| { | |
| "name": "SLAKE", | |
| "release_date": "2026-05-25", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "short_answer_qa", | |
| "score_pipeline": [ | |
| "exact-match", | |
| "rule-match", | |
| "llm-judge" | |
| ], | |
| "score_protocol": { | |
| "reference": "No dedicated lmms-eval or VLMEvalKit task on disk (image_shortqa.py grep for SLAKE = empty; confirmed absent). SLAKE official metric is med-VQA accuracy: closed-ended (yes/no) exact-match accuracy + open-ended answer accuracy. rows/SLAKE.json shows short closed and open answers ('MRI','Abdomen','T2'). Classified by analogy to VLMEvalKit ShortQA grader used for PathVQA (image_shortqa.py:87-142).", | |
| "note": "Official SLAKE = closed accuracy + open accuracy (exact/token match). Modeled as rule_llm_judge: exact match for closed yes/no, LLM judge for open medical answers. English subset only (mdwiratathya/SLAKE-vqa-english); official SLAKE is bilingual EN+ZH." | |
| }, | |
| "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": "__index__" | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "answer", | |
| "optional": true | |
| }, | |
| "source": { | |
| "format": "huggingface", | |
| "url": { | |
| "train": "https://huggingface.co/datasets/mdwiratathya/SLAKE-vqa-english", | |
| "validation": "https://huggingface.co/datasets/mdwiratathya/SLAKE-vqa-english", | |
| "test": "https://huggingface.co/datasets/mdwiratathya/SLAKE-vqa-english" | |
| } | |
| } | |
| }, | |
| "prompt_template_source": { | |
| "origin": "fallback", | |
| "reference": "T1", | |
| "notes": "Tier 5: no higher tier publishes a prompt. T1: SLAKE (2021) is classification-era (BAN/MEVF baselines), no generative prompt; LLaVA-Med (closest official-adjacent eval, model_vqa_med.py#L199-L223) feeds the raw question through its own conversation template — a harness convention, not a benchmark prompt spec. T2: no prompt column in mdwiratathya/SLAKE-vqa-english. T3: no VLMEvalKit slake dataset. T4: no lmms-eval tasks/slake (clone @047ec52). Template is the canonical single-image short-answer fallback." | |
| } | |
| } | |
| } | |
| } |