cistine commited on
Commit
f02f0be
·
verified ·
1 Parent(s): 640f87c

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +240 -48
README.md CHANGED
@@ -1,48 +1,240 @@
1
- ## ReasonBench:用于复杂图形推理的视觉语言模型评估基准
2
-
3
- ReasonBench 旨在评估视觉语言模型(VLMs)在复杂图形推理中的表现。数据集包含从真实智力测试中收集的 1,613 个问题,覆盖 11 个核心认知推理维度和 29 种任务类型,为评估 VLMs 的空间、关系和抽象推理能力提供综合框架。
4
-
5
- 数据集类型:视觉语言推理、图形推理、基准测试
6
-
7
-
8
- |11个核心认知维度 | 29种任务类型 |数量 |
9
- |--------------|----------------|-----|
10
- | 位置规律 | 平移 |94|
11
- | | 旋转 |56|
12
- | | 组合 |30|
13
- | 样式规律 | 穿越 |54|
14
- | | 加减法 |67|
15
- | | 黑白运算 |63|
16
- | 属性规律 | 对称 |109|
17
- | | 开闭状态 |19|
18
- | | 组合 |6|
19
- | 数量规律 | 线 |173|
20
- | | 面 |137|
21
- | | |66|
22
- | | 元素 |94|
23
- | | 组合 |50|
24
- | 空间规律 | 立方体 |109|
25
- | | 3d |46|
26
- | | 多面体 |17|
27
- | | 三视图 |40|
28
- | | 剖视图 |35|
29
- | | 空间数量变换 |10|
30
- | 特殊规律 | 2d 组合 |31|
31
- | | 图形关系 |40|
32
- | 字母数字 | 字母数字 |27|
33
- | 黑白块 | 黑白块 |32|
34
- | 其他规律 | 综合 |34|
35
- | 门萨 |任务1 |35|
36
- | |任务2 |39|
37
- | 瑞文 |任务1 |40|
38
- | |任务2 |60|
39
-
40
- 本基准考虑了输入图片的整体式和分隔式两种输入格式对模型的影响。
41
- |格式|介绍|
42
- |-------------|-------|
43
- |集成格式(Integrated format)|问题与选项呈现在单个图形中,便于模型整体处理|
44
- |分离格式(Separated format)|问题与选项拆分为多个图形,依次输入模型,测试分步推理能力|
45
-
46
- 为未来研究者方便复现以及后续研究,我们将所有格式的图片url全部公开,分别对应题目,4-8个选项,题目➕选项。
47
-
48
- 同时,我们公开了人类基准的准确率。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - zh
5
+ - en
6
+ tags:
7
+ - vlm
8
+ - benchmark
9
+ - graphic-reasoning
10
+ - intelligence-test
11
+ ---
12
+ # 🧠 ReasonBench: Benchmark for Complex Visual Reasoning
13
+
14
+ ## 🌐 Overview
15
+ **ReasonBench** is a comprehensive benchmark designed to evaluate Visual Language Models (VLMs) on complex graphical reasoning tasks. It contains **1,613 problems** collected from real-world intelligence tests, covering **11 core cognitive dimensions** and **29 task types**. This benchmark provides a robust framework for assessing VLMs' spatial, relational, and abstract reasoning capabilities.
16
+
17
+ **Dataset Type**: Visual Language Reasoning · Graphical Reasoning · Benchmark Evaluation
18
+
19
+ ## 📊 Dataset Structure
20
+ ### Core Cognitive Dimensions & Task Types
21
+ | Cognitive Dimension | Task Type | Count |
22
+ |--------------------------|-----------------------------|-------|
23
+ | **Positional Patterns** | Translation | 94 |
24
+ | | Rotation | 56 |
25
+ | | Combination | 30 |
26
+ | **Stylistic Patterns** | Crossing | 54 |
27
+ | | Addition/Subtraction | 67 |
28
+ | | Black/White Operation | 63 |
29
+ | **Attribute Patterns** | Symmetry | 109 |
30
+ | | Open/Close State | 19 |
31
+ | | Combination | 6 |
32
+ | **Quantitative Patterns**| Lines | 173 |
33
+ | | Faces | 137 |
34
+ | | Points | 66 |
35
+ | | Elements | 94 |
36
+ | | Combination | 50 |
37
+ | **Spatial Patterns** | Cubes | 109 |
38
+ | | 3D | 46 |
39
+ | | Polyhedrons | 17 |
40
+ | | Three Views | 40 |
41
+ | | Cross-Sections | 35 |
42
+ | | Spatial Quantitative Trans. | 10 |
43
+ | **Special Patterns** | 2D Combination | 31 |
44
+ | | Figure Relations | 40 |
45
+ | **Alphanumeric** | Alphanumeric | 27 |
46
+ | **B&W Blocks** | Black & White Blocks | 32 |
47
+ | **Other Patterns** | Comprehensive | 34 |
48
+ | **MENSA** | Task 1 | 35 |
49
+ | | Task 2 | 39 |
50
+ | **Raven** | Task 1 | 40 |
51
+ | | Task 2 | 60 |
52
+
53
+ ### 🖼️ Input Formats
54
+ | Format | Description |
55
+ |-----------------------|-------------|
56
+ | **Integrated Format** | Presents questions and options in a single image for holistic processing |
57
+ | **Separated Format** | Splits questions and options into multiple images for step-by-step reasoning |
58
+
59
+ ## 🔍 Key Features
60
+ - **Multi-format Evaluation**: Supports both integrated and separated input formats
61
+ - **Full Accessibility**: Provides public URLs for all images (questions, options, and combined sets)
62
+ - **Human Baseline**: Includes human performance metrics for comparison
63
+ - **Diverse Tasks**: Covers 29 distinct reasoning task types across 11 cognitive dimensions
64
+
65
+ ## 🚀 Usage(GPT-4o example)
66
+ ```python
67
+ import base64
68
+ import requests
69
+ import os
70
+ from openai import OpenAI # Requires openai>=1.0.0
71
+
72
+ # Configuration
73
+ api_key = os.getenv("OPENAI_API_KEY")
74
+ if not api_key:
75
+ raise ValueError("Missing OPENAI_API_KEY environment variable")
76
+
77
+ # Initialize client (official SDK approach)
78
+ client = OpenAI(api_key=api_key)
79
+
80
+ def process_image_question(image_path: str, question: str, max_tokens=300):
81
+ """Send image and question to GPT-4o API"""
82
+ # Encode image to base64
83
+ base64_image = base64.b64encode(open(image_path, "rb").read()).decode("utf-8")
84
+
85
+ # Construct messages payload
86
+ messages = [
87
+ {
88
+ "role": "user",
89
+ "content": [
90
+ {"type": "text", "text": question},
91
+ {
92
+ "type": "image_url",
93
+ "image_url": {
94
+ "url": f"data:image/jpeg;base64,{base64_image}",
95
+ "detail": "auto" # Options: low, high, auto
96
+ }
97
+ }
98
+ ]
99
+ }
100
+ ]
101
+
102
+ # Make API request
103
+ response = client.chat.completions.create(
104
+ model="gpt-4o",
105
+ messages=messages,
106
+ max_tokens=max_tokens
107
+ )
108
+
109
+ return response.choices[0].message.content
110
+
111
+ # Example usage
112
+ if __name__ == "__main__":
113
+ image_path = "path/to/your/image.jpg" # Update with actual path
114
+ user_question = "What's in this image?" # Customize your question
115
+
116
+ try:
117
+ answer = process_image_question(image_path, user_question)
118
+ print("AI Response:", answer)
119
+ except Exception as e:
120
+ print(f"Error: {str(e)}")
121
+ ---
122
+
123
+ # 🧠 ReasonBench:复杂图形推理的视觉语言模型评估基准
124
+
125
+ ## 🌐 概述
126
+ **ReasonBench** 是一个用于评估视觉语言模型(VLMs)在复杂图形推理任务表现的基准测试。数据集包含从真实智力测试中收集的 **1,613个问题**,覆盖**11个核心认知维度**和**29种任务类型**,为评估VLMs的空间、关系和抽象推理能力提供综合框架。
127
+
128
+ **数据集类型**:视觉语言推理 · 图形推理 · 基准评估
129
+
130
+ ## 📊 数据结构
131
+ ### 核心认知维度与任务类型
132
+ | 认知维度 | 任务类型 | 数量 |
133
+ |---------------------|------------------------|------|
134
+ | **位置规律** | 平移 | 94 |
135
+ | | 旋转 | 56 |
136
+ | | 组合 | 30 |
137
+ | **样式规律** | 穿越 | 54 |
138
+ | | 加减法 | 67 |
139
+ | | 黑白运算 | 63 |
140
+ | **属性规律** | 对称 | 109 |
141
+ | | 开闭状态 | 19 |
142
+ | | 组合 | 6 |
143
+ | **数量规律** | 线 | 173 |
144
+ | | 面 | 137 |
145
+ | | 点 | 66 |
146
+ | | 元素 | 94 |
147
+ | | 组合 | 50 |
148
+ | **空间规律** | 立方体 | 109 |
149
+ | | 3D | 46 |
150
+ | | 多面体 | 17 |
151
+ | | 三视图 | 40 |
152
+ | | 剖视图 | 35 |
153
+ | | 空间数量变换 | 10 |
154
+ | **特殊规律** | 2D组合 | 31 |
155
+ | | 图形关系 | 40 |
156
+ | **字母数字** | 字母数字 | 27 |
157
+ | **黑白块** | 黑白块 | 32 |
158
+ | **其他规律** | 综合 | 34 |
159
+ | **门萨** | 任务1 | 35 |
160
+ | | 任务2 | 39 |
161
+ | **瑞文** | 任务1 | 40 |
162
+ | | 任务2 | 60 |
163
+
164
+ ### 🖼️ 输入格式
165
+ | 格式 | 描述 |
166
+ |---------------------|------|
167
+ | **集成格式** | 问题与选项呈现在单个图形中,便于模型整体处理 |
168
+ | **分离格式** | 将问题与选项拆分为多个图形,测试分步推理能力 |
169
+
170
+ ## 🔍 核心特性
171
+ - **多格式评估**:支持整体式和分隔式两种输入格式
172
+ - **完全开放**:公开所有格式的图片URL(题目、选项、题目+选项)
173
+ - **人类基准**:提供人类准确率作为参考基准
174
+ - **多样化任务**:覆盖11个认知维度的29种推理任务
175
+
176
+ ## 🚀 使用示例(以openai GPT-4o为例)
177
+ ```python
178
+ import base64
179
+ import requests
180
+ import os
181
+
182
+ # 配置OpenAI API密钥
183
+ api_key = os.getenv("OPENAI_API_KEY") # 建议将密钥存储在环境变量中
184
+ if not api_key:
185
+ raise ValueError("请设置OPENAI_API_KEY环境变量")
186
+
187
+ # 图像处理函数
188
+ def encode_image(image_path):
189
+ """将本地图像编码为base64字符串"""
190
+ with open(image_path, "rb") as image_file:
191
+ return base64.b64encode(image_file.read()).decode('utf-8')
192
+
193
+ # 示例图像路径和问题
194
+ image_path = "path/to/your/image.jpg" # 替换为你的图像路径
195
+ question = "描述这张图片的内容" # 替换为你的问题
196
+
197
+ # 构建API请求
198
+ headers = {
199
+ "Content-Type": "application/json",
200
+ "Authorization": f"Bearer {api_key}"
201
+ }
202
+
203
+ payload = {
204
+ "model": "gpt-4o", # 使用支持图像的模型
205
+ "messages": [
206
+ {
207
+ "role": "user",
208
+ "content": [
209
+ {
210
+ "type": "text",
211
+ "text": question
212
+ },
213
+ {
214
+ "type": "image_url",
215
+ "image_url": {
216
+ "url": f"data:image/jpeg;base64,{encode_image(image_path)}"
217
+ }
218
+ }
219
+ ]
220
+ }
221
+ ],
222
+ "max_tokens": 300 # 控制响应长度
223
+ }
224
+
225
+ # 发送请求
226
+ response = requests.post(
227
+ "https://api.openai.com/v1/chat/completions",
228
+ headers=headers,
229
+ json=payload
230
+ )
231
+
232
+ # 处理响应
233
+ if response.status_code == 200:
234
+ result = response.json()
235
+ answer = result['choices'][0]['message']['content']
236
+ print("AI回答:", answer)
237
+ else:
238
+ print("请求失败,状态码:", response.status_code)
239
+ print("错误信息:", response.text)
240
+ ```