metadata: 2026-07-07 revised taxonomy (task_type format vocabulary multiple_choice/yes_no/short_answer/free_form; score_type incl. none; re-verified protocol evidence; repaired prompt provenance)

metadata.json

@@ -9,10 +9,10 @@
       "modalities": [
         "single_image_start"
       ],
-      "task_type": "vqa",
+      "task_type": "short_answer",
       "score_type": "vqa_accuracy",
       "score_protocol": {
-        "reference": "lmms-eval@lmms_eval/tasks/ok_vqa/utils.py:13-36 (ok_vqa_process_results) — EvalAIAnswerProcessor normalization then acc=min(1,#matches/3) averaged over 10 answers. rows/OK-VQA.json confirms published `answer` retains the full 10-item list (e.g. id 1000005).",
+        "reference": "lmms-eval@lmms_eval/tasks/ok_vqa/utils.py:13-36 (ok_vqa_process_results) verified: EvalAIAnswerProcessor normalization then acc=min(1,#matches/3) averaged over the 10-answer list. Published answer maps from source `answers` (10-item list) — confirmed retained in rows.",
         "note": "Fractional per-sample VQA accuracy averaged over held-out-one of 10 reference answers; headline is mean over samples."
       },
       "prompt_template": "<image>{{ question }}\nAnswer the question using a single word or phrase.",