{ "name": "Winoground", "release_date": "2026-07-07", "subsets": { "main": { "language": [ "en" ], "modalities": [ "single_image_start" ], "task_type": "yes_no_qa", "score_pipeline": [ "exact-match", "rule-match" ], "score_protocol": { "reference": "Official Winoground defines text/image/group scores over ITM score comparisons for contrastive models (Thrush et al. 2022, §4, eqs. 1-3); no official generative protocol exists — probes graded by deterministic yes/no match.", "note": "GENERATIVE ADAPTATION: official Winoground text/image/group scores are defined over ITM score preferences and cannot be computed from generative outputs. This copy ships 4 balanced yes/no probes per example (image_i x caption_j, gt Yes iff i==j; group_id preserved). A group-score analog = all 4 probes of a group_id correct. Do not compare accuracies to official Winoground scores." }, "prompt_template": "{{ question }}", "prompt_template_source": { "origin": "fallback", "reference": "T2", "notes": "Tier 5: no official generative prompt exists (Winoground is a contrastive ITM benchmark, paper §4; no VLMEvalKit/lmms-eval task). Bare-question fallback; the probe question carries the Yes/No instruction." }, "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 }, "extra": { "group_id": { "from": "group_id" }, "image_index": { "from": "image_index" }, "caption_index": { "from": "caption_index" }, "caption_0": { "from": "caption_0" }, "caption_1": { "from": "caption_1" }, "tag": { "from": "tag" }, "secondary_tag": { "from": "secondary_tag" }, "collapsed_tag": { "from": "collapsed_tag" }, "num_main_preds": { "from": "num_main_preds" } }, "source": { "format": "json", "url": { "test": "https://huggingface.co/datasets/facebook/winoground" } } } } } }