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

@@ -9,7 +9,7 @@
       "modalities": [
         "single_image_start"
       ],
-      "task_type": "multiple_choice_vqa",
+      "task_type": "multiple_choice",
       "score_type": "llm_extract",
       "score_protocol": {
         "reference": "Published copy (HF datasets-server, mm-eval/Visual7W test, 69817 rows): choices A-D embedded in question text with trailer 'Answer with the letter.', gold = letter. Original Visual7W protocol (Zhu et al., CVPR 2016) is log-likelihood candidate ranking accuracy — not executable for generative/API models; classified per the generic VLMEvalKit MCQ letter-extraction protocol (vlmeval/dataset/utils/multiple_choice.py:359-406: rule prefetch + GPT fallback, exact letter compare).",
@@ -46,9 +46,9 @@
         }
       },
       "prompt_template_source": {
-        "origin": "official",
-        "reference": "https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/lmms_eval/tasks/visual7w/utils.py (visual7w_doc_to_text — bare question; 7W telling task)",
-        "notes": "Tier 4: lmms-eval Visual7W canonical evaluation prompt."
+        "origin": "fallback",
+        "reference": "T2",
+        "notes": "Tier 5: no higher tier publishes a prompt. T1: yukezhu/visual7w-toolkit is classification-era (evaluate.py consumes prediction JSON; no model-input prompt); T2: no source prompt column (choices are embedded in the question text of nimapourjafar/mm_visual7w, which itself triggers the T2 bare-question rule); T3: no VLMEvalKit visual7w dataset (tree + code search, 2026-07-07); T4: no lmms-eval tasks/visual7w in current main or v0.2.4-v0.3.1 tags."
       }
     }
   }