{
  "name": "MMBench",
  "release_date": "2024-01-26",
  "subsets": {
    "cc": {
      "language": [
        "zh"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "multiple_choice_qa",
      "score_pipeline": [
        "rule-match",
        "llm-match"
      ],
      "score_protocol": {
        "reference": "vlmevalkit@vlmeval/dataset/image_mcq.py:269-274 — listinstr(['mmbench','ccbench',...]) forces circular=True -> vlmevalkit@vlmeval/dataset/utils/multiple_choice.py:511-582 mcq_circular_eval: rotation group (index % 1e6) scores hit=1 only if EVERY rotation is answered correctly; per-rotation grading = can_infer prefetch + GPT extractor (vlmeval/dataset/utils/multiple_choice.py:359-407)",
        "note": "Official protocol is CircularEval: a question counts correct only if all option rotations are correct. The mm-eval copy ships pre-rotated rows (VLMEvalKit index convention: rotation ids = base_id + k*1e6, verified via HF rows, e.g. mmbench_en_dev_1001484 / mmbench_cc_test_1000524); a per-sample scorer reports vanilla per-rotation accuracy, systematically higher than official CircularEval. Circular group id is derivable as int(id_suffix) % 1e6. cc subset is CCBench (zh cultural), also circular per the same listinstr branch; test-split answers ARE present (verified 'A' etc.)."
      },
      "prompt_template": "<image>{% if hint %}Hint: {{ hint }}\n{% endif %}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
        },
        "hint": {
          "from": "hint",
          "optional": true
        },
        "options": {
          "from": "options",
          "optional": true,
          "note": "list source values are normalized to {A,B,...} dict"
        },
        "extra": {
          "source_id": {
            "from": "source_id"
          },
          "category": {
            "from": "category"
          },
          "L2_category": {
            "from": "L2_category"
          },
          "source": {
            "from": "source"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "cc_test": "https://huggingface.co/datasets/lmms-lab/MMBench"
          }
        }
      },
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_mcq.py#L212-L247 (ImageMCQDataset.build_prompt)",
        "notes": "Tier 3: VLMEvalKit canonical MCQ template (byte-for-byte)"
      }
    },
    "cn": {
      "language": [
        "zh"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "multiple_choice_qa",
      "score_pipeline": [
        "rule-match",
        "llm-match"
      ],
      "score_protocol": {
        "reference": "vlmevalkit@vlmeval/dataset/image_mcq.py:269-274 — listinstr(['mmbench','ccbench',...]) forces circular=True -> vlmevalkit@vlmeval/dataset/utils/multiple_choice.py:511-582 mcq_circular_eval: rotation group (index % 1e6) scores hit=1 only if EVERY rotation is answered correctly; per-rotation grading = can_infer prefetch + GPT extractor (vlmeval/dataset/utils/multiple_choice.py:359-407)",
        "note": "Official protocol is CircularEval: a question counts correct only if all option rotations are correct. The mm-eval copy ships pre-rotated rows (VLMEvalKit index convention: rotation ids = base_id + k*1e6, verified via HF rows, e.g. mmbench_en_dev_1001484 / mmbench_cc_test_1000524); a per-sample scorer reports vanilla per-rotation accuracy, systematically higher than official CircularEval. Circular group id is derivable as int(id_suffix) % 1e6. cn test split ships empty answers (official answers withheld, submission-only) — only dev is locally scorable (verified via HF rows)."
      },
      "prompt_template": "<image>{% if hint %}Hint: {{ hint }}\n{% endif %}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
        },
        "hint": {
          "from": "hint",
          "optional": true
        },
        "options": {
          "from": "options",
          "optional": true,
          "note": "list source values are normalized to {A,B,...} dict"
        },
        "extra": {
          "source_id": {
            "from": "source_id"
          },
          "category": {
            "from": "category"
          },
          "L2_category": {
            "from": "L2_category"
          },
          "source": {
            "from": "source"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "cn_dev": "https://huggingface.co/datasets/lmms-lab/MMBench",
            "cn_test": "https://huggingface.co/datasets/lmms-lab/MMBench"
          }
        }
      },
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_mcq.py#L212-L247 (ImageMCQDataset.build_prompt)",
        "notes": "Tier 3: VLMEvalKit canonical MCQ template (byte-for-byte); English wrapper preserved for Chinese subset per VLMEvalKit convention"
      }
    },
    "en": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "multiple_choice_qa",
      "score_pipeline": [
        "rule-match",
        "llm-match"
      ],
      "score_protocol": {
        "reference": "vlmevalkit@vlmeval/dataset/image_mcq.py:269-274 — listinstr(['mmbench','ccbench',...]) forces circular=True -> vlmevalkit@vlmeval/dataset/utils/multiple_choice.py:511-582 mcq_circular_eval: rotation group (index % 1e6) scores hit=1 only if EVERY rotation is answered correctly; per-rotation grading = can_infer prefetch + GPT extractor (vlmeval/dataset/utils/multiple_choice.py:359-407)",
        "note": "Official protocol is CircularEval: a question counts correct only if all option rotations are correct. The mm-eval copy ships pre-rotated rows (VLMEvalKit index convention: rotation ids = base_id + k*1e6, verified via HF rows, e.g. mmbench_en_dev_1001484 / mmbench_cc_test_1000524); a per-sample scorer reports vanilla per-rotation accuracy, systematically higher than official CircularEval. Circular group id is derivable as int(id_suffix) % 1e6. en test split ships empty answers (official answers withheld, submission-only) — only dev is locally scorable (verified via HF rows)."
      },
      "prompt_template": "<image>{% if hint %}Hint: {{ hint }}\n{% endif %}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
        },
        "hint": {
          "from": "hint",
          "optional": true
        },
        "options": {
          "from": "options",
          "optional": true,
          "note": "list source values are normalized to {A,B,...} dict"
        },
        "extra": {
          "source_id": {
            "from": "source_id"
          },
          "category": {
            "from": "category"
          },
          "L2_category": {
            "from": "L2_category"
          },
          "source": {
            "from": "source"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "en_dev": "https://huggingface.co/datasets/lmms-lab/MMBench",
            "en_test": "https://huggingface.co/datasets/lmms-lab/MMBench"
          }
        }
      },
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_mcq.py#L212-L247 (ImageMCQDataset.build_prompt)",
        "notes": "Tier 3: VLMEvalKit canonical MCQ template (byte-for-byte)"
      }
    }
  }
}