dataset_info.json

{
  "name": "AIForge-Doc",
  "version": "2.0.0",
  "description": "Paired benchmark of GPT-Image-2 document forgeries; every forged image has its authentic source and a pixel-precise tampered-region mask in DocTamper-compatible format.",
  "created": "2026-04-28",
  "license": "CC BY 4.0",
  "format": "DocTamper-compatible",
  "mask_convention": {
    "0": "authentic pixel",
    "255": "tampered pixel"
  },
  "filename_format": "9-digit zero-padded integer (000000001.png)",
  "splits": {
    "TrainingSet": {
      "images": 2453,
      "masks": 2453
    },
    "TestingSet": {
      "images": 613,
      "masks": 613
    },
    "total_forged": 3066
  },
  "inpainting_tools": {
    "openai-gpt-image-2": {
      "description": "OpenAI GPT-Image-2 (April 2026 release), accessed via an API gateway exposing the openai:gpt-image@2 model.",
      "count": 3066,
      "share": 1.0
    }
  },
  "source_datasets": {
    "cord": {
      "doc_type": "receipt",
      "count": 983,
      "languages": [
        "id"
      ],
      "license": "CC BY 4.0",
      "citation": "Park et al., 2019"
    },
    "wildreceipt": {
      "doc_type": "receipt",
      "count": 1336,
      "languages": [
        "en"
      ],
      "license": "Apache 2.0",
      "citation": "Sun et al., 2021"
    },
    "sroie": {
      "doc_type": "receipt",
      "count": 329,
      "languages": [
        "en"
      ],
      "license": "ICDAR 2019 research-use",
      "citation": "Huang et al., 2019"
    },
    "xfund": {
      "doc_type": "form",
      "count": 418,
      "languages": [
        "de",
        "es",
        "fr",
        "it",
        "ja",
        "pt",
        "zh"
      ],
      "license": "CC BY 4.0",
      "citation": "Xu et al., 2022"
    }
  },
  "by_language": {
    "en": 1665,
    "id": 983,
    "it": 80,
    "de": 78,
    "es": 67,
    "pt": 62,
    "fr": 56,
    "zh": 38,
    "ja": 37
  },
  "by_doctype": {
    "receipt": 2648,
    "form": 418
  },
  "relation_to_v1": {
    "shared_specs": "v2 reuses v1 forgery specifications spec-for-spec; every successful v2 image has a same-spec v1 counterpart where v1 generation was successful.",
    "differences": "v1 used Gemini-2.5 Flash Image and Ideogram-v2 Edit; v2 uses OpenAI GPT-Image-2 only. v1 used a red-fill marker; v2 uses a 3-pixel green-outline marker."
  },
  "baseline_auc": {
    "trufor": 0.599,
    "doctamper": 0.585,
    "self_judge": 0.532,
    "human_2afc": 0.501
  },
  "calibration_auc": {
    "trufor_traditional": 0.962,
    "doctamper_traditional": 0.852
  }
}