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

metadata.json

@@ -1,66 +1,69 @@
 {
- "name": "PuzzleVQA",
- "release_date": "2024-03-21",
- "subsets": {
-  "default": {
-   "language": [
-    "en"
-   ],
-   "modalities": [
-    "single_image_start"
-   ],
-   "task_type": "multiple_choice_qa",
-   "score_type": "rule",
-   "score_protocol": {
-    "reference": "official@github.com/declare-lab/LLM-PuzzleTest PuzzleVQA/main.py — prompter.get_answer parses option from raw output, ExactScorer returns 1.0 iff pred == answer; the only model call in the loop is a second-stage re-prompt of the EVALUATED model itself (inference-side self-extraction), no external judge/extractor LLM. Cross-check: VLMEvalKit@vlmeval/dataset/utils/puzzlevqa.py:85-152 PuzzleVQA_acc — pure regex extraction ('answer: X', boxed, answer-flag phrases), unparseable -> 'Z', per-category accuracy.",
-    "note": "Official prompting is two-stage (CoT, then the same model restates the answer); mm-eval prompt asks for the letter directly, so single-stage rule extraction is faithful at scoring time. VLMEvalKit aggregates per-puzzle-category accuracy (20 categories) alongside overall."
-   },
-   "prompt_template": "<image>{{ question }}\n{% for k, v in options.items() %}{{ k }}. {{ v }}\n{% endfor %}Answer with the option's letter from the given choices directly.",
-   "prompt_template_source": {
-    "origin": "official",
-    "reference": "https://huggingface.co/datasets/declare-lab/PuzzleVQA",
-    "notes": "Tier 1: PuzzleVQA is single-image MCQ with options list; answer is the option text. We map answer to the letter for letter-eval convention."
-   },
-   "mapping_from_source": {
-    "media": {
-     "from": "images",
-     "type": "list",
-     "min_items": 1,
-     "max_items": 1
-    },
-    "id": {
-     "from": "id"
-    },
-    "question": {
-     "from": "question"
-    },
-    "answer": {
-     "from": "answer",
-     "optional": true
-    },
-    "options": {
-     "from": "options",
-     "optional": true,
-     "note": "list source values are normalized to {A,B,...} dict"
-    },
-    "extra": {
-     "n_images": {
-      "from": "n_images"
-     },
-     "category": {
-      "from": "category"
-     },
-     "answer_text": {
-      "from": "answer_text"
-     }
-    },
-    "source": {
-     "format": "json",
-     "url": {
-      "train": "https://huggingface.co/datasets/declare-lab/PuzzleVQA"
-     }
+  "name": "PuzzleVQA",
+  "release_date": "2024-03-21",
+  "subsets": {
+    "default": {
+      "language": [
+        "en"
+      ],
+      "modalities": [
+        "single_image_start"
+      ],
+      "task_type": "multiple_choice_qa",
+      "score_pipeline": [
+        "exact-match",
+        "rule-match"
+      ],
+      "score_protocol": {
+        "reference": "official@github.com/declare-lab/LLM-PuzzleTest PuzzleVQA/main.py — prompter.get_answer parses option from raw output, ExactScorer returns 1.0 iff pred == answer; the only model call in the loop is a second-stage re-prompt of the EVALUATED model itself (inference-side self-extraction), no external judge/extractor LLM. Cross-check: VLMEvalKit@vlmeval/dataset/utils/puzzlevqa.py:85-152 PuzzleVQA_acc — pure regex extraction ('answer: X', boxed, answer-flag phrases), unparseable -> 'Z', per-category accuracy.",
+        "note": "Official prompting is two-stage (CoT, then the same model restates the answer); mm-eval prompt asks for the letter directly, so single-stage rule extraction is faithful at scoring time. VLMEvalKit aggregates per-puzzle-category accuracy (20 categories) alongside overall."
+      },
+      "prompt_template": "<image>{{ question }}\n{% for k, v in options.items() %}{{ k }}. {{ v }}\n{% endfor %}Answer with the option's letter from the given choices directly.",
+      "prompt_template_source": {
+        "origin": "official",
+        "reference": "https://huggingface.co/datasets/declare-lab/PuzzleVQA",
+        "notes": "Tier 1: PuzzleVQA is single-image MCQ with options list; answer is the option text. We map answer to the letter for letter-eval convention."
+      },
+      "mapping_from_source": {
+        "media": {
+          "from": "images",
+          "type": "list",
+          "min_items": 1,
+          "max_items": 1
+        },
+        "id": {
+          "from": "id"
+        },
+        "question": {
+          "from": "question"
+        },
+        "answer": {
+          "from": "answer",
+          "optional": true
+        },
+        "options": {
+          "from": "options",
+          "optional": true,
+          "note": "list source values are normalized to {A,B,...} dict"
+        },
+        "extra": {
+          "n_images": {
+            "from": "n_images"
+          },
+          "category": {
+            "from": "category"
+          },
+          "answer_text": {
+            "from": "answer_text"
+          }
+        },
+        "source": {
+          "format": "json",
+          "url": {
+            "train": "https://huggingface.co/datasets/declare-lab/PuzzleVQA"
+          }
+        }
+      }
     }
-   }
   }
- }
-}
+}