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_video_start"
       ],
       "task_type": "multiple_choice_vqa",
+      "score_type": "rule_llm_judge",
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/tempcompass/utils.py:115-172 (hand-crafted letter-match rules first; on failure ChatGPT judges Correct/Incorrect) + :480 (judge = gpt-3.5-turbo-1106) — mirrors official llyx97/TempCompass eval; vlmevalkit@vlmeval/dataset/tempcompass.py:63-260 (TempCompass_MCQ, DEFAULT_JUDGE chatgpt-1106) agrees.",
+        "note": "Official judge model gpt-3.5-turbo-1106. Official reports accuracy per temporal dimension (extra.dim) and overall average across the four task formats; per-sample scoring reproduces per-subset accuracy only."
+      },
       "prompt_template": "<video>{{ question }}\nPlease directly give the best option:",
       "mapping_from_source": {
         "media": {
@@ -62,6 +67,11 @@
         "single_video_start"
       ],
       "task_type": "vqa",
+      "score_type": "rule_llm_judge",
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/tempcompass/utils.py:174-225 (extract_pred yes/no rule first; ChatGPT gpt-3.5-turbo-1106 judge fallback) — mirrors official llyx97/TempCompass; vlmevalkit@vlmeval/dataset/tempcompass.py:458-630 (TempCompass_YorN) agrees.",
+        "note": "Official judge model gpt-3.5-turbo-1106; official aggregation is per-dim + cross-format average."
+      },
       "prompt_template": "<video>{{ question }}\nPlease answer yes or no:",
       "mapping_from_source": {
         "media": {
@@ -114,6 +124,11 @@
         "single_video_start"
       ],
       "task_type": "multiple_choice_vqa",
+      "score_type": "rule_llm_judge",
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/tempcompass/utils.py:227-277 (eval_rule deterministic caption-option matching first; ChatGPT gpt-3.5-turbo-1106 judge fallback) — mirrors official llyx97/TempCompass caption_matching eval.",
+        "note": "Options are full caption sentences embedded in the question text (no options dict); official judge model gpt-3.5-turbo-1106."
+      },
       "prompt_template": "<video>{{ question }}\nPlease directly give the best option:",
       "mapping_from_source": {
         "media": {
@@ -166,6 +181,15 @@
         "single_video_start"
       ],
       "task_type": "vqa",
+      "score_type": "llm_judge",
+      "score_params": {
+        "rubric": "binary",
+        "official_judge_model": "gpt-3.5-turbo-1106"
+      },
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/tempcompass/utils.py:280-345 + :409 (every generated caption is graded by ChatGPT: the judge reads the caption and answers the row's mc_question; rating = 1 iff the judge's choice equals mc_answer) — official llyx97/TempCompass caption-generation eval.",
+        "note": "Official protocol requires per-row mc_question and mc_answer; the mm-eval mapping keeps only answer/video_id/dim — as published, the official caption-matching judge prompt cannot be reconstructed. Not caption_metrics (no CIDEr/BLEU): grading is LLM-QA-based, binary per sample."
+      },
       "prompt_template": "<video>{{ question }}",
       "mapping_from_source": {
         "media": {