{
  "name": "PMC-VQA",
  "release_date": "2026-05-26",
  "subsets": {
    "main": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "multiple_choice_qa",
      "score_pipeline": [
        "exact-match",
        "rule-match"
      ],
      "score_protocol": {
        "reference": "official@github.com/xiaoman-zhang/PMC-VQA src/MedVInT_TD/test.py — find_most_similar_index via difflib.SequenceMatcher maps prediction to nearest choice; ACC++ when index_pred==index_label; deterministic string-similarity, no LLM.",
        "note": "Official metric maps the free-text prediction to the closest of the 4 choices by difflib fuzzy similarity (not exact letter). mm-eval copy uses a VLMEvalKit-style letter prompt; the exact/template matcher on the emitted letter is a reasonable rule equivalent. VLMEvalKit does not ship PMC-VQA."
      },
      "prompt_template": "<image>Question: {{ question }}\nOptions:\n{% for k, v in options.items() %}{{ k }}. {{ v }}\n{% endfor %}Please select the correct answer from the options above. ",
      "mapping_from_source": {
        "media": {
          "from": "image",
          "type": "list",
          "min_items": 1,
          "max_items": 1
        },
        "id": {
          "from": "id"
        },
        "question": {
          "from": "question"
        },
        "answer": {
          "from": "answer",
          "optional": true
        },
        "options": {
          "from": "options",
          "optional": true,
          "note": "list source values are normalized to {A,B,...} dict"
        },
        "extra": {
          "source_id": {
            "from": "source_id"
          }
        },
        "source": {
          "format": "json",
          "url": {
            "test": "https://huggingface.co/datasets/xmcmic/PMC-VQA"
          }
        }
      },
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_mcq.py (PMC-VQA uses ImageMCQDataset.build_prompt canonical MCQ template)",
        "notes": "Tier 3: VLMEvalKit ImageMCQDataset.build_prompt; PMC-VQA is evaluated via VLMEvalKit's standard MCQ flow."
      }
    }
  }
}