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 +10,19 @@
         "multi_image_start"
       ],
       "task_type": "vqa",
+      "score_type": "llm_judge",
+      "score_params": {
+        "rubric": "scale",
+        "scale": [
+          0,
+          10
+        ],
+        "official_judge_model": "gpt-4o"
+      },
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/live_bench/utils.py:49-131,160-186 (v1, covers the 2024-05 era) and utils_v2.py:49-121,159-186 — this is lmms-lab/LiveBench (live web-news screenshots), NOT Abacus/NYU text LiveBench nor OpenCompass LiveMMBench; GPT-4o rates every answer 0-10 against ground truth using a per-sample `criteria` field with the screenshot images attached; aggregate = mean(rating/10)*100 with subtask breakdown.",
+        "note": "mm-eval copy does not carry the per-sample `criteria` column that the official judge prompt interpolates (extra keeps only subtask/website/data_generator), and the official judge is multimodal (sees the screenshots) — a text-only judge without criteria deviates from official. 0-10 rating averaged then scaled x10 to 0-100."
+      },
       "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}",
       "prompt_template_source": {
         "origin": "official",
@@ -60,6 +73,19 @@
         "multi_image_start"
       ],
       "task_type": "vqa",
+      "score_type": "llm_judge",
+      "score_params": {
+        "rubric": "scale",
+        "scale": [
+          0,
+          10
+        ],
+        "official_judge_model": "gpt-4o"
+      },
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/live_bench/live_bench.yaml + utils_v2.py:49-121,159-186 — task live_bench_2406; GPT-4o rates every answer 0-10 against ground truth using per-sample `criteria` and the screenshot images; aggregate = mean(rating/10)*100.",
+        "note": "mm-eval copy omits the per-sample `criteria` column used by the official judge prompt; official judge is multimodal (sees the screenshots). Some rows are MCQ/fill-in-the-blank phrased inline but all are judged with the same 0-10 GPT rubric."
+      },
       "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}",
       "prompt_template_source": {
         "origin": "official",
@@ -110,6 +136,19 @@
         "multi_image_start"
       ],
       "task_type": "vqa",
+      "score_type": "llm_judge",
+      "score_params": {
+        "rubric": "scale",
+        "scale": [
+          0,
+          10
+        ],
+        "official_judge_model": "gpt-4o"
+      },
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/live_bench/live_bench.yaml + utils_v2.py:49-121,159-186 — task live_bench_2407; same GPT-4o 0-10 judge with per-sample criteria and images.",
+        "note": "mm-eval copy omits the per-sample `criteria` column used by the official judge prompt; official judge is multimodal (sees the screenshots)."
+      },
       "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}",
       "prompt_template_source": {
         "origin": "official",
@@ -160,6 +199,19 @@
         "multi_image_start"
       ],
       "task_type": "vqa",
+      "score_type": "llm_judge",
+      "score_params": {
+        "rubric": "scale",
+        "scale": [
+          0,
+          10
+        ],
+        "official_judge_model": "gpt-4o"
+      },
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/live_bench/utils_v2.py:49-121,159-186 — same GPT-4o 0-10 judge protocol; 2024-08 config is not listed in live_bench.yaml task group.",
+        "note": "Only 8 rows in this config (source upload appears truncated) — results statistically meaningless; mm-eval copy omits the per-sample `criteria` column; official judge is multimodal."
+      },
       "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}",
       "prompt_template_source": {
         "origin": "official",
@@ -210,6 +262,19 @@
         "multi_image_start"
       ],
       "task_type": "vqa",
+      "score_type": "llm_judge",
+      "score_params": {
+        "rubric": "scale",
+        "scale": [
+          0,
+          10
+        ],
+        "official_judge_model": "gpt-4o"
+      },
+      "score_protocol": {
+        "reference": "lmms-eval@lmms_eval/tasks/live_bench/live_bench.yaml + utils_v2.py:49-121,159-186 — task live_bench_2409; GPT-4o 0-10 judge with per-sample criteria and images; subtask taxonomy v2 (Concrete Recognition/Analytical Questions/Divergent Thinking/Real-world Assistance).",
+        "note": "mm-eval copy omits the per-sample `criteria` column used by the official judge prompt; official judge is multimodal (sees the screenshots)."
+      },
       "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}",
       "prompt_template_source": {
         "origin": "official",