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": "rule",
+      "score_protocol": {
+        "reference": "official mathllm/MATH-V@evaluation/evaluate.py — purely rule-based: boxed/'the answer is' extraction, find_math_answer, is_equal (latex2sympy symbolic compare), no LLM call; identical logic vendored at lmms-eval@lmms_eval/tasks/mathvision/utils.py:79-137 (mathvision_standard_eval) + eval_utils.py.",
+        "note": "Official is_equal does latex2sympy symbolic equivalence and also accepts the option TEXT for MCQ rows (gt letter or option value); the scorer's exact/template chain reproduces letter/number matches but not symbolic LaTeX equivalence. VLMEvalKit disagrees with the official protocol (utils/mathv.py:116-160: rule prefetch then GPT answer extraction = llm_extract); official rule protocol preferred. The published prompt is the lmms-eval boxed-CoT prompt, so boxed template extraction applies."
+      },
       "prompt_template": "Please first solve the problem step by step, then put your final answer or a single letter (if it is a multiple choice question) in one \"\\boxed{}\". \n<image>{{ question }}",
       "mapping_from_source": {
         "media": {