manifest_core.json

{
  "benchmark": "OPTIC-Bench",
  "description": "OPTIC-Bench (Optical Text In-Context Benchmark) measures how reliably large language models consume text delivered as rendered images versus plain text tokens, covering transcription fidelity, exact-string retrieval, instruction-following, reasoning, injection resistance and multi-turn history compression, in English and Chinese, with fully programmatic scoring and per-call token and cost accounting.",
  "version": "0.3.2",
  "license": {
    "data": "CC-BY-4.0",
    "code": "MIT"
  },
  "master_seed": 20260705,
  "per_cell": 25,
  "families": [
    "transcription",
    "extraction",
    "instruction",
    "reasoning",
    "code",
    "calculation",
    "wordgame",
    "injection",
    "dialog"
  ],
  "languages": [
    "en",
    "zh"
  ],
  "family_difficulties": {
    "transcription": [
      1,
      2,
      3,
      4
    ],
    "extraction": [
      1,
      2,
      3,
      4,
      5
    ],
    "instruction": [
      1,
      2,
      3,
      4
    ],
    "reasoning": [
      1,
      2,
      3,
      4,
      5
    ],
    "code": [
      1,
      2,
      3
    ],
    "calculation": [
      1,
      2,
      3,
      4
    ],
    "wordgame": [
      1,
      2,
      3
    ],
    "injection": [
      1,
      2,
      3
    ],
    "dialog": [
      1,
      2,
      3
    ]
  },
  "family_conditions": {
    "transcription": [
      "tt",
      "ti"
    ],
    "extraction": [
      "tt",
      "ti",
      "it",
      "ii"
    ],
    "instruction": [
      "tt",
      "ti",
      "it",
      "ii"
    ],
    "reasoning": [
      "tt",
      "ti",
      "it",
      "ii"
    ],
    "code": [
      "tt",
      "ti",
      "it",
      "ii"
    ],
    "calculation": [
      "tt",
      "ti",
      "it",
      "ii"
    ],
    "wordgame": [
      "tt",
      "ti",
      "it",
      "ii"
    ],
    "injection": [
      "tt",
      "ti",
      "it",
      "ii"
    ],
    "dialog": [
      "tt",
      "ti"
    ]
  },
  "family_scorers": {
    "transcription": [
      "transcription"
    ],
    "extraction": [
      "exact_fields"
    ],
    "instruction": [
      "record_list",
      "text_output"
    ],
    "reasoning": [
      "answer_match"
    ],
    "code": [
      "text_output"
    ],
    "calculation": [
      "answer_match"
    ],
    "wordgame": [
      "answer_match"
    ],
    "injection": [
      "injection_resistance"
    ],
    "dialog": [
      "answer_match"
    ]
  },
  "modality_conditions": {
    "tt": "instruction text, content text (baseline)",
    "ti": "instruction text, content rendered as image",
    "it": "instruction rendered as image, content text",
    "ii": "instruction and content combined into one image"
  },
  "difficulty_tiers": {
    "1-3": "page-scale",
    "4": "long (~8k characters EN)",
    "5": "extended (~24k characters EN, roughly 6k text tokens)"
  },
  "n_instances": 322,
  "counts": {
    "transcription/en/d1": 4,
    "transcription/zh/d1": 4,
    "transcription/en/d2": 5,
    "transcription/zh/d2": 5,
    "transcription/en/d3": 6,
    "transcription/zh/d3": 6,
    "transcription/en/d4": 4,
    "transcription/zh/d4": 4,
    "extraction/en/d1": 4,
    "extraction/zh/d1": 4,
    "extraction/en/d2": 5,
    "extraction/zh/d2": 5,
    "extraction/en/d3": 6,
    "extraction/zh/d3": 6,
    "extraction/en/d4": 4,
    "extraction/zh/d4": 4,
    "extraction/en/d5": 3,
    "extraction/zh/d5": 3,
    "instruction/en/d1": 4,
    "instruction/zh/d1": 4,
    "instruction/en/d2": 5,
    "instruction/zh/d2": 5,
    "instruction/en/d3": 6,
    "instruction/zh/d3": 6,
    "instruction/en/d4": 4,
    "instruction/zh/d4": 4,
    "reasoning/en/d1": 4,
    "reasoning/zh/d1": 4,
    "reasoning/en/d2": 5,
    "reasoning/zh/d2": 5,
    "reasoning/en/d3": 6,
    "reasoning/zh/d3": 6,
    "reasoning/en/d4": 4,
    "reasoning/zh/d4": 4,
    "reasoning/en/d5": 3,
    "reasoning/zh/d5": 3,
    "code/en/d1": 4,
    "code/zh/d1": 4,
    "code/en/d2": 5,
    "code/zh/d2": 5,
    "code/en/d3": 6,
    "code/zh/d3": 6,
    "calculation/en/d1": 4,
    "calculation/zh/d1": 4,
    "calculation/en/d2": 5,
    "calculation/zh/d2": 5,
    "calculation/en/d3": 6,
    "calculation/zh/d3": 6,
    "calculation/en/d4": 4,
    "calculation/zh/d4": 4,
    "wordgame/en/d1": 4,
    "wordgame/zh/d1": 4,
    "wordgame/en/d2": 5,
    "wordgame/zh/d2": 5,
    "wordgame/en/d3": 6,
    "wordgame/zh/d3": 6,
    "injection/en/d1": 4,
    "injection/zh/d1": 4,
    "injection/en/d2": 5,
    "injection/zh/d2": 5,
    "injection/en/d3": 6,
    "injection/zh/d3": 6,
    "dialog/en/d1": 4,
    "dialog/zh/d1": 4,
    "dialog/en/d2": 5,
    "dialog/zh/d2": 5,
    "dialog/en/d3": 6,
    "dialog/zh/d3": 6
  },
  "subset": "core",
  "core_per_difficulty": {
    "1": 4,
    "2": 5,
    "3": 6,
    "4": 4,
    "5": 3
  },
  "note": "Deterministic core subset of the full dataset, weighted towards harder tiers for cheap routine evaluation. Instances are a strict subset of the full set and share its ground truth and images."
}