MaRVL / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
b0a72bb verified
Raw
History Blame Contribute Delete
3.29 kB
{
"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)."
}
}
}
}