{
  "name": "WildVision-Bench",
  "release_date": "2024-06-17",
  "subsets": {
    "main": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "open_ended_qa",
      "score_pipeline": [
        "llm-judge"
      ],
      "score_params": {
        "rubric": "scale",
        "scale": [
          -2,
          2
        ],
        "official_judge_model": "gpt-4o"
      },
      "score_protocol": {
        "reference": "VLMEvalKit@vlmeval/dataset/wildvision.py:15-56,96-220 — GPT-4o judge with arena-style system prompt compares claude-3-sonnet reference (A) vs model (B), 5-way verdict A>>B..B>>A mapped to -2..2; headline Reward = (100*much_better + 50*better - 50*worse - 100*much_worse)/n and win rate. Reference answers: official WildVision-AI/WildVision-Bench GitHub data/vision_bench_0617/model_answers/claude-3-sonnet-20240229.jsonl (byte-identical to the claude3_sonnet column of VLMEvalKit WildVision.tsv, md5 b38f80156d49411c594772866b0d0b52).",
        "note": "Official protocol is PAIRWISE preference vs a stored claude-3-sonnet reference answer with win-rate/reward headline — not absolute per-sample grading. Fixed 2026-07-07: per-row `reference_response` extra now carries the claude-3-sonnet-20240229 reference answer. Coverage: test_0617 = 500/500 (official answers file covers exactly this split; VLMEvalKit's 500-sample WildVision set is the 0617 set). test_0701 = 191/500 (only the rows shared with 0617; no public claude-3-sonnet answers exist for the remaining 309 rows — the WildVision org repo carrying model responses, WildVision/wildvision-arena-data, is gated, ships only 0617-era responses, and the official GitHub repo has no 0701 answers). Rows without a reference omit the field. The source questions ship no ground-truth answer column, so `answer` is empty by design."
      },
      "prompt_template": "<image>{{ question }}",
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_vqa.py (WildVision-Bench — judge-scored; bare question)",
        "notes": "Tier 3: WildVision-Bench judge-scored eval: bare question per the official WildVision-Bench paper."
      },
      "mapping_from_source": {
        "media": {
          "from": "image",
          "type": "list",
          "min_items": 1,
          "max_items": 1
        },
        "id": {
          "from": "question_id"
        },
        "question": {
          "from": "instruction"
        },
        "extra": {
          "reference_response": {
            "from": "claude3_sonnet",
            "note": "claude-3-sonnet-20240229 reference answer joined by question_id from the official WildVision-Bench GitHub answers file (data/vision_bench_0617/model_answers/claude-3-sonnet-20240229.jsonl); present on 500/500 test_0617 rows and 191/500 test_0701 rows (rows absent from the 0617 answer set omit the field)."
          }
        },
        "source": {
          "format": "json",
          "url": {
            "test_0701": "https://huggingface.co/datasets/WildVision/wildvision-bench",
            "test_0617": "https://huggingface.co/datasets/WildVision/wildvision-bench"
          }
        }
      }
    }
  }
}