Anonymous723 commited on
Commit
f322e2b
·
verified ·
1 Parent(s): bc2fc73

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +150 -146
README.md CHANGED
@@ -1,146 +1,150 @@
1
- ---
2
- license: cc-by-4.0
3
- language:
4
- - zh
5
- - en
6
- tags:
7
- - multimodal-mathematical-reasoning
8
- - geometry
9
- - tikz
10
- - latex
11
- - image-to-tikz
12
- - benchmark
13
- size_categories:
14
- - 10K<n<100K
15
- task_categories:
16
- - image-to-text
17
- - visual-question-answering
18
- - text-generation
19
- pretty_name: TriGeoBench
20
- ---
21
-
22
- # TriGeoBench
23
-
24
- TriGeoBench is a geometry-centric multimodal mathematics benchmark designed for evaluating mathematical reasoning with visual diagrams and image-to-TikZ generation. The dataset contains mathematical problems, solutions, diagram images, and corresponding TikZ annotations.
25
-
26
- This repository is anonymized for peer review. Author and institution information will be added upon acceptance.
27
-
28
- ## Dataset Files
29
-
30
- The dataset contains four Parquet files:
31
-
32
- ```text
33
- image2tikz_train.parquet
34
- image2tikz_test.parquet
35
- question_train.parquet
36
- question_test.parquet
37
- ````
38
-
39
- The dataset supports two tasks:
40
-
41
- 1. **Image-to-TikZ generation**: generating TikZ code from a geometric diagram image.
42
- 2. **Multimodal mathematical reasoning**: solving math problems with textual questions, solutions, and associated figures.
43
-
44
- ## Image-to-TikZ Data
45
-
46
- Files:
47
-
48
- ```text
49
- image2tikz_train.parquet
50
- image2tikz_test.parquet
51
- ```
52
-
53
- Each row corresponds to one diagram image and its ground-truth TikZ code.
54
-
55
- ### Fields
56
-
57
- | Field | Description |
58
- | ------------ | ------------------------------------------------------------ |
59
- | `key` | Unique figure identifier. It is composed of `<problem_id>_<position>_<figure_index>`, where `position` indicates whether the figure appears in the question or the solution. This key can be linked to the corresponding problem in the question-level data. |
60
- | `image` | Base64-encoded image. |
61
- | `latex_gt` | Ground-truth TikZ code corresponding to the image. |
62
- | `difficulty` | Figure complexity level. Possible values are `容易`, `中等`, and `困难`. |
63
-
64
- ## Question-Level Data
65
-
66
- Files:
67
-
68
- ```text
69
- question_train.parquet
70
- question_test.parquet
71
- ```
72
-
73
- Each row corresponds to one mathematical problem, including the problem text, solution, metadata, and associated figures.
74
-
75
- ### Fields
76
-
77
- | Field | Description |
78
- | ----------------- | ------------------------------------------------------------ |
79
- | `sample_id` | Unique problem identifier. It can be linked to the `key` field in the image-to-TikZ data. |
80
- | `difficulty` | Problem difficulty level. Possible values are `容易`, `中等`, and `困难`. |
81
- | `question_type` | Problem type. Possible values include `选择题`, `填空题`, `解答题`, and `证明题`. |
82
- | `knowledge_point` | Main mathematical knowledge area. Possible values include `向量`, `函数`, `平面几何`, `立体几何`, and `解析几何`. |
83
- | `question` | Problem statement in LaTeX format. |
84
- | `solution` | Solution or answer in LaTeX format. |
85
- | `q_figX` | Base64-encoded image of the X-th figure appearing in the question. |
86
- | `q_figX_latex_gt` | Ground-truth TikZ code of the X-th question figure. |
87
- | `s_figY` | Base64-encoded image of the Y-th figure appearing in the solution. |
88
- | `s_figY_latex_gt` | Ground-truth TikZ code of the Y-th solution figure. |
89
-
90
- Here, `X` and `Y` denote figure indices. A problem may contain different numbers of question-side and solution-side figures.
91
-
92
- ## Data Splits
93
-
94
- The dataset is split into training and test sets for both tasks:
95
-
96
- | Task | Train File | Test File |
97
- | ---------------------- | -------------------------- | ------------------------- |
98
- | Image-to-TikZ | `image2tikz_train.parquet` | `image2tikz_test.parquet` |
99
- | Mathematical Reasoning | `question_train.parquet` | `question_test.parquet` |
100
-
101
- ## Loading the Dataset
102
-
103
- The Parquet files can be loaded with `pandas`:
104
-
105
- ```python
106
- import pandas as pd
107
-
108
- image2tikz_train = pd.read_parquet("image2tikz_train.parquet")
109
- image2tikz_test = pd.read_parquet("image2tikz_test.parquet")
110
-
111
- question_train = pd.read_parquet("question_train.parquet")
112
- question_test = pd.read_parquet("question_test.parquet")
113
- ```
114
-
115
- Base64-encoded images can be decoded as follows:
116
-
117
- ```python
118
- import base64
119
- from io import BytesIO
120
- from PIL import Image
121
-
122
- def decode_base64_image(image_base64):
123
- image_bytes = base64.b64decode(image_base64)
124
- return Image.open(BytesIO(image_bytes)).convert("RGB")
125
-
126
- img = decode_base64_image(image2tikz_train.iloc[0]["image"])
127
- img.show()
128
- ```
129
-
130
- ## Intended Use
131
-
132
- TriGeoBench is intended for research on:
133
-
134
- * multimodal mathematical reasoning;
135
- * geometry-centric visual question answering;
136
- * image-to-TikZ generation;
137
- * evaluating whether models can reason over precise geometric structures;
138
- * studying the interaction between textual math problems, visual diagrams, and symbolic diagram representations.
139
-
140
- ## Limitations
141
-
142
- The dataset focuses on geometry-centric middle- and high-school mathematics problems. The annotations include LaTeX-formatted problem texts and TikZ code for figures. Although the dataset has been processed and checked, residual annotation errors may remain.
143
-
144
- ## Anonymous Review Notice
145
-
146
- This repository is anonymized for peer review. Please do not attempt to identify the authors during the review process.
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ language:
4
+ - zh
5
+ - en
6
+ tags:
7
+ - multimodal-mathematical-reasoning
8
+ - geometry
9
+ - tikz
10
+ - latex
11
+ - image-to-tikz
12
+ - benchmark
13
+ size_categories:
14
+ - 10K<n<100K
15
+ task_categories:
16
+ - image-to-text
17
+ - visual-question-answering
18
+ - text-generation
19
+ pretty_name: TriGeoBench
20
+ ---
21
+
22
+ # TriGeoBench
23
+
24
+ TriGeoBench is a geometry-centric multimodal mathematics benchmark designed for evaluating mathematical reasoning with visual diagrams and image-to-TikZ generation. The dataset contains mathematical problems, solutions, diagram images, and corresponding TikZ annotations.
25
+
26
+ This repository is anonymized for peer review. Author and institution information will be added upon acceptance.
27
+
28
+ ## Dataset Files
29
+
30
+ The dataset contains four Parquet files:
31
+
32
+ ```text
33
+ TriGeoBench
34
+ ├── README.md
35
+ ├── image2tikz/
36
+ │ ├── train.parquet
37
+ │ └── test.parquet
38
+ └── question/
39
+ ├── train.parquet
40
+ └── test.parquet
41
+ ````
42
+
43
+ The dataset supports two tasks:
44
+
45
+ 1. **Image-to-TikZ generation**: generating TikZ code from a geometric diagram image.
46
+ 2. **Multimodal mathematical reasoning**: solving math problems with textual questions, solutions, and associated figures.
47
+
48
+ ## Image-to-TikZ Data
49
+
50
+ Files:
51
+
52
+ ```text
53
+ image2tikz/train.parquet
54
+ image2tikz/test.parquet
55
+ ```
56
+
57
+ Each row corresponds to one diagram image and its ground-truth TikZ code.
58
+
59
+ ### Fields
60
+
61
+ | Field | Description |
62
+ | ------------ | ------------------------------------------------------------ |
63
+ | `key` | Unique figure identifier. It is composed of `<problem_id>_<position>_<figure_index>`, where `position` indicates whether the figure appears in the question or the solution. This key can be linked to the corresponding problem in the question-level data. |
64
+ | `image` | Base64-encoded image. |
65
+ | `latex_gt` | Ground-truth TikZ code corresponding to the image. |
66
+ | `difficulty` | Figure complexity level. Possible values are `容易`, `中等`, and `困难`. |
67
+
68
+ ## Question-Level Data
69
+
70
+ Files:
71
+
72
+ ```text
73
+ question/train.parquet
74
+ question/test.parquet
75
+ ```
76
+
77
+ Each row corresponds to one mathematical problem, including the problem text, solution, metadata, and associated figures.
78
+
79
+ ### Fields
80
+
81
+ | Field | Description |
82
+ | ----------------- | ------------------------------------------------------------ |
83
+ | `sample_id` | Unique problem identifier. It can be linked to the `key` field in the image-to-TikZ data. |
84
+ | `difficulty` | Problem difficulty level. Possible values are `容易`, `中等`, and `困难`. |
85
+ | `question_type` | Problem type. Possible values include `选择题`, `填空题`, `解答题`, and `证明题`. |
86
+ | `knowledge_point` | Main mathematical knowledge area. Possible values include `向量`, `函数`, `平面几何`, `立体几何`, and `解析几何`. |
87
+ | `question` | Problem statement in LaTeX format. |
88
+ | `solution` | Solution or answer in LaTeX format. |
89
+ | `q_figX` | Base64-encoded image of the X-th figure appearing in the question. |
90
+ | `q_figX_latex_gt` | Ground-truth TikZ code of the X-th question figure. |
91
+ | `s_figY` | Base64-encoded image of the Y-th figure appearing in the solution. |
92
+ | `s_figY_latex_gt` | Ground-truth TikZ code of the Y-th solution figure. |
93
+
94
+ Here, `X` and `Y` denote figure indices. A problem may contain different numbers of question-side and solution-side figures.
95
+
96
+ ## Data Splits
97
+
98
+ The dataset is split into training and test sets for both tasks:
99
+
100
+ | Task | Train File | Test File |
101
+ | ---------------------- | -------------------------- | ------------------------- |
102
+ | Image-to-TikZ | `image2tikz_train.parquet` | `image2tikz_test.parquet` |
103
+ | Mathematical Reasoning | `question_train.parquet` | `question_test.parquet` |
104
+
105
+ ## Loading the Dataset
106
+
107
+ The Parquet files can be loaded with `pandas`:
108
+
109
+ ```python
110
+ import pandas as pd
111
+
112
+ image2tikz_train = pd.read_parquet("image2tikz_train.parquet")
113
+ image2tikz_test = pd.read_parquet("image2tikz_test.parquet")
114
+
115
+ question_train = pd.read_parquet("question_train.parquet")
116
+ question_test = pd.read_parquet("question_test.parquet")
117
+ ```
118
+
119
+ Base64-encoded images can be decoded as follows:
120
+
121
+ ```python
122
+ import base64
123
+ from io import BytesIO
124
+ from PIL import Image
125
+
126
+ def decode_base64_image(image_base64):
127
+ image_bytes = base64.b64decode(image_base64)
128
+ return Image.open(BytesIO(image_bytes)).convert("RGB")
129
+
130
+ img = decode_base64_image(image2tikz_train.iloc[0]["image"])
131
+ img.show()
132
+ ```
133
+
134
+ ## Intended Use
135
+
136
+ TriGeoBench is intended for research on:
137
+
138
+ * multimodal mathematical reasoning;
139
+ * geometry-centric visual question answering;
140
+ * image-to-TikZ generation;
141
+ * evaluating whether models can reason over precise geometric structures;
142
+ * studying the interaction between textual math problems, visual diagrams, and symbolic diagram representations.
143
+
144
+ ## Limitations
145
+
146
+ The dataset focuses on geometry-centric middle- and high-school mathematics problems. The annotations include LaTeX-formatted problem texts and TikZ code for figures. Although the dataset has been processed and checked, residual annotation errors may remain.
147
+
148
+ ## Anonymous Review Notice
149
+
150
+ This repository is anonymized for peer review. Please do not attempt to identify the authors during the review process.