VQAv2 / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
a023503 verified
Raw
History Blame Contribute Delete
2.56 kB
{
"name": "VQAv2",
"release_date": "2026-05-26",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "short_answer_qa",
"score_pipeline": [
"vqa-accuracy"
],
"score_protocol": {
"reference": "lmms-eval lmms_eval/tasks/vqav2/utils.py:15-50 (vqav2_process_results) — official VQA accuracy: EvalAIAnswerProcessor punctuation/digit-article normalization on response and each reference, then per-annotator acc=min(1, #matches-among-other-9/3), averaged over the 10-answer list (lines 31-42).",
"note": "Per-sample scores are fractional (multiples of 1/3 capped at 1, averaged over 10 leave-one-out folds). `answer` carries the full 10-reference list from the source `answers` column (annotator dicts {answer, answer_confidence, answer_id} reduced to their `answer` strings, order preserved; answer_confidence/answer_id dropped). The majority-vote `multiple_choice_answer` is preserved as the pass-through field `mc_answer`."
},
"prompt_template": "<image>{{ question }}\nAnswer the question using a single word or phrase.",
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/vqav2/utils.py (vqav2_doc_to_text — canonical short-answer)",
"notes": "Tier 4: lmms-eval VQAv2 canonical evaluation prompt."
},
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "question_id"
},
"question": {
"from": "question"
},
"answer": {
"from": "answers",
"optional": true,
"note": "Source `answers` is a list of 10 {answer, answer_confidence, answer_id} dicts; reduced to the list of 10 `answer` strings (order preserved) during pre-processing."
},
"extra": {
"mc_answer": {
"from": "multiple_choice_answer"
},
"image_id": {
"from": "image_id"
},
"question_type": {
"from": "question_type"
},
"answer_type": {
"from": "answer_type"
}
},
"source": {
"format": "huggingface",
"url": {
"validation": "https://huggingface.co/datasets/lmms-lab/VQAv2"
}
}
}
}
}
}