Hugging Face's logo

Datasets:
mm-eval
/
SpatialEval

Modalities:
Image
Text
Formats:
parquet
Size:
1K - 10K
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
1
ZTWHHH commited on
Commit
083e849
·
verified ·
1 Parent(s): 1311916

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

Browse files
Files changed (1) hide show
  1. metadata.json +112 -106
metadata.json CHANGED
@@ -1,110 +1,116 @@
1
  {
2
- "name": "SpatialEval",
3
- "release_date": "2024-09-09",
4
- "subsets": {
5
- "vqa": {
6
- "language": [
7
- "en"
8
- ],
9
- "modalities": [
10
- "single_image_start"
11
- ],
12
- "task_type": "multiple_choice_qa",
13
- "score_type": "rule",
14
- "score_protocol": {
15
- "reference": "official@github.com/jiayuww/SpatialEval evals/evaluation.py — task-specific regex extraction (extract_answer_from_text_spatialmap L26 /_mazenav L140 /_spatialgrid L205) then lowercase substring match of the oracle answer ('ref_ans.lower() in model_answer.lower()', L286-288); no LLM anywhere.",
16
- "note": "Official grading matches the oracle ANSWER CONTENT by substring after regex extraction; the mm-eval copy stores the option letter (oracle_option, e.g. 'C') as ground truth with options embedded in the question text ('Available options:\\nA. 4...'), keeping oracle_answer/oracle_full_answer in extra. Scoring the published copy is therefore letter-selection matching, not content-substring matching. Official also defines a text-only tqa mode not shipped by the org (vqa/vtqa only)."
17
- },
18
- "prompt_template": "<image>{{ question }}",
19
- "prompt_template_source": {
20
- "origin": "source_column",
21
- "reference": "https://huggingface.co/datasets/MilaWang/SpatialEval",
22
- "notes": "Tier 2: SpatialEval text column pre-renders the question with embedded option list; oracle_option holds the gold letter. Template prepends <image> and uses text verbatim."
23
- },
24
- "mapping_from_source": {
25
- "media": {
26
- "from": "image",
27
- "type": "list",
28
- "min_items": 1,
29
- "max_items": 1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
30
  },
31
- "id": {
32
- "from": "id"
33
- },
34
- "question": {
35
- "from": "text"
36
- },
37
- "answer": {
38
- "from": "oracle_option",
39
- "optional": true
40
- },
41
- "extra": {
42
- "oracle_full_answer": {
43
- "from": "oracle_full_answer"
44
- },
45
- "oracle_answer": {
46
- "from": "oracle_answer"
47
- }
48
- },
49
- "source": {
50
- "format": "huggingface",
51
- "url": {
52
- "test": "https://huggingface.co/datasets/MilaWang/SpatialEval"
53
- }
54
- }
55
- }
56
- },
57
- "vtqa": {
58
- "language": [
59
- "en"
60
- ],
61
- "modalities": [
62
- "single_image_start"
63
- ],
64
- "task_type": "multiple_choice_qa",
65
- "score_type": "rule",
66
- "score_protocol": {
67
- "reference": "official@github.com/jiayuww/SpatialEval evals/evaluation.py — same rule-based regex extraction + lowercase substring match of the oracle answer (run with --mode 'vtqa'); no LLM.",
68
- "note": "Same as SpatialEval/vqa: published GT is the option letter while the official script substring-matches oracle answer content; tqa mode not shipped."
69
- },
70
- "prompt_template": "<image>{{ question }}",
71
- "prompt_template_source": {
72
- "origin": "source_column",
73
- "reference": "https://huggingface.co/datasets/MilaWang/SpatialEval",
74
- "notes": "Tier 2: SpatialEval text column pre-renders the question with embedded option list; oracle_option holds the gold letter. Template prepends <image> and uses text verbatim."
75
- },
76
- "mapping_from_source": {
77
- "media": {
78
- "from": "image",
79
- "type": "list",
80
- "min_items": 1,
81
- "max_items": 1
82
- },
83
- "id": {
84
- "from": "id"
85
- },
86
- "question": {
87
- "from": "text"
88
- },
89
- "answer": {
90
- "from": "oracle_option",
91
- "optional": true
92
- },
93
- "extra": {
94
- "oracle_full_answer": {
95
- "from": "oracle_full_answer"
96
- },
97
- "oracle_answer": {
98
- "from": "oracle_answer"
99
- }
100
- },
101
- "source": {
102
- "format": "huggingface",
103
- "url": {
104
- "test": "https://huggingface.co/datasets/MilaWang/SpatialEval"
105
- }
106
  }
107
- }
108
  }
109
- }
110
- }
 
1
  {
2
+ "name": "SpatialEval",
3
+ "release_date": "2024-09-09",
4
+ "subsets": {
5
+ "vqa": {
6
+ "language": [
7
+ "en"
8
+ ],
9
+ "modalities": [
10
+ "single_image_start"
11
+ ],
12
+ "task_type": "multiple_choice_qa",
13
+ "score_pipeline": [
14
+ "exact-match",
15
+ "rule-match"
16
+ ],
17
+ "score_protocol": {
18
+ "reference": "official@github.com/jiayuww/SpatialEval evals/evaluation.py — task-specific regex extraction (extract_answer_from_text_spatialmap L26 /_mazenav L140 /_spatialgrid L205) then lowercase substring match of the oracle answer ('ref_ans.lower() in model_answer.lower()', L286-288); no LLM anywhere.",
19
+ "note": "Official grading matches the oracle ANSWER CONTENT by substring after regex extraction; the mm-eval copy stores the option letter (oracle_option, e.g. 'C') as ground truth with options embedded in the question text ('Available options:\\nA. 4...'), keeping oracle_answer/oracle_full_answer in extra. Scoring the published copy is therefore letter-selection matching, not content-substring matching. Official also defines a text-only tqa mode not shipped by the org (vqa/vtqa only)."
20
+ },
21
+ "prompt_template": "<image>{{ question }}",
22
+ "prompt_template_source": {
23
+ "origin": "source_column",
24
+ "reference": "https://huggingface.co/datasets/MilaWang/SpatialEval",
25
+ "notes": "Tier 2: SpatialEval text column pre-renders the question with embedded option list; oracle_option holds the gold letter. Template prepends <image> and uses text verbatim."
26
+ },
27
+ "mapping_from_source": {
28
+ "media": {
29
+ "from": "image",
30
+ "type": "list",
31
+ "min_items": 1,
32
+ "max_items": 1
33
+ },
34
+ "id": {
35
+ "from": "id"
36
+ },
37
+ "question": {
38
+ "from": "text"
39
+ },
40
+ "answer": {
41
+ "from": "oracle_option",
42
+ "optional": true
43
+ },
44
+ "extra": {
45
+ "oracle_full_answer": {
46
+ "from": "oracle_full_answer"
47
+ },
48
+ "oracle_answer": {
49
+ "from": "oracle_answer"
50
+ }
51
+ },
52
+ "source": {
53
+ "format": "huggingface",
54
+ "url": {
55
+ "test": "https://huggingface.co/datasets/MilaWang/SpatialEval"
56
+ }
57
+ }
58
+ }
59
  },
60
+ "vtqa": {
61
+ "language": [
62
+ "en"
63
+ ],
64
+ "modalities": [
65
+ "single_image_start"
66
+ ],
67
+ "task_type": "multiple_choice_qa",
68
+ "score_pipeline": [
69
+ "exact-match",
70
+ "rule-match"
71
+ ],
72
+ "score_protocol": {
73
+ "reference": "official@github.com/jiayuww/SpatialEval evals/evaluation.py — same rule-based regex extraction + lowercase substring match of the oracle answer (run with --mode 'vtqa'); no LLM.",
74
+ "note": "Same as SpatialEval/vqa: published GT is the option letter while the official script substring-matches oracle answer content; tqa mode not shipped."
75
+ },
76
+ "prompt_template": "<image>{{ question }}",
77
+ "prompt_template_source": {
78
+ "origin": "source_column",
79
+ "reference": "https://huggingface.co/datasets/MilaWang/SpatialEval",
80
+ "notes": "Tier 2: SpatialEval text column pre-renders the question with embedded option list; oracle_option holds the gold letter. Template prepends <image> and uses text verbatim."
81
+ },
82
+ "mapping_from_source": {
83
+ "media": {
84
+ "from": "image",
85
+ "type": "list",
86
+ "min_items": 1,
87
+ "max_items": 1
88
+ },
89
+ "id": {
90
+ "from": "id"
91
+ },
92
+ "question": {
93
+ "from": "text"
94
+ },
95
+ "answer": {
96
+ "from": "oracle_option",
97
+ "optional": true
98
+ },
99
+ "extra": {
100
+ "oracle_full_answer": {
101
+ "from": "oracle_full_answer"
102
+ },
103
+ "oracle_answer": {
104
+ "from": "oracle_answer"
105
+ }
106
+ },
107
+ "source": {
108
+ "format": "huggingface",
109
+ "url": {
110
+ "test": "https://huggingface.co/datasets/MilaWang/SpatialEval"
111
+ }
112
+ }
113
+ }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
114
  }
 
115
  }
116
+ }