| { | |
| "name": "OCRBench", | |
| "release_date": "2026-07-07", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "short_answer_qa", | |
| "score_pipeline": [ | |
| "exact-match", | |
| "rule-match" | |
| ], | |
| "score_params": { | |
| "string_match": "contains" | |
| }, | |
| "score_protocol": { | |
| "reference": "VLMEvalKit@vlmeval/dataset/utils/ocrbench.py:8-42 and lmms-eval@lmms_eval/tasks/ocrbench/utils.py:33-66 (both mirroring official Yuliang-Liu/MultimodalOCR eval) — lowercase, newline-normalized substring containment of ANY of the ground-truth alternatives in the prediction (answer is a list; VLMEvalKit ocrbench.py:29-42, lmms-eval utils.py:39-62); HME100k category is case-sensitive with all spaces stripped.", | |
| "note": "Official headline is Final Score out of 1000 (sum of correct over 10 categories, VLMEvalKit ocrbench.py:44-60) = per-sample accuracy x1000. Re-converted 2026-07-07: answer now preserves the full official list of acceptable alternatives (148/1000 rows have >1; the earlier copy space-joined them into one unmatched string), and the official 10-way category (question_type) plus upstream source-dataset field are restored, enabling the per-category breakdown and the HME100k special case (source_dataset == 'HME100k': case-sensitive containment with all spaces stripped, lmms-eval utils.py:39-50) which the generic lowercase-contains rule does not itself apply." | |
| }, | |
| "prompt_template": "<image>{{ question }}", | |
| "prompt_template_source": { | |
| "origin": "official", | |
| "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/ocrbench/utils.py (ocrbench_doc_to_text — bare question; OCR-specific evaluation handled at scoring)", | |
| "notes": "Tier 4: lmms-eval OCRBench canonical evaluation prompt." | |
| }, | |
| "mapping_from_source": { | |
| "media": { | |
| "from": "image", | |
| "type": "list", | |
| "min_items": 1, | |
| "max_items": 1 | |
| }, | |
| "id": { | |
| "from": "id", | |
| "note": "Stringified 0-based row index of the official echo840/OCRBench test split (the source ships no id column); identical to the previously published ids." | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "answer", | |
| "optional": true, | |
| "note": "Kept as the full official list of acceptable alternatives (1-4 entries; 148/1000 rows have >1). Official graders score a hit when ANY alternative is contained in the prediction." | |
| }, | |
| "extra": { | |
| "question_type": { | |
| "from": "question_type" | |
| }, | |
| "source_dataset": { | |
| "from": "dataset" | |
| } | |
| }, | |
| "source": { | |
| "format": "json", | |
| "url": { | |
| "test": "https://huggingface.co/datasets/echo840/OCRBench" | |
| } | |
| } | |
| } | |
| } | |
| } | |
| } |