| { | |
| "name": "ScreenQA", | |
| "release_date": "2026-05-25", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "short_answer_qa", | |
| "score_pipeline": [ | |
| "exact-match", | |
| "rule-match" | |
| ], | |
| "score_params": { | |
| "string_match": "exact" | |
| }, | |
| "score_protocol": { | |
| "reference": "Official google-research-datasets/screen_qa README + code/metrics.py (SQA-Short task): Exact Match and SQuAD-style token F1 against multiple ground-truth answers with SQuAD preprocessing, averaged over samples (also ScreenQA paper arXiv:2209.08199 §metrics). Neither lmms-eval nor VLMEvalKit implements ScreenQA (re-confirmed 2026-07-07: no screenqa task dir in either local clone; only screenspot* grounding tasks exist in lmms-eval).", | |
| "note": "Official headline reports BOTH EM and SQuAD F1; per-sample binary exact matching reproduces EM only — the fractional token-F1 aggregate is not reproduced. Official EM applies SQuAD normalization (lowercase, strip articles/punctuation), which is looser than the scorer's default exact matcher. Published copy is the third-party mirror rootsautomation/RICO-ScreenQA-Short (not the official google-research-datasets release); multi-reference answer lists are preserved. 597/8427 test rows (7.1%) have the single reference '<no answer>'." | |
| }, | |
| "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": "screen_id" | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "ground_truth", | |
| "optional": true | |
| }, | |
| "source": { | |
| "format": "huggingface", | |
| "url": { | |
| "train": "https://huggingface.co/datasets/rootsautomation/RICO-ScreenQA-Short", | |
| "validation": "https://huggingface.co/datasets/rootsautomation/RICO-ScreenQA-Short", | |
| "test": "https://huggingface.co/datasets/rootsautomation/RICO-ScreenQA-Short" | |
| } | |
| } | |
| }, | |
| "prompt_template_source": { | |
| "origin": "fallback", | |
| "reference": "T1", | |
| "notes": "Tier 5: no higher tier publishes a prompt. T1: github.com/google-research-datasets/screen_qa publishes data + metrics-only code/metrics.py (grepped: no prompt strings); T2: no prompt column in rootsautomation/RICO-ScreenQA-Short; T3: no VLMEvalKit screenqa; T4: no lmms-eval tasks/screenqa (clone @047ec52; screenspot* tasks are a different benchmark). Template is the canonical single-image short-answer fallback." | |
| } | |
| } | |
| } | |
| } |