{
  "name": "MMStar",
  "release_date": "2026-05-15",
  "subsets": {
    "main": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "multiple_choice_vqa",
      "prompt_template": "<image>{{ question }}\nPlease select the correct answer from the options above.",
      "mapping_from_source": {
        "media": {
          "from": "image",
          "type": "list",
          "min_items": 1,
          "max_items": 1
        },
        "id": {
          "from": "index"
        },
        "question": {
          "from": "question",
          "note": "MCQ options are embedded as `Options: A: ..., B: ..., ...` in source"
        },
        "answer": {
          "from": "answer",
          "optional": true
        },
        "extra": {
          "category": {
            "from": "category"
          },
          "l2_category": {
            "from": "l2_category"
          }
        },
        "source": {
          "format": "json",
          "url": {
            "val": "https://huggingface.co/datasets/Lin-Chen/MMStar"
          }
        }
      },
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_mcq.py (ImageMCQDataset.build_prompt — bare question + 'Please select the correct answer from the options above. ' trailer; source `question` column already contains 'Options: A: ..., B: ..., C: ..., D: ...' inline so the template avoids re-rendering options)",
        "notes": "Tier 3: VLMEvalKit ImageMCQDataset.build_prompt. The MMStar source `question` column already includes the options inline (e.g., 'Options: A: ..., B: ..., C: ..., D: ...'), so the template appends only the answer-format trailer. lmms-eval's mmstar task uses a different trailer ('Answer with the option\\'s letter from the given choices directly') and is not the canonical reference for the current prompt."
      }
    }
  }
}