Geometry3K / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
c4c6433 verified
Raw
History Blame Contribute Delete
2.87 kB
{
"name": "Geometry3K",
"release_date": "2021-05-30",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "short_answer_qa",
"score_pipeline": [],
"score_protocol": {
"reference": "Verified absence, checked 2026-07-07: the published copy is row-identical to hiyouga/geometry3k (matching problem/answer text and split sizes 2101/300/601), an RL-TRAINING reformat 'converted from https://github.com/lupantech/InterGPS' whose dataset card publishes NO eval protocol or metric; it is the tutorial training set of EasyR1 (hiyouga/EasyR1 README: 'Run Qwen2.5-VL GRPO on Geometry3K Dataset'), where rule-based mathruler boxed-extraction reward is a training signal, not an eval protocol. The official Inter-GPS (ACL 2021) protocol and lmms-eval geometry3k (geometry3k.yaml: dataset_path Yang130/geometry3k_4choices_mixed; utils.py:66-125 A-D letter extraction) both grade 4-choice MCQ variants and are inapplicable to this free-form copy.",
"note": "No official grading protocol exists for this variant; the scorer's default exact,template chain applies and results must be stamped non-official (config.official_protocol=false). Even then, LaTeX ground truths ('2 \\sqrt { 221 }') need math-expression equivalence to grade correctly — exact/template matching under-scores them. This is an RL training dataset (train/validation/test), not an eval-only benchmark; numbers are not comparable to official Inter-GPS Geometry3K results."
},
"prompt_template": "{{ question }}",
"mapping_from_source": {
"media": {
"from": "images",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "id"
},
"question": {
"from": "problem"
},
"answer": {
"from": "answer",
"optional": true
},
"source": {
"format": "json",
"url": {
"train": "https://huggingface.co/datasets/hiyouga/geometry3k",
"validation": "https://huggingface.co/datasets/hiyouga/geometry3k",
"test": "https://huggingface.co/datasets/hiyouga/geometry3k"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/geometry3k/utils.py (geometry3k_doc_to_text — bare question; choices and trailer baked into source question column)",
"notes": "Tier 4: lmms-eval Geometry3K canonical evaluation prompt. The question column already contains <image> placeholders and problem+choices+trailer baked in from the source dataset, so the template is bare {{ question }} to avoid duplicating the placeholder."
}
}
}
}