"
  },
  "Disfl-QA": {
    "source": "https://github.com/google-research-datasets/Disfl-QA",
    "revision": "1f0c16171c77b3d3408be92c485f11b8998a9189",
    "license_declared": "CC-BY-4.0 for dataset"
  },
  "stage1_manifest": {
    "stage": "pilot-v1-not-natural-mixed-gold",
    "seed": 20261104,
    "rows": 9645,
    "sources": {
      "authored-synthetic CC0-1.0": 2407,
      "CS2W CC-BY-SA-4.0": 3983,
      "CS2W-derived identity CC-BY-SA-4.0": 1653,
      "Disfl-QA-derived identity CC-BY-4.0": 359,
      "Disfl-QA CC-BY-4.0": 1200,
      "authored CC0-1.0": 43
    },
    "tasks": {
      "filler": 1326,
      "cleanup": 5183,
      "identity": 2596,
      "empty": 325,
      "mixed": 192,
      "repair": 18,
      "list": 5
    },
    "rejected": {
      "cs2w_out_of_stage1_task": 1225,
      "cs2w_identity_or_non_deletion_review": 545,
      "heldout_near_jaccard_0.8": 22,
      "heldout_exact_cross_side": 8,
      "duplicate_pair": 79,
      "conservative_disflqa_id_family": 862,
      "conflicting_target_group_rows": 9
    },
    "source_sha256": {
      "CS2W-train.jsonl": "acf249f7e8077337fc8e4029e7c381e117f3bc780fef48c820e4f328e78f563a",
      "CS2W-val.jsonl": "1d89f155c7caecee9d63fc938c42226d43b52fe43f5644ee6c9190b38272ee9a",
      "CS2W-test.jsonl": "32ea92028c3f990b304e2dc5838f487a86a7725b99c9f12fc9696446ecfaab80",
      "Disfl-QA-train.json": "5407199d0c039de5b50cfc16d1ed4a3299c9127cb549da4e4a650b30f4e000eb",
      "Disfl-QA-dev.json": "b60e075b810b27a5130fd0aa2cfbc85753b71a0b30dcd2585f540f0a6afe6492",
      "Disfl-QA-test.json": "404801de916ebcb2caa82661dfd189c0520e2766db6838f6ff548088650e565e"
    },
    "development_rows": 328,
    "evaluation_policy": "Official test excluded from training and development selection; acceptance set used for development, not independent final proof",
    "limits": [
      "Synthetic mixed/noise templates are not natural mixed-language gold",
      "Target-to-target identities require semantic quality review",
      "Deletion-only source filter does not itself guarantee semantic preservation",
      "Disfl-QA 22-character ID grouping is conservative, not a confirmed source paragraph mapping"
    ]
  },
  "stage2_manifest": {
    "stage": "v2 safety replay",
    "rows": 14428,
    "tasks": {
      "identity": 2596,
      "list_identity": 500,
      "cleanup": 5183,
      "identity_replay": 2500,
      "filler": 1325,
      "mixed": 192,
      "focus_identity": 64,
      "quoted_identity": 384,
      "empty": 325,
      "list": 505,
      "demonstrative_identity": 240,
      "repetition_identity": 32,
      "demonstrative_cleanup": 120,
      "acknowledgement_identity": 132,
      "emphasis_identity": 96,
      "quoted_cleanup": 192,
      "repair": 18,
      "deictic_list_identity": 24
    },
    "synthetic_additions": 4796,
    "limits": "Acknowledgement and safety repetitions are deliberate sampling weights; additions are synthetic, not independent human gold."
  },
  "notes": "Only filtered training splits and labelled authored/synthetic replay entered SFT. Evaluation splits excluded; exact cross-side and near-duplicate filtering applied as documented. Sampling repetitions are intentional. Dataset texts are not redistributed in this model repository.",
  "stage3_manifest": {
    "rows": 7407,
    "sha256": "b6992bda2e32bce2ce493239ffd6f1c1d272fe02b717183588c5654fa6fd27d7",
    "policy": "numbered lists; technical models/version/parameter sizes Arabic; coarse round quantities Chinese; idioms preserved",
    "limits": "Authored synthetic plus replay, not natural human-labelled gold; user report examples held out.",
    "prior_training_replay_rows": 2500,
    "new_authored_synthetic_rows": 4907,
    "audit": {
      "rows": 7407,
      "unique_inputs_with_structure": 7331,
      "conflicting_input_labels": 0,
      "empty_targets": 51,
      "tasks": {
        "long_numbered_list": 1100,
        "numbered_list": 1800,
        "numeric_digits": 150,
        "identity": 446,
        "empty": 51,
        "technical_number": 579,
        "filler": 232,
        "numeric_quantity": 396,
        "identity_replay": 429,
        "cleanup": 916,
        "numeric_version": 150,
        "acknowledgement_identity": 22,
        "numbered_list_identity": 300,
        "numeric_time": 184,
        "mixed": 29,
        "repair": 4,
        "quoted_identity": 69,
        "demonstrative_identity": 45,
        "list_identity": 85,
        "coarse_quantity_identity": 160,
        "demonstrative_cleanup": 27,
        "list": 86,
        "idiom_identity": 88,
        "emphasis_identity": 11,
        "quoted_cleanup": 34,
        "repetition_identity": 4,
        "focus_identity": 6,
        "deictic_list_identity": 4
      }
    }
  },
  "stage3_notes": "Additional synthetic numbered-list and numeric-style supervision plus prior training replay. Reported/development examples were excluded from training inputs after lexical normalization. No private recordings or user input history were used for training. Diagnostic probes are not an independent blind evaluation."
}