{
  "name": "NLVR2",
  "release_date": "2026-05-25",
  "subsets": {
    "main": {
      "language": [
        "en"
      ],
      "modalities": [
        "multi_image_start"
      ],
      "task_type": "yes_no_qa",
      "score_pipeline": [
        "exact-match",
        "rule-match"
      ],
      "score_protocol": {
        "reference": "official@github.com/lil-lab/nlvr nlvr2/eval/metrics.py — deterministic comparison of predicted True/False labels against ground truth ('prediction.lower() == example[\"label\"].lower()'); reports accuracy and consistency; no LLM. (File re-verified on GitHub master 2026-07-07.)",
        "note": "Ground truth is 'True'/'False' strings (statement verification), prompt appends 'Answer with True or False.'. Official also reports CONSISTENCY (all image-pairs of the same sentence correct) and per-split (balanced/unbalanced) accuracies — the org ships 6 HF splits (balanced/unbalanced x dev/test_public/test_unseen); headline is usually test_public accuracy. Per-sample scoring reproduces accuracy only. Metadata's cited lmms-eval tasks/nlvr2/utils.py does not exist in current lmms-eval main."
      },
      "prompt_template": "<image><image>{{ question }}\nAnswer with True or False.",
      "mapping_from_source": {
        "media": {
          "from": "images",
          "type": "list",
          "min_items": 2,
          "max_items": 2
        },
        "id": {
          "from": "id"
        },
        "question": {
          "from": "question"
        },
        "answer": {
          "from": "answer",
          "optional": true
        },
        "extra": {
          "synset": {
            "from": "synset"
          },
          "writer": {
            "from": "writer"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "balanced_dev": "https://huggingface.co/datasets/lmms-lab/NLVR2",
            "balanced_test_public": "https://huggingface.co/datasets/lmms-lab/NLVR2",
            "balanced_test_unseen": "https://huggingface.co/datasets/lmms-lab/NLVR2",
            "unbalanced_dev": "https://huggingface.co/datasets/lmms-lab/NLVR2",
            "unbalanced_test_unseen": "https://huggingface.co/datasets/lmms-lab/NLVR2",
            "unbalanced_test_public": "https://huggingface.co/datasets/lmms-lab/NLVR2"
          }
        }
      },
      "prompt_template_source": {
        "origin": "fallback",
        "reference": "T2-adapted",
        "notes": "Tier 5: no higher tier publishes a generative prompt. T1: lil-lab/nlvr is pre-VLM (data + accuracy/consistency scripts only); lmms-lab/NLVR2 HF card states no prompt. T2: no prompt column. T3: no VLMEvalKit nlvr2. T4: lmms-eval never had a standalone nlvr2 task (code+commit search; only NLVR2_Mantis inside llava_interleave_bench with the prompt baked into the data). Template = two <image> tags + bare statement + mm-eval convention trailer 'Answer with True or False.' (T2-adapted)."
      }
    }
  }
}