metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)

metadata.json

@@ -1,73 +1,76 @@
 {
- "name": "Q-Bench-Plus",
- "release_date": "2024-04-15",
- "subsets": {
-  "default": {
-   "language": [
-    "en"
-   ],
-   "modalities": [
-    "multi_image_start"
-   ],
-   "task_type": "multiple_choice_qa",
-   "score_type": "llm_extract",
-   "score_protocol": {
-    "reference": "official paper@arXiv:2402.07116 (Q-Bench+) — GPT-assisted evaluation matches the model's MCQ answer to the choices, with 5-round voting (raises match accuracy 93.2%->98.4%). Cross-check: lmms-eval@lmms_eval/tasks/qbench/qbench2_dev.yaml + utils.py:42-100,103-119 — rule-only MMMU-style parse then exact compare; VLMEvalKit hosts the sibling Q-Bench1/A-Bench as ImageMCQDataset (image_mcq.py:89-94) = rule prefetch + GPT extraction.",
-    "note": "Official uses 5-round GPT voting for answer matching; our llm_extract chain does rule prefetch + single-pass LLM extraction. lmms-eval variant is rule-only and falls back to a RANDOM choice when unparseable (not reproduced). Published copy = dev split (public-labeled), options stored as extra fields option0-option3 with 'N/A' padding possible."
-   },
-   "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}\nA. {{ options.option0 }}\nB. {{ options.option1 }}\nC. {{ options.option2 }}\nD. {{ options.option3 }}\nAnswer with the option's letter from the given choices directly.",
-   "prompt_template_source": {
-    "origin": "official",
-    "reference": "https://huggingface.co/datasets/q-future/Q-Bench2-HF",
-    "notes": "Tier 1: Q-Bench2 (Q-Bench+) paired-image quality MCQ; dev split is the public-labeled split."
-   },
-   "mapping_from_source": {
-    "media": {
-     "from": "images",
-     "type": "list",
-     "min_items": 2,
-     "max_items": 2
-    },
-    "id": {
-     "from": "id"
-    },
-    "question": {
-     "from": "question"
-    },
-    "answer": {
-     "from": "answer",
-     "optional": true
-    },
-    "extra": {
-     "n_images": {
-      "from": "n_images"
-     },
-     "option0": {
-      "from": "option0"
-     },
-     "option1": {
-      "from": "option1"
-     },
-     "option2": {
-      "from": "option2"
-     },
-     "option3": {
-      "from": "option3"
-     },
-     "question_type": {
-      "from": "question_type"
-     },
-     "question_concern": {
-      "from": "question_concern"
-     }
-    },
-    "source": {
-     "format": "json",
-     "url": {
-      "dev": "https://huggingface.co/datasets/q-future/Q-Bench2-HF"
-     }
+  "name": "Q-Bench-Plus",
+  "release_date": "2024-04-15",
+  "subsets": {
+    "default": {
+      "language": [
+        "en"
+      ],
+      "modalities": [
+        "multi_image_start"
+      ],
+      "task_type": "multiple_choice_qa",
+      "score_pipeline": [
+        "rule-match",
+        "llm-match"
+      ],
+      "score_protocol": {
+        "reference": "official paper@arXiv:2402.07116 (Q-Bench+) — GPT-assisted evaluation matches the model's MCQ answer to the choices, with 5-round voting (raises match accuracy 93.2%->98.4%). Cross-check: lmms-eval@lmms_eval/tasks/qbench/qbench2_dev.yaml + utils.py:42-100,103-119 — rule-only MMMU-style parse then exact compare; VLMEvalKit hosts the sibling Q-Bench1/A-Bench as ImageMCQDataset (image_mcq.py:89-94) = rule prefetch + GPT extraction.",
+        "note": "Official uses 5-round GPT voting for answer matching; our llm_extract chain does rule prefetch + single-pass LLM extraction. lmms-eval variant is rule-only and falls back to a RANDOM choice when unparseable (not reproduced). Published copy = dev split (public-labeled), options stored as extra fields option0-option3 with 'N/A' padding possible."
+      },
+      "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}\nA. {{ options.option0 }}\nB. {{ options.option1 }}\nC. {{ options.option2 }}\nD. {{ options.option3 }}\nAnswer with the option's letter from the given choices directly.",
+      "prompt_template_source": {
+        "origin": "official",
+        "reference": "https://huggingface.co/datasets/q-future/Q-Bench2-HF",
+        "notes": "Tier 1: Q-Bench2 (Q-Bench+) paired-image quality MCQ; dev split is the public-labeled split."
+      },
+      "mapping_from_source": {
+        "media": {
+          "from": "images",
+          "type": "list",
+          "min_items": 2,
+          "max_items": 2
+        },
+        "id": {
+          "from": "id"
+        },
+        "question": {
+          "from": "question"
+        },
+        "answer": {
+          "from": "answer",
+          "optional": true
+        },
+        "extra": {
+          "n_images": {
+            "from": "n_images"
+          },
+          "option0": {
+            "from": "option0"
+          },
+          "option1": {
+            "from": "option1"
+          },
+          "option2": {
+            "from": "option2"
+          },
+          "option3": {
+            "from": "option3"
+          },
+          "question_type": {
+            "from": "question_type"
+          },
+          "question_concern": {
+            "from": "question_concern"
+          }
+        },
+        "source": {
+          "format": "json",
+          "url": {
+            "dev": "https://huggingface.co/datasets/q-future/Q-Bench2-HF"
+          }
+        }
+      }
     }
-   }
   }
- }
-}
+}