| { | |
| "name": "VisualMRC", | |
| "release_date": "2026-05-26", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "open_ended_qa", | |
| "score_pipeline": { | |
| "unsupported": "caption-metrics", | |
| "reason": "requires corpus-level caption metrics (CIDEr/BLEU/ROUGE/BERTScore); use the benchmark's own evaluator" | |
| }, | |
| "score_params": { | |
| "metrics": [ | |
| "rouge_l", | |
| "bleu4", | |
| "meteor", | |
| "cider" | |
| ] | |
| }, | |
| "score_protocol": { | |
| "reference": "Official VisualMRC paper (arXiv:2101.11272, AAAI 2021, 'Evaluation metrics' §: 'we used BLEU, METEOR, ROUGE-L, and CIDEr' following generative MRC/captioning; BERTScore additionally reported in result tables; ROUGE-L used for model selection). No lmms-eval or VLMEvalKit task exists (re-confirmed 2026-07-07: no visualmrc task in either local clone).", | |
| "note": "Generative reading-comprehension answers graded with captioning-style reference metrics over a single reference answer per question; paper declares no single headline — ROUGE-L listed first here since the paper selects models by ROUGE-L. Official tables also report BERTScore (now representable in the caption_metrics enum but kept out of params to match the paper's core four). metadata.json prompt citation of lmms_eval/tasks/visualmrc is bogus — must be corrected." | |
| }, | |
| "prompt_template": "<image>{{ question }}", | |
| "mapping_from_source": { | |
| "media": { | |
| "from": "image", | |
| "type": "list", | |
| "min_items": 1, | |
| "max_items": 1 | |
| }, | |
| "id": { | |
| "from": "id" | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "answer", | |
| "optional": true | |
| }, | |
| "extra": { | |
| "qa_idx": { | |
| "from": "qa_idx" | |
| }, | |
| "url": { | |
| "from": "url" | |
| }, | |
| "screenshot_filename": { | |
| "from": "screenshot_filename" | |
| }, | |
| "image_filename": { | |
| "from": "image_filename" | |
| } | |
| }, | |
| "source": { | |
| "format": "huggingface", | |
| "url": { | |
| "test": "https://huggingface.co/datasets/jeepliu/VisualMRC" | |
| } | |
| } | |
| }, | |
| "prompt_template_source": { | |
| "origin": "fallback", | |
| "reference": "T2", | |
| "notes": "Tier 5: no higher tier publishes a prompt. T1: nttmdlab-nlp/VisualMRC is a data-only repo (AAAI 2021; LayoutT5/LayoutBART fine-tuned seq2seq baselines, pre-LLM, no prompt template); T2: no prompt column in jeepliu/VisualMRC; T3: no VLMEvalKit visualmrc; T4: no lmms-eval tasks/visualmrc (clone @047ec52). Bare-question T2 chosen because answers are abstractive full sentences graded with reference text metrics — a short-answer cue would suppress the expected long-form output." | |
| } | |
| } | |
| } | |
| } |