| { | |
| "name": "WHOOPS!", | |
| "release_date": "2023-03-13", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "short_answer_qa", | |
| "score_pipeline": [ | |
| "exact-match", | |
| "rule-match", | |
| "llm-judge" | |
| ], | |
| "score_protocol": { | |
| "reference": "WHOOPS! paper (Bitton-Guetta et al., ICCV 2023; arxiv 2303.07274) — the VQA sub-task is graded with BEM (BERT-based answer-equivalence classifier judging prediction vs reference correct/incorrect), reported alongside exact match. mm-eval ships the VQA sub-task (WHOOPS.json explodes `qa` into question/answer).", | |
| "note": "Official headline metric is BEM (learned BERT answer-equivalence), not a rule or an LLM judge. Modeled as rule_llm_judge (exact match short-circuit + LLM judge for semantic equivalence) as the closest reproducible analog; results will differ numerically from BEM. WHOOPS also has captioning/explanation/matching sub-tasks not shipped here." | |
| }, | |
| "prompt_template": "<image>{{ question }}\nAnswer the question using a single word or phrase.", | |
| "mapping_from_source": { | |
| "media": { | |
| "from": "image", | |
| "type": "list", | |
| "min_items": 1, | |
| "max_items": 1 | |
| }, | |
| "id": { | |
| "from": "id", | |
| "explode_from": "qa", | |
| "id_template": "{id}_q{idx}" | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "answer", | |
| "optional": true | |
| }, | |
| "extra": { | |
| "commonsense_category": { | |
| "from": "commonsense_category" | |
| }, | |
| "designer_explanation": { | |
| "from": "designer_explanation" | |
| } | |
| }, | |
| "source": { | |
| "format": "huggingface", | |
| "url": { | |
| "test": "https://huggingface.co/datasets/nlphuji/whoops" | |
| } | |
| } | |
| }, | |
| "prompt_template_source": { | |
| "origin": "fallback", | |
| "reference": "T1", | |
| "notes": "Tier 5: no higher tier publishes a prompt. T1: whoops-benchmark.github.io and the ICCV 2023 paper publish no reusable model-input prompt (paper VQA numbers use BLIP2 whose 'Question: {} Short answer:' is BLIP2's own convention, not a WHOOPS-authored prompt); T2: no prompt column in nlphuji/whoops; T3: no VLMEvalKit whoops; T4: no lmms-eval tasks/whoops (clone @047ec52). Template is the canonical single-image short-answer fallback." | |
| } | |
| } | |
| } | |
| } |