| { |
| "name": "IconQA", |
| "release_date": null, |
| "subsets": { |
| "choose_img": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "multi_image_start" |
| ], |
| "task_type": "multiple_choice_qa", |
| "score_pipeline": [ |
| "exact-match", |
| "rule-match" |
| ], |
| "score_protocol": { |
| "reference": "official@github.com/lupantech/IconQA run_choose_img/eval.py — plain accuracy of the selected choice, no LLM. Cross-check: lmms-eval@lmms_eval/tasks/iconqa/_default_template_docvqa_yaml + utils.py:49-53 — generative eval with metric anls (no LLM at any step).", |
| "note": "lmms-eval scores IconQA with ANLS (an artifact of reusing the DocVQA task template) instead of the official accuracy; for letter answers ANLS~exact match. Official protocol is deterministic accuracy — classified rule. Published choose_img rows give choices ['A','B'] (letter labels referring to candidate images) and a letter gold answer. WARNING: ~15% of published choose_img rows have an EMPTY gold answer (sampled 2026-07-07 via datasets-server: 48/300 val rows and 45/300 test rows with answer='') — these rows are ungradable as published." |
| }, |
| "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}\nAnswer with the option letter (A, B, C, ...).", |
| "mapping_from_source": { |
| "media": { |
| "from": "image", |
| "type": "list" |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer" |
| }, |
| "options": { |
| "from": "options" |
| }, |
| "extra": { |
| "n_images": { |
| "from": "n_images" |
| }, |
| "grade": { |
| "from": "grade" |
| }, |
| "skills": { |
| "from": "skills" |
| } |
| }, |
| "source": { |
| "format": "huggingface", |
| "url": { |
| "val": "https://huggingface.co/datasets/lmms-lab/ICON-QA", |
| "test": "https://huggingface.co/datasets/lmms-lab/ICON-QA" |
| } |
| } |
| }, |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/iconqa/utils.py (iconqa choose_img task; multi-image with letter-option suffix 'Answer with the option letter (A, B, C, ...)')", |
| "notes": "Tier 4: lmms-eval IconQA choose_img canonical evaluation prompt." |
| } |
| }, |
| "choose_txt": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "single_image_start" |
| ], |
| "task_type": "multiple_choice_qa", |
| "score_pipeline": [ |
| "exact-match", |
| "rule-match" |
| ], |
| "score_protocol": { |
| "reference": "official@github.com/lupantech/IconQA run_choose_txt/eval.py — plain accuracy, no LLM. Cross-check: lmms-eval@lmms_eval/tasks/iconqa (ANLS metric, rule-only).", |
| "note": "lmms-eval scores ANLS due to the reused DocVQA template; official metric is accuracy. Published rows: options dict A-E, letter gold answer." |
| }, |
| "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.", |
| "mapping_from_source": { |
| "media": { |
| "from": "image", |
| "type": "list" |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer" |
| }, |
| "options": { |
| "from": "options" |
| }, |
| "extra": { |
| "grade": { |
| "from": "grade" |
| }, |
| "skills": { |
| "from": "skills" |
| } |
| }, |
| "source": { |
| "format": "huggingface", |
| "url": { |
| "val": "https://huggingface.co/datasets/lmms-lab/ICON-QA", |
| "test": "https://huggingface.co/datasets/lmms-lab/ICON-QA" |
| } |
| } |
| }, |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/iconqa/_default_template_docvqa_yaml (iconqa choose_txt task; canonical MCQ trailer)", |
| "notes": "Tier 4: lmms-eval IconQA choose_txt canonical evaluation prompt." |
| } |
| }, |
| "fill_in_blank": { |
| "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@github.com/lupantech/IconQA run_fill_in_blank/eval.py — accuracy by exact match of the short answer, no LLM (compute_test_acc: pred_ans == gt_ans over an answer vocabulary, eval.py:31-45). Cross-check: lmms-eval@lmms_eval/tasks/iconqa/utils.py:49-53 (ANLS on answer strings, rule-only).", |
| "note": "lmms-eval scores ANLS (DocVQA-template artifact); official is exact-match accuracy. Published answers are short numerics/words ('3', '4', '5')." |
| }, |
| "prompt_template": "<image>{{ question }}\nAnswer the question using a single word or phrase.", |
| "mapping_from_source": { |
| "media": { |
| "from": "image", |
| "type": "list" |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer" |
| }, |
| "extra": { |
| "grade": { |
| "from": "grade" |
| }, |
| "skills": { |
| "from": "skills" |
| } |
| }, |
| "source": { |
| "format": "huggingface", |
| "url": { |
| "val": "https://huggingface.co/datasets/lmms-lab/ICON-QA", |
| "test": "https://huggingface.co/datasets/lmms-lab/ICON-QA" |
| } |
| } |
| }, |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/iconqa/_default_template_yaml (iconqa fill_in_blank — open-ended short-answer with 'Answer the question using a single word or phrase.' trailer)", |
| "notes": "Tier 4: lmms-eval IconQA fill_in_blank canonical evaluation prompt — VLMEvalKit-style short-answer trailer." |
| } |
| } |
| } |
| } |
