133 lines
3.4 KiB
Markdown
133 lines
3.4 KiB
Markdown
# DocVQA
|
||
|
||
|
||
## 概述
|
||
|
||
DocVQA（Document Visual Question Answering，文档视觉问答）是一个用于评估 AI 系统根据文档图像（如扫描页面、表单、发票和报告）回答问题能力的基准测试。该任务不仅要求简单的文本提取，还需要理解复杂的文档布局、结构和视觉元素。
|
||
|
||
## 任务描述
|
||
|
||
- **任务类型**：文档视觉问答
|
||
- **输入**：文档图像 + 自然语言问题
|
||
- **输出**：从文档中提取的单个词或短语作为答案
|
||
- **领域**：文档理解、OCR、版面理解
|
||
|
||
## 主要特点
|
||
|
||
- 涵盖多种文档类型（表单、发票、信函、报告）
|
||
- 需要理解文档的版面与结构
|
||
- 同时考察文本提取能力和上下文推理能力
|
||
- 问题要求定位并解读特定信息
|
||
- 结合了 OCR 能力与视觉理解能力
|
||
|
||
## 评估说明
|
||
|
||
- 默认使用 **验证集**（validation split）进行评估
|
||
- 主要指标：**ANLS**（Average Normalized Levenshtein Similarity，平均归一化编辑距离相似度）
|
||
- 答案格式应为 `"ANSWER: [ANSWER]"`
|
||
- ANLS 指标可容忍轻微的 OCR 或拼写差异
|
||
- 每个问题可能接受多个有效答案
|
||
|
||
## 属性
|
||
|
||
| 属性 | 值 |
|
||
|----------|-------|
|
||
| **基准测试名称** | `docvqa` |
|
||
| **数据集 ID** | [lmms-lab/DocVQA](https://modelscope.cn/datasets/lmms-lab/DocVQA/summary) |
|
||
| **论文** | N/A |
|
||
| **标签** | `Knowledge`, `MultiModal`, `QA` |
|
||
| **指标** | `anls` |
|
||
| **默认示例数量** | 0-shot |
|
||
| **评估划分** | `validation` |
|
||
|
||
|
||
## 数据统计
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 总样本数 | 5,349 |
|
||
| 提示词长度（平均） | 254.82 字符 |
|
||
| 提示词长度（最小/最大） | 220 / 354 字符 |
|
||
|
||
**图像统计信息：**
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 总图像数 | 5,000 |
|
||
| 每样本图像数 | 最小: 1, 最大: 1, 平均: 1 |
|
||
| 分辨率范围 | 593x294 - 5367x7184 |
|
||
| 格式 | png |
|
||
|
||
|
||
## 样例示例
|
||
|
||
**子集**: `DocVQA`
|
||
|
||
```json
|
||
{
|
||
"input": [
|
||
{
|
||
"id": "002390bd",
|
||
"content": [
|
||
{
|
||
"text": "Answer the question according to the image using a single word or phrase.\nWhat is the ‘actual’ value per 1000, during the year 1975?\nThe last line of your response should be of the form \"ANSWER: [ANSWER]\" (without quotes) where [ANSWER] is the answer to the question."
|
||
},
|
||
{
|
||
"image": "[BASE64_IMAGE: png, ~1.2MB]"
|
||
}
|
||
]
|
||
}
|
||
],
|
||
"target": "[\"0.28\"]",
|
||
"id": 0,
|
||
"group_id": 0,
|
||
"metadata": {
|
||
"questionId": "49153",
|
||
"question_types": [
|
||
"figure/diagram"
|
||
],
|
||
"docId": 14465,
|
||
"ucsf_document_id": "pybv0228",
|
||
"ucsf_document_page_no": "81"
|
||
}
|
||
}
|
||
```
|
||
|
||
## 提示模板
|
||
|
||
**提示模板：**
|
||
```text
|
||
Answer the question according to the image using a single word or phrase.
|
||
{question}
|
||
The last line of your response should be of the form "ANSWER: [ANSWER]" (without quotes) where [ANSWER] is the answer to the question.
|
||
```
|
||
|
||
## 使用方法
|
||
|
||
### 使用 CLI
|
||
|
||
```bash
|
||
evalscope eval \
|
||
--model YOUR_MODEL \
|
||
--api-url OPENAI_API_COMPAT_URL \
|
||
--api-key EMPTY_TOKEN \
|
||
--datasets docvqa \
|
||
--limit 10 # 正式评估时请删除此行
|
||
```
|
||
|
||
### 使用 Python
|
||
|
||
```python
|
||
from evalscope import run_task
|
||
from evalscope.config import TaskConfig
|
||
|
||
task_cfg = TaskConfig(
|
||
model='YOUR_MODEL',
|
||
api_url='OPENAI_API_COMPAT_URL',
|
||
api_key='EMPTY_TOKEN',
|
||
datasets=['docvqa'],
|
||
limit=10, # 正式评估时请删除此行
|
||
)
|
||
|
||
run_task(task_cfg=task_cfg)
|
||
``` |