| { |
| "name": "EMMA", |
| "release_date": "2026-07-07", |
| "subsets": { |
| "main": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "multi_image_interleave", |
| "text" |
| ], |
| "task_type": "multiple_choice_qa", |
| "score_pipeline": [ |
| "exact-match", |
| "rule-match", |
| "llm-judge" |
| ], |
| "score_protocol": { |
| "reference": "official EMMA-Bench/EMMA@evaluation/evaluate.py — rule-based fast_extract_answer (letter/number/boxed/'answer is' patterns) + is_equal compare by default; optional --gpt_eval mode where GPT-4o judges and 'correct' in the response decides the score; evaluation/calculate_acc.py aggregates. lmms-eval@lmms_eval/tasks/emma/utils.py:159-200 (LLM judge with fast_extract_answer+is_equal fallback, emma_all.yaml use_lmms_judge: True).", |
| "note": "Framework implementations lean judge-heavier than the official default: lmms-eval judges every sample (use_lmms_judge), VLMEvalKit EMMADataset inherits ImageShortQADataset's all-LLM binary judge (image_shortqa.py:15-48). Classified rule_llm_judge: deterministic extraction resolves direct letter/short answers (the published prompt requests direct answers), LLM resolves the remainder — consistent with official fast-eval + LLMs-eval pairing. EMMA is mixed MCQ/open (options optional); task_type multiple_choice as majority format (2002/2788 rows are Multiple Choice)." |
| }, |
| "prompt_template": "{% if context %}{{ context }}\n{% endif %}{{ question }}{% if options %}\n\n{% for k, v in options.items() %}({{ k }}) {{ v }}\n{% endfor %}\nAnswer with the option's letter from the given choices directly.{% else %}\n\nAnswer the question using a single word or phrase.{% endif %}", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://huggingface.co/datasets/luckychao/EMMA; context+question+options structure per official EMMA-Bench/EMMA@data_utils.py L25-52 (build_query: '{context}\\n{question}\\n{options}\\n...') and lmms-eval@lmms_eval/tasks/emma/utils.py L27-31", |
| "notes": "Tier 1: EMMA uses MMMU-style <image_N> markers in question/options; MCQ letter-answer convention. Re-conversion 2026-07-07: restored the options block data (options were dropped in the previous upload) and prepended '{% if context %}{{ context }}\\n{% endif %}' because the official build_query renders context before the question (57 Physics rows carry a non-empty context block of physical constants that the previous upload dropped). The direct-answer trailers are kept from the previously published template; the official EMMA eval prompt additionally requests \\boxed{} + CoT (configs/gpt.yaml) — intentional deviation preserved from the published copy." |
| }, |
| "mapping_from_source": { |
| "source": { |
| "format": "json", |
| "url": { |
| "test": "https://huggingface.co/datasets/luckychao/EMMA" |
| } |
| }, |
| "id": { |
| "from": "id", |
| "note": "source column 'pid' (e.g. Math_1, chem_1, coding_1, phy_1), copied verbatim during pre-download." |
| }, |
| "question": { |
| "from": "question", |
| "note": "<image_N> refs normalized to the canonical <image> placeholder; referenced images materialized into media in textual order." |
| }, |
| "options": { |
| "from": "options", |
| "optional": true, |
| "note": "list source values are normalized to {A,B,...} dict. 301 rows reference images inside option strings via <image_N>; those refs are normalized to <image> and the referenced images are appended to media in option order (after question refs), so placeholder count == media count. 4 rows repeat a ref (chem_115, chem_139, Math_105, Math_544) — the repeated image is duplicated in media, whereas the official pipeline deduplicates and sends it once." |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "media": { |
| "from": "media", |
| "type": "list", |
| "min_items": 0, |
| "max_items": 5, |
| "note": "materialized during pre-download from source columns image_1..image_5, one entry per <image_N> occurrence across context/question/options in template render order. chem_146 references no images anywhere (its 2 source images are unreferenced) and is a text-only row with media=[], mirroring the official pipeline which sends no image for it." |
| }, |
| "extra": { |
| "context": { |
| "from": "context" |
| }, |
| "type": { |
| "from": "type" |
| }, |
| "subject": { |
| "from": "subject" |
| }, |
| "task": { |
| "from": "task" |
| }, |
| "category": { |
| "from": "category" |
| }, |
| "source_dataset": { |
| "from": "source_dataset" |
| }, |
| "solution": { |
| "from": "solution" |
| } |
| } |
| } |
| } |
| } |
| } |
