| { | |
| "name": "SEED-Bench-H", | |
| "release_date": "2024-04-08", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "multiple_choice_qa", | |
| "score_pipeline": [ | |
| "rule-match", | |
| "llm-match" | |
| ], | |
| "score_protocol": { | |
| "reference": "Neither VLMEvalKit nor lmms-eval implements SEED-Bench-H; protocol inherited from the SEED-Bench family: official AILab-CVC/SEED-Bench eval.py/EVALUATION.md (likelihood ranking) and, for generation pipelines, the SEEDBench MCQ letter protocol (vlmevalkit@vlmeval/dataset/image_mcq.py:249-324 (ImageMCQDataset.evaluate_heuristic) -> vlmevalkit@vlmeval/dataset/utils/multiple_choice.py:475-508 mcq_vanilla_eval: can_infer rule prefetch, then GPT extractor (extract_answer_from_item @vlmeval/dataset/utils/multiple_choice.py:359-407) maps verbose output to a letter, rule-compare vs GT)", | |
| "note": "SEED-Bench authors' original protocol is generation-free answer-likelihood ranking ('for each choice ... compute the likelihood ... select the choice with the highest likelihood', AILab-CVC/SEED-Bench EVALUATION.md) — not reproducible in a generation pipeline. The published mm-eval copy uses the generation-based ecosystem protocol (lmms-eval prompt + letter compare / VLMEvalKit MCQ extraction), which is how all current leaderboards run SEED-Bench on API/chat models." | |
| }, | |
| "prompt_template": "<image>{{ question }}\nOptions:\n{% for k, v in options.items() %}{{ k }}. {{ v }}{% if not loop.last %}\n{% endif %}{% endfor %}\nAnswer with the option's letter from the given choices directly.", | |
| "mapping_from_source": { | |
| "media": { | |
| "from": "image", | |
| "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" | |
| }, | |
| "source": { | |
| "format": "json", | |
| "url": { | |
| "test": "https://huggingface.co/datasets/AILab-CVC/SEED-Bench-H" | |
| } | |
| } | |
| }, | |
| "prompt_template_source": { | |
| "origin": "official", | |
| "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/tree/main/lmms_eval/tasks/seedbench_2_plus (SEED-Bench-H follows SEEDBench-2-Plus / SEEDBench format)", | |
| "notes": "Tier 4: lmms-eval SEED-Bench-H canonical evaluation prompt (same SEEDBench format)." | |
| } | |
| } | |
| } | |
| } |