{
  "name": "POPE",
  "release_date": "2026-05-15",
  "subsets": {
    "main": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "yes_no_qa",
      "score_pipeline": [
        "exact-match",
        "rule-match"
      ],
      "score_protocol": {
        "reference": "lmms-eval@lmms_eval/tasks/pope/utils.py:17-29 — pred.lower().strip() exact-compared to gt yes/no, no LLM; aggregations utils.py:32-90 compute accuracy, precision, recall, F1, yes-ratio.",
        "note": "Official headline is accuracy+precision+recall+F1 reported per category (adversarial/popular/random, carried in extra.category); per-sample accuracy reproduces only accuracy — F1/precision/recall must be recomputed at aggregation (VLMEvalKit yorn.py:148-187 POPE_rating explodes per category). VLMEvalKit variant adds GPT yes/no extraction fallback (yorn.py:254-272 YOrN_Extraction + YOrN_auxeval); lmms-eval/official use pure rules — rule chosen per official protocol."
      },
      "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": {
          "category": {
            "from": "category"
          }
        },
        "source": {
          "format": "json",
          "url": {
            "test": "https://huggingface.co/datasets/lmms-lab/POPE"
          }
        }
      },
      "prompt_template_source": {
        "origin": "official",
        "reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_yorn.py (ImageYORNDataset.build_prompt — bare yes/no question)",
        "notes": "Tier 3: VLMEvalKit POPE ImageYORNDataset.build_prompt: bare yes/no question."
      }
    }
  }
}