metadata.json

{
  "name": "Visual7W",
  "release_date": "2026-05-27",
  "subsets": {
    "main": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "multiple_choice_qa",
      "score_pipeline": [
        "rule-match",
        "llm-match"
      ],
      "score_protocol": {
        "reference": "Published copy (HF datasets-server, mm-eval/Visual7W test, 69817 rows): choices A-D embedded in question text with trailer 'Answer with the letter.', gold = letter. Original Visual7W protocol (Zhu et al., CVPR 2016) is log-likelihood candidate ranking accuracy — not executable for generative/API models; classified per the generic VLMEvalKit MCQ letter-extraction protocol (vlmeval/dataset/utils/multiple_choice.py:359-406: rule prefetch + GPT fallback, exact letter compare).",
        "note": "No current framework implements Visual7W (metadata cites lmms_eval/tasks/visual7w which does not exist in present lmms-eval main or v0.2.4-v0.3.1 tags). Official 2016 metric is candidate-scoring accuracy; the published generative letter-answer format is graded as generic MCQ. Published options dict is empty (choices live in the question text)."
      },
      "prompt_template": "<image>{{ question }}",
      "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
        },
        "extra": {
          "source_id": {
            "from": "source_id"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "test": "https://huggingface.co/datasets/nimapourjafar/mm_visual7w"
          }
        }
      },
      "prompt_template_source": {
        "origin": "fallback",
        "reference": "T2",
        "notes": "Tier 5: no higher tier publishes a prompt. T1: yukezhu/visual7w-toolkit is classification-era (evaluate.py consumes prediction JSON; no model-input prompt); T2: no source prompt column (choices are embedded in the question text of nimapourjafar/mm_visual7w, which itself triggers the T2 bare-question rule); T3: no VLMEvalKit visual7w dataset (tree + code search, 2026-07-07); T4: no lmms-eval tasks/visual7w in current main or v0.2.4-v0.3.1 tags."
      }
    }
  }
}