HRBench4K / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
eace796 verified
Raw
History Blame Contribute Delete
3.18 kB
{
"name": "HRBench4K",
"release_date": "2026-05-15",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "multiple_choice_qa",
"score_pipeline": [
"rule-match",
"llm-match"
],
"score_protocol": {
"reference": "vlmevalkit@vlmeval/dataset/image_mcq.py:1302-1364 HRBenchDataset.evaluate -> mcq_vanilla_eval (can_infer + GPT extractor; DEFAULT_JUDGE ['chatgpt-0125','gpt-4-0125'] @vlmeval/dataset/image_mcq.py:1303) + report_acc_hrbench @vlmeval/dataset/utils/hrbench.py:7-29 (per-cycle x category accuracy, averaged over 4 cycles). HR-Bench authors (DreamMr/HR-Bench) vendor VLMEvalKit, so this is tier-1. Cross-check: lmms-eval hrbench also GPT-extracts (lmms_eval/tasks/hrbench/utils.py:24)",
"note": "Official report averages accuracy per cycle_category (0-3) and per category (single/cross); the mm-eval copy ships all 4 rotations as rows (800 = 4 x 200, cycle_category field preserved in extra), so overall per-sample accuracy equals the official 'all' average (cycles partition the file); the per-category/per-cycle breakdown itself is not reproduced per-sample. Note this is NOT CircularEval — each rotation counts independently."
},
"prompt_template": "<image>Question: {{ question }}\nOptions:\nA. {{ options.A }}\nB. {{ options.B }}\nC. {{ options.C }}\nD. {{ options.D }}\nPlease select the correct answer from the options above. \n",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "index"
},
"question": {
"from": "question"
},
"answer": {
"from": "answer",
"optional": true
},
"extra": {
"category": {
"from": "category"
},
"A": {
"from": "A"
},
"B": {
"from": "B"
},
"C": {
"from": "C"
},
"D": {
"from": "D"
},
"cycle_category": {
"from": "cycle_category"
}
},
"source": {
"format": "json",
"url": {
"hrbench_4k": "https://huggingface.co/datasets/DreamMr/HR-Bench"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_mcq.py (ImageMCQDataset.build_prompt — Question:/Options:/A.B.C.D./'Please select the correct answer from the options above. ' trailer; HR-Bench authors' eval at https://github.com/DreamMr/HR-Bench/blob/main/hrbench/vlmeval/dataset/image_mcq.py inherits this via HRBenchDataset)",
"notes": "Tier 1/3: HR-Bench authors use VLMEvalKit's ImageMCQDataset.build_prompt; mm-eval prompt reproduces it byte-for-byte. lmms-eval's hrbench_doc_to_text uses a different trailer ('Answer the option letter directly.') and is not the canonical reference."
}
}
}
}