| { |
| "name": "SAT", |
| "release_date": "2025-12-04", |
| "subsets": { |
| "real": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "multi_image_start" |
| ], |
| "task_type": "multiple_choice_qa", |
| "score_pipeline": [ |
| "exact-match", |
| "rule-match" |
| ], |
| "score_protocol": { |
| "reference": "lmms-eval@lmms_eval/tasks/sat/sat.yaml:17-31 + utils.py (MultiChoiceRegexFilter) — deterministic filter maps the response to a choice letter by regex / choice-text matching, then exact_match; no LLM at any step. Official repo (github.com/arijitray1993/SAT) has not published a dedicated scoring script (README: 'work in progress').", |
| "note": "lmms-eval letters the options (A., B.) and grades letters; the mm-eval copy renders un-lettered choice texts ('Choose between the following options: X, or Y.') and stores the full choice text as answer, so grading is deterministic matching of the chosen option text rather than a letter compare. Rows carry extra field 'question_type' with SAT category names (obj_movement, action_sequence, ...) that are NOT scorer vocabulary — the scorer must ignore them (unrecognized values fall through to gt-shape per spec)." |
| }, |
| "prompt_template": "{% for i in range(n_images) %}<image>{% endfor %}{{ question }} Choose between the following options: {% for c in choices %}{{ c }}{% if not loop.last %}, or {% endif %}{% endfor %}.", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://arxiv.org/pdf/2412.07755 (SAT paper, Appendix A.3, Question-answer format, p.21)", |
| "notes": "Tier 1: SAT paper Appendix specifies the binary-choice text format with <image> tokens at start; LLaVA conversation scaffolding is dropped (handled by inference backend)." |
| }, |
| "mapping_from_source": { |
| "media": { |
| "from": "media", |
| "type": "list", |
| "min_items": 1, |
| "max_items": 2 |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "choices": { |
| "from": "choices" |
| }, |
| "extra": { |
| "question_type": { |
| "from": "question_type" |
| }, |
| "n_images": { |
| "from": "n_images" |
| } |
| }, |
| "source": { |
| "format": "huggingface", |
| "url": { |
| "real_test": "https://huggingface.co/datasets/array/SAT" |
| } |
| } |
| } |
| }, |
| "synthetic": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "multi_image_start" |
| ], |
| "task_type": "multiple_choice_qa", |
| "score_pipeline": [ |
| "exact-match", |
| "rule-match" |
| ], |
| "score_protocol": { |
| "reference": "lmms-eval@lmms_eval/tasks/sat/sat.yaml:17-31 + utils.py (MultiChoiceRegexFilter) — same deterministic choice-matching protocol as the real subset; no LLM.", |
| "note": "Same as SAT/real: full-text answers instead of letters; extra 'question_type' field carries SAT category names, not scorer vocabulary. lmms-eval's sat task evaluates the (nv-njb/SAT) test set; the org ships real (150 rows) and synthetic (4001 rows) as separate configs from array/SAT." |
| }, |
| "prompt_template": "{% for i in range(n_images) %}<image>{% endfor %}{{ question }} Choose between the following options: {% for c in choices %}{{ c }}{% if not loop.last %}, or {% endif %}{% endfor %}.", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://arxiv.org/pdf/2412.07755 (SAT paper, Appendix A.3, Question-answer format, p.21)", |
| "notes": "Tier 1: SAT paper Appendix specifies the binary-choice text format with <image> tokens at start; LLaVA conversation scaffolding is dropped (handled by inference backend)." |
| }, |
| "mapping_from_source": { |
| "media": { |
| "from": "media", |
| "type": "list", |
| "min_items": 1, |
| "max_items": 2 |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "choices": { |
| "from": "choices" |
| }, |
| "extra": { |
| "question_type": { |
| "from": "question_type" |
| }, |
| "n_images": { |
| "from": "n_images" |
| } |
| }, |
| "source": { |
| "format": "huggingface", |
| "url": { |
| "synthetic_test": "https://huggingface.co/datasets/array/SAT" |
| } |
| } |
| } |
| } |
| } |
| } |
