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": "multiple_choice_vqa",
+      "score_type": "rule",
+      "score_protocol": {
+        "reference": "official We-Math/We-Math@README + evaluation/four_dimensional_metrics.py & evaluation/accuracy.py — 'we use string matching to directly extract answers, which eliminates the high cost of using additional models for further answer extraction' (rule-based option-letter extraction, exact compare).",
+        "note": "Official headline is the four-dimensional metric set (IK/IG/CM/RM) plus step-wise accuracy, computed by grouping one-step sub-problems with their multi-step composite problem (problem_id/step_key, vendored in VLMEvalKit vlmeval/dataset/utils/wemath.py:16-238); a per-sample accuracy scorer will not reproduce those group scores (grouping keys ARE shipped in extra). VLMEvalKit grades WeMath via mcq_vanilla_eval (utils/multiple_choice.py:475, rule can_infer + GPT extraction fallback), i.e. llm_extract — disagrees with the official rule-only protocol; official preferred."
+      },
       "prompt_template": "<image>{{ question }}\nOptions:\n{% for k, v in options.items() %}{{ k }}. {{ v }}{% if not loop.last %}\n{% endif %}{% endfor %}\nPlease select the correct answer from the options above. \n",
       "mapping_from_source": {
         "media": {