{ "name": "MME-RealWorld-Lite", "release_date": "2026-05-15", "subsets": { "main": { "language": [ "en" ], "modalities": [ "single_image_start" ], "task_type": "multiple_choice_vqa", "score_type": "rule", "score_protocol": { "reference": "VLMEvalKit@vlmeval/dataset/image_mcq.py:1053-1103 — MMERealWorld.evaluate extracts the choice with extract_characters_regex (utils/multiple_choice.py:585, choices (A)-(E)) after optional ... unwrapping; unmatched predictions score 0; exact letter compare; NO LLM at any step (mirrors the official yfzhang114/MME-RealWorld rule-based protocol).", "note": "Official reporting uses per-task/l2-category dimension ratings (get_dimension_rating) over Perception/Reasoning; per-sample mean reproduces only overall accuracy. Five options with E typically 'The image does not feature the object.'; published choices carry '(A) ...' prefixes with a bare letter as answer." }, "prompt_template": "{{ question }} The choices are listed below:\n{% for c in choices %}{{ c }}\n{% endfor %}Select the best answer to the above multiple-choice question based on the image. Respond with only the letter (A, B, C, D, or E) of the correct option. The best answer is:", "mapping_from_source": { "media": { "from": "image", "type": "list", "min_items": 1, "max_items": 1 }, "id": { "from": "question_id" }, "question": { "from": "text" }, "answer": { "from": "answer", "optional": true }, "choices": { "from": "choices", "optional": true }, "extra": { "category": { "from": "category" }, "subtask": { "from": "subtask" }, "task": { "from": "task" } }, "source": { "format": "json", "url": { "test": "https://huggingface.co/datasets/yifanzhang114/MME-RealWorld-Lite" } } }, "prompt_template_source": { "origin": "official", "reference": "https://github.com/yfzhang114/MME-RealWorld README", "notes": "Tier 1: byte-for-byte from cited source" } } } }