metadata: add score_type/score_params/score_protocol (official grading protocol taxonomy); task_type is now format-only (judge_scored migrated to vqa + llm_judge). See dataset-upload skill references/metadata-json.md.

metadata.json

@@ -10,6 +10,11 @@
         "single_image_start"
       ],
       "task_type": "vqa",
+      "score_type": "llm_extract",
+      "score_protocol": {
+        "reference": "VLMEvalKit@vlmeval/dataset/dynamath.py:58-109 — rule parse of the answer first (JSON 'short answer' / direct parse_answer), auxiliary LLM extraction only on parse failure, then rule compare: float abs(diff)<=0.001, MC letter equality, text containment. Official DynaMath@evaluation/gpt4o/gpt_4o_json_eval.py:98-123 uses the same JSON 'short answer' parse + abs(diff)<=0.001 rule compare.",
+        "note": "Official grading is rule-based over a JSON-formatted answer forced by their inference prompt; the mm-eval prompt (lmms-eval style question+Options trailer) does not force JSON, so LLM extraction fallback (VLMEvalKit protocol) is required. Official float tolerance is ABSOLUTE 0.001, not relative. Official headline metrics are average-case AND worst-case accuracy over 10 generated variants per seed question; the mm-eval copy ships 501 rows (seed variant only), so worst-case accuracy is not reproducible."
+      },
       "prompt_template": "<image>{{ question }}{% if options %}\nOptions:\n{% for k, v in options.items() %}{{ k }}. {{ v }}{% if not loop.last %}\n{% endif %}{% endfor %}{% endif %}\n",
       "mapping_from_source": {
         "media": {