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": "lmms-eval@lmms_eval/tasks/geometry3k/geometry3k.yaml:23-28 (metric exact_match) + utils.py:62-125 (rule letter extraction, no LLM) — nearest framework protocol; original Inter-GPS (ACL 2021) evaluates by selecting among the 4 choices with rule matching. The published copy is the hiyouga/geometry3k RL variant whose ecosystem grading (EasyR1/verl mathruler) is rule-based boxed extraction + math equality.",
+        "note": "The mm-eval copy is NOT the official Inter-GPS MCQ release and NOT the lmms-eval 4-choices variant: answers are free-form numerics/LaTeX ('3', '120', '2 \\sqrt { 221 }') with no choices in the question, and the metadata prompt_template_source note ('choices and trailer baked into source question column') is inaccurate for this copy. Rule compare needs math-expression equivalence for LaTeX answers; exact/template matching under-scores them. Also ships train(2101)/validation/test splits — an RL training set, not an eval-only benchmark."
+      },
       "prompt_template": "{{ question }}",
       "mapping_from_source": {
         "media": {