VSR / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
07c9a71 verified
Raw
History Blame Contribute Delete
2.68 kB
{
"name": "VSR",
"release_date": "2023-05-04",
"subsets": {
"zeroshot": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "yes_no_qa",
"score_pipeline": [
"rule-match",
"llm-match"
],
"score_protocol": {
"reference": "VLMEvalKit@vlmeval/dataset/image_yorn.py:19-21,41-103 — VSR-zeroshot graded via YOrN pipeline: rule extraction (utils/yorn.py:254-261 YOrN_Extraction) then GPT extraction fallback for 'Unknown' (yorn.py:264-272 YOrN_auxeval), exact compare to Yes/No; VSR_rating (yorn.py:190-210) reports acc plus precision/recall/F1.",
"note": "Original VSR repo (cambridgeltl/visual-spatial-reasoning) evaluates finetuned classifiers by label probability — no official generative-VLM protocol exists, so VLMEvalKit's YORN pipeline is the de facto standard (tier 2). Aggregation extras (precision/recall/F1 over Yes-class) not reproduced per-sample; headline is accuracy. mm-eval gt stored as Yes/No strings with numeric label in extra (rows verified)."
},
"prompt_template": "<image>{{ question }} Please answer yes or no. Please answer strictly Yes or No.",
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_yorn.py (VSR-zeroshot Y/N TSV; question = '<caption> Please answer yes or no. Please answer strictly Yes or No.')",
"notes": "Tier 3: VLMEvalKit ImageYORNDataset's pre-rendered Y/N prompt format for VSR-zeroshot (canonical). Caption used directly as question; answer derived from VSR label (1=Yes, 0=No)."
},
"mapping_from_source": {
"media": {
"from": "media",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "id"
},
"question": {
"from": "caption"
},
"answer": {
"from": "answer",
"optional": true
},
"extra": {
"label": {
"from": "label"
},
"relation": {
"from": "relation"
},
"subj": {
"from": "subj"
},
"obj": {
"from": "obj"
},
"source_image": {
"from": "source_image"
},
"image_link": {
"from": "image_link"
}
},
"source": {
"format": "huggingface",
"url": {
"test": "https://huggingface.co/datasets/cambridgeltl/vsr_zeroshot"
}
}
}
}
}
}