File size: 6,468 Bytes
455a754
eb26a37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d3c20d3
eb26a37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d3c20d3
455a754
eb26a37
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
{
  "name": "MathVerse",
  "release_date": "2024-03-22",
  "subsets": {
    "testmini": {
      "language": [
        "en"
      ],
      "modalities": [
        "single_image_start"
      ],
      "task_type": "multiple_choice_qa",
      "score_pipeline": [
        "exact-match",
        "rule-match",
        "llm-judge"
      ],
      "score_protocol": {
        "reference": "VLMEvalKit@vlmeval/dataset/utils/mathverse.py:123-167 — post_check_score exact-match prefetch, else GPT answer extraction (MathVerse_auxeval_extract) then GPT binary 0/1 correctness judgement (MathVerse_auxeval_score); same two-stage extract+score design as official ZrrSkywalker/MathVerse evaluation (extract_answer.py + score_answer.py). lmms-eval@lmms_eval/tasks/mathverse/mathverse_evals.py:92-110 judges every sample binary via LLM (quick_match optional exact match). [Re-verified in clone 2026-07-07: post_check_score at mathverse.py:126-133, auxeval_extract/score at 136-170; lmms score_answer at mathverse_evals.py:99-110.]",
        "note": "Official protocol judges with an LLM after LLM extraction (binary 0/1); VLMEvalKit pre-resolves exact string matches by rule, which this classification adopts. mm-eval publishes the answer-only protocol (query_wo prompts, 'please directly answer'), not the official CoT eval (query_cot + multi-step CoT judging) — scores are comparable to VLMEvalKit MathVerse_MINI, not to the paper's CoT-E numbers. The HF repo also carries a testmini_text_only config (788 rows, no image) that metadata.json does not describe as a subset."
      },
      "prompt_template": "<image>{{ question }}",
      "prompt_template_source": {
        "origin": "source_column",
        "reference": "https://huggingface.co/datasets/AI4Math/MathVerse (query_wo column = pre-rendered eval prompt)",
        "notes": "Tier 2: MathVerse provides query_wo column which is the pre-rendered single-answer evaluation prompt produced by the official eval pipeline; we use it verbatim and prepend <image>."
      },
      "mapping_from_source": {
        "media": {
          "from": "image",
          "type": "list",
          "min_items": 1,
          "max_items": 1
        },
        "id": {
          "from": "sample_index"
        },
        "question": {
          "from": "query_wo"
        },
        "answer": {
          "from": "answer",
          "optional": true
        },
        "extra": {
          "question_type": {
            "from": "question_type"
          },
          "problem_version": {
            "from": "problem_version"
          },
          "problem_index": {
            "from": "problem_index"
          },
          "source": {
            "from": "source",
            "note": "v2 message pass-through moved to the flat top level (format v3 re-emission)"
          },
          "subfield": {
            "from": "subfield",
            "note": "v2 message pass-through moved to the flat top level (format v3 re-emission)"
          },
          "subject": {
            "from": "subject",
            "note": "v2 message pass-through moved to the flat top level (format v3 re-emission)"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "testmini": "https://huggingface.co/datasets/AI4Math/MathVerse"
          }
        }
      }
    },
    "testmini_text_only": {
      "language": [
        "en"
      ],
      "modalities": [
        "text"
      ],
      "task_type": "multiple_choice_qa",
      "score_pipeline": [
        "exact-match",
        "rule-match",
        "llm-judge"
      ],
      "score_protocol": {
        "reference": "VLMEvalKit@vlmeval/dataset/utils/mathverse.py:123-167 — post_check_score exact-match prefetch, else GPT answer extraction (MathVerse_auxeval_extract) then GPT binary 0/1 correctness judgement (MathVerse_auxeval_score); same two-stage extract+score design as official ZrrSkywalker/MathVerse evaluation (extract_answer.py + score_answer.py). lmms-eval@lmms_eval/tasks/mathverse/mathverse_evals.py:92-110 judges every sample binary via LLM (quick_match optional exact match). [Re-verified in clone 2026-07-07: post_check_score at mathverse.py:126-133, auxeval_extract/score at 136-170; lmms score_answer at mathverse_evals.py:99-110.]",
        "note": "Same protocol as the testmini subset (answer-only query_wo prompts, two-stage LLM extract+score). This subset IS the repo's testmini_text_only config (788 rows, text-dominant problems without the diagram image), previously published as data but undescribed in metadata.json; declared as a subset at the format-v3 re-emission."
      },
      "prompt_template": "{{ question }}",
      "prompt_template_source": {
        "origin": "source_column",
        "reference": "https://huggingface.co/datasets/AI4Math/MathVerse (query_wo column = pre-rendered eval prompt)",
        "notes": "Tier 2: MathVerse provides query_wo column which is the pre-rendered single-answer evaluation prompt produced by the official eval pipeline; we use it verbatim. Text-only variant rows carry no media, so no <image> placeholder is prepended."
      },
      "mapping_from_source": {
        "media": {
          "from": "image",
          "type": "list",
          "min_items": 0,
          "max_items": 0
        },
        "id": {
          "from": "sample_index"
        },
        "question": {
          "from": "query_wo"
        },
        "answer": {
          "from": "answer",
          "optional": true
        },
        "extra": {
          "question_type": {
            "from": "question_type"
          },
          "problem_version": {
            "from": "problem_version"
          },
          "problem_index": {
            "from": "problem_index"
          },
          "source": {
            "from": "source",
            "note": "v2 message pass-through moved to the flat top level (format v3 re-emission)"
          },
          "subfield": {
            "from": "subfield",
            "note": "v2 message pass-through moved to the flat top level (format v3 re-emission)"
          },
          "subject": {
            "from": "subject",
            "note": "v2 message pass-through moved to the flat top level (format v3 re-emission)"
          }
        },
        "source": {
          "format": "huggingface",
          "url": {
            "testmini_text_only": "https://huggingface.co/datasets/AI4Math/MathVerse"
          }
        }
      }
    }
  }
}