{
  "title": "SPC/COCO public source-metadata overlap audit",
  "counts": {
    "paragraph_rows": 19561,
    "paragraph_distinct_vg_ids": 19551,
    "official_split_distinct_vg_ids": 19551,
    "paragraph_ids_equal_official_split_union": true,
    "paragraph_ids_not_in_split_union": [],
    "split_ids_not_in_paragraphs": [],
    "duplicate_paragraph_image_id_counts": {
      "2349250": 2,
      "2385924": 2,
      "2408653": 2,
      "2412652": 2,
      "2340853": 2,
      "2377108": 2,
      "2414810": 2,
      "2327869": 2,
      "2363167": 2,
      "2370496": 2
    },
    "paragraph_vg_ids_missing_from_vg_metadata": [],
    "paragraph_distinct_vg_ids_with_coco_id": 9595,
    "coco_metadata_images_by_official_filepath": {
      "val2014": 40504,
      "train2014": 82783
    },
    "spc_distinct_vg_ids_joined_to_coco_trainval": 9595,
    "spc_unique_coco_ids_joined_to_trainval": 9593,
    "vg_ids_per_duplicate_coco_id": {
      "112040": 2,
      "322194": 2
    },
    "overlap_distinct_vg_ids_by_official_filepath": {
      "train2014": 6259,
      "val2014": 3336
    },
    "overlap_unique_coco_ids_by_official_filepath": {
      "train2014": 6258,
      "val2014": 3335
    }
  },
  "method": [
    "Load all paragraph rows from the same published paragraphs_v1.json source used by the released dataset loader.",
    "Check its distinct Visual Genome image IDs against the union of the official SPC train, validation, and test split IDs.",
    "Join Visual Genome image_id to coco_id using published image_data.json metadata.",
    "Join coco_id to cocoid in the original authors' dataset_coco.json; use filepath train2014/val2014, not the later Karpathy split label.",
    "Count both distinct Visual Genome IDs and unique COCO IDs; separately report any duplicate mappings."
  ],
  "sources": [
    {
      "url": "https://cs.stanford.edu/people/ranjaykrishna/im2p/train_split.json",
      "sha256": "c1a5685dfd6994b15b40b59cf422e12a665f6bf4475bca16b863be5d15e17898"
    },
    {
      "url": "https://cs.stanford.edu/people/ranjaykrishna/im2p/val_split.json",
      "sha256": "d7e735db44b73530f65fbfe1087c35552fd774caa5b2169bde47274779116402"
    },
    {
      "url": "https://cs.stanford.edu/people/ranjaykrishna/im2p/test_split.json",
      "sha256": "aa93f1636dce87bff4798bb9b4ac738eef25a953e0f6e9d5f34fc11d83965bb4"
    },
    {
      "url": "https://homes.cs.washington.edu/~ranjay/visualgenome/data/dataset/image_data.json.zip",
      "sha256": "b87a94918cb2ff4d952cf1dfeca0b9cf6cd6fd204c2f8704645653be1163681a"
    },
    {
      "url": "https://cs.stanford.edu/people/karpathy/deepimagesent/caption_datasets.zip",
      "sha256": "4cfd70132527b80933105e5829dc9034eaab9573482e2e680abbab6130244817"
    },
    {
      "url": "https://homes.cs.washington.edu/~ranjay/visualgenome/data/dataset/paragraphs_v1.json.zip",
      "sha256": "8407f1068e0711e5f6550d641c523ee26984ee8e02c1d51564369cccd56e8b26"
    }
  ],
  "released_code_commit": "fdfad84ea488f29702a61a455be71622c4ff5d49",
  "released_code_locations": [
    "https://github.com/dominik-schnaus/unpaired-rosetta/blob/fdfad84ea488f29702a61a455be71622c4ff5d49/experiments/cross_dataset.py",
    "https://github.com/dominik-schnaus/unpaired-rosetta/blob/fdfad84ea488f29702a61a455be71622c4ff5d49/experiments/alignment.py#L64-L69",
    "https://github.com/dominik-schnaus/unpaired-rosetta/blob/fdfad84ea488f29702a61a455be71622c4ff5d49/unpaired_rosetta/evaluation.py#L58-L65",
    "https://github.com/dominik-schnaus/unpaired-rosetta/blob/fdfad84ea488f29702a61a455be71622c4ff5d49/modalities/image_captions/datasets.py#L173-L180"
  ],
  "uncertainty_and_scope": [
    "The released loader and splitting logic were inspected. The full published experiments have not been replicated; the subsequent 15-fit reduced pilot only validates the control pipeline.",
    "These are source-image identity overlaps, not proof that the aligner was given pairing labels or row-order correspondence.",
    "The numerical impact on reported performance is unknown without a deduplicated rerun.",
    "The published metadata and released code may differ from unpublished cached inputs used for the paper.",
    "COCO validation overlap means SPC fitting text describes those validation source images; it does not mean their original COCO captions were provided.",
    "This finding concerns the COCO-to-SPC protocol and does not invalidate the separate disjoint-half COCO experiment."
  ]
}