| { | |
| "name": "PuzzleVQA", | |
| "release_date": "2024-03-21", | |
| "subsets": { | |
| "default": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "multiple_choice_qa", | |
| "score_pipeline": [ | |
| "exact-match", | |
| "rule-match" | |
| ], | |
| "score_protocol": { | |
| "reference": "official@github.com/declare-lab/LLM-PuzzleTest PuzzleVQA/main.py — prompter.get_answer parses option from raw output, ExactScorer returns 1.0 iff pred == answer; the only model call in the loop is a second-stage re-prompt of the EVALUATED model itself (inference-side self-extraction), no external judge/extractor LLM. Cross-check: VLMEvalKit@vlmeval/dataset/utils/puzzlevqa.py:85-152 PuzzleVQA_acc — pure regex extraction ('answer: X', boxed, answer-flag phrases), unparseable -> 'Z', per-category accuracy.", | |
| "note": "Official prompting is two-stage (CoT, then the same model restates the answer); mm-eval prompt asks for the letter directly, so single-stage rule extraction is faithful at scoring time. VLMEvalKit aggregates per-puzzle-category accuracy (20 categories) alongside overall." | |
| }, | |
| "prompt_template": "<image>{{ question }}\n{% for k, v in options.items() %}{{ k }}. {{ v }}\n{% endfor %}Answer with the option's letter from the given choices directly.", | |
| "prompt_template_source": { | |
| "origin": "official", | |
| "reference": "https://huggingface.co/datasets/declare-lab/PuzzleVQA", | |
| "notes": "Tier 1: PuzzleVQA is single-image MCQ with options list; answer is the option text. We map answer to the letter for letter-eval convention." | |
| }, | |
| "mapping_from_source": { | |
| "media": { | |
| "from": "images", | |
| "type": "list", | |
| "min_items": 1, | |
| "max_items": 1 | |
| }, | |
| "id": { | |
| "from": "id" | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "answer", | |
| "optional": true | |
| }, | |
| "options": { | |
| "from": "options", | |
| "optional": true, | |
| "note": "list source values are normalized to {A,B,...} dict" | |
| }, | |
| "extra": { | |
| "n_images": { | |
| "from": "n_images" | |
| }, | |
| "category": { | |
| "from": "category" | |
| }, | |
| "answer_text": { | |
| "from": "answer_text" | |
| } | |
| }, | |
| "source": { | |
| "format": "json", | |
| "url": { | |
| "train": "https://huggingface.co/datasets/declare-lab/PuzzleVQA" | |
| } | |
| } | |
| } | |
| } | |
| } | |
| } |