LEGO-Puzzles / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
3daa3a8 verified
Raw
History Blame Contribute Delete
3.14 kB
{
"name": "LEGO-Puzzles",
"release_date": "2025-03-25",
"subsets": {
"default": {
"language": [
"en"
],
"modalities": [
"multi_image_start"
],
"task_type": "multiple_choice_qa",
"score_pipeline": [
"rule-match",
"llm-match"
],
"score_protocol": {
"reference": "VLMEvalKit@vlmeval/dataset/image_mcq.py:1614-1695 (class LEGO(ImageMCQDataset), official harness of the authors) + vlmeval/dataset/utils/multiple_choice.py:359-406 (extract_answer_from_item uses can_infer_lego rule prefetch, GPT fallback via build_prompt_LEGO:321-344) + vlmeval/utils/matching_util.py:130-136 (can_infer_lego: question_type 'sort' -> permutation-sequence inference, else option-letter inference).",
"note": "Mixed row types: ~10/11 tasks are single-letter MCQ (question_type 'multiple-choice'), the Ordering task rows are question_type 'sort' with a 4-letter permutation answer (e.g. 'BDAC') compared as a full sequence. Published prompt_template appends \"Answer with the option's letter from the given choices directly.\" to ALL rows, whereas VLMEvalKit uses a sort-specific instruction (\"respond with only the sequence of letters (e.g., 'BDAC')\") for sort rows — published sort prompts diverge from official. Published per-row question_type values ('multiple-choice'/'sort') are outside the scorer vocabulary."
},
"prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}\nA. {{ options.A }}\nB. {{ options.B }}\nC. {{ options.C }}{% if options.D %}\nD. {{ options.D }}{% endif %}{% if options.E %}\nE. {{ options.E }}{% endif %}\nAnswer with the option's letter from the given choices directly.",
"prompt_template_source": {
"origin": "official",
"reference": "https://huggingface.co/datasets/KexianTang/LEGO-Puzzles",
"notes": "Tier 1: LEGO-Puzzles MCQ A-E (D/E often empty). Image column has mixed encoding: single rows store bare base64, multi rows store stringified list; preprocessor handles both."
},
"mapping_from_source": {
"media": {
"from": "images",
"type": "list",
"min_items": 1,
"max_items": 8
},
"id": {
"from": "id"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true
},
"extra": {
"n_images": {
"from": "n_images"
},
"category": {
"from": "category"
},
"question_type": {
"from": "question_type"
},
"A": {
"from": "A"
},
"B": {
"from": "B"
},
"C": {
"from": "C"
},
"D": {
"from": "D"
},
"E": {
"from": "E"
}
},
"source": {
"format": "json",
"url": {
"train": "https://huggingface.co/datasets/KexianTang/LEGO-Puzzles"
}
}
}
}
}
}