{
  "name": "MaRVL",
  "release_date": "2021-09-29",
  "subsets": {
    "main": {
      "language": [
        "en"
      ],
      "modalities": [
        "multi_image_interleave"
      ],
      "task_type": "yes_no_qa",
      "score_pipeline": [
        "exact-match",
        "rule-match"
      ],
      "score_protocol": {
        "reference": "Official: Liu et al., EMNLP 2021 (MaRVL paper, arXiv 2109.13238 — 'the task consists of discriminating whether each grounded statement is true or false', scored by accuracy; abstract re-verified 2026-07-07) — eval code in github.com/marvl-challenge/marvl-code (experiments/ per-model, per-language test scripts; repo tree re-verified 2026-07-07). Deterministic label comparison, no LLM.",
        "note": "Published ground truth is a JSON BOOLEAN (true/false), not a string — must be normalized to 'True'/'False' to match the prompt's requested output ('Answer with only \"True\" or \"False\"'). Official reporting is per-language accuracy; the org ships 5 language HF splits (id/sw/ta/tr/zh) under one subset. Metadata's cited lmms-eval tasks/marvl/utils.py does not exist in current lmms-eval main."
      },
      "prompt_template": "<image><image>Statement: {{ question }}\nIs the statement True or False based on the two images? Answer with only \"True\" or \"False\".",
      "mapping_from_source": {
        "media": {
          "from": "images",
          "type": "list",
          "min_items": 2,
          "max_items": 2
        },
        "id": {
          "from": "id"
        },
        "question": {
          "from": "hypothesis"
        },
        "answer": {
          "from": "label",
          "optional": true
        },
        "extra": {
          "hypothesis_en": {
            "from": "hypothesis_en"
          },
          "language": {
            "from": "language"
          },
          "chapter": {
            "from": "chapter"
          },
          "concept": {
            "from": "concept"
          },
          "left_img_id": {
            "from": "left_img_id"
          },
          "right_img_id": {
            "from": "right_img_id"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "id": "https://huggingface.co/datasets/floschne/marvl",
            "sw": "https://huggingface.co/datasets/floschne/marvl",
            "ta": "https://huggingface.co/datasets/floschne/marvl",
            "tr": "https://huggingface.co/datasets/floschne/marvl",
            "zh": "https://huggingface.co/datasets/floschne/marvl"
          }
        }
      },
      "prompt_template_source": {
        "origin": "fallback",
        "reference": "T-yesno-adapted (statement True/False verification, mm-eval convention)",
        "notes": "Tier 5: MaRVL's own eval targets finetuned classifiers (no model-input prompt); no VLMEvalKit/lmms-eval task exists. The third-party M5B suite publishes a Yes/No variant (github.com/floschne/m5b src/m5b/data/lit_marvl_wds.py#L43) but M5B is not an established tier source and its Yes/No format conflicts with the boolean True/False gt shipped here — kept the deployed True/False template as an explicit fallback. Fixed 2026-07-07 (previous citation pointed at a nonexistent lmms-eval task)."
      }
    }
  }
}