SeePhys / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
8a936c7 verified
Raw
History Blame Contribute Delete
2.78 kB
{
"name": "SeePhys",
"release_date": "2025-05-15",
"subsets": {
"default": {
"language": [
"en"
],
"modalities": [
"multi_image_start"
],
"task_type": "short_answer_qa",
"score_pipeline": [
"exact-match",
"rule-match",
"llm-judge"
],
"score_protocol": {
"reference": "VLMEvalKit@vlmeval/dataset/utils/seephys.py:187-260 — post_check rule prefetch (extract_by_rule + sympy quick_compare, abs tol 1e-6) scores 1 without any LLM; otherwise LLM answer extraction then LLM binary 0/1 judgement (score_func with prompt_scoring); default judge 'deepseek' (image_vqa.py:1658). Official seephys-project repo delegates evaluation to this VLMEvalKit integration (--judge deepseek, valid_type LLM); lmms-eval@lmms_eval/tasks/seephys/seephys_utils.py:62-110 mirrors (quick_extract else LLM judge). [Re-verified in clone 2026-07-07: post_check at seephys.py:191-206, extract/score_func at :212-260, default model 'deepseek' at image_vqa.py:1658, seephys_process_results at seephys_utils.py:66-110.]",
"note": "Official judge model is deepseek. Official prompt appends a CoT <think>/<answer> instruction and a per-row significant-figures requirement (sig_figs shipped in extra); the mm-eval prompt template is bare images+question without those instructions. Numeric prefetch tolerance is absolute 1e-6 (not relative)."
},
"prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}",
"prompt_template_source": {
"origin": "official",
"reference": "https://huggingface.co/datasets/SeePhys/SeePhys",
"notes": "Tier 1: SeePhys open-ended physics VQA; variable image count per row (1-4), single-prompt eval format."
},
"mapping_from_source": {
"media": {
"from": "images",
"type": "list",
"min_items": 1,
"max_items": 4
},
"id": {
"from": "id"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true
},
"extra": {
"n_images": {
"from": "n_images"
},
"reasoning": {
"from": "reasoning"
},
"sig_figs": {
"from": "sig_figs"
},
"level": {
"from": "level"
},
"subject": {
"from": "subject"
},
"language_field": {
"from": "language_field"
}
},
"source": {
"format": "json",
"url": {
"train": "https://huggingface.co/datasets/SeePhys/SeePhys"
}
}
}
}
}
}