13274243a0
Keep K3 suite selection and report-schema scoring in bash, merge K3/vision dataset_args into dpv4 yamls, and pin EvalScope at 735d920ee911 with local patches. Co-authored-by: Cursor <cursoragent@cursor.com>
135 lines
3.9 KiB
Markdown
135 lines
3.9 KiB
Markdown
# VisFactor
|
||
|
||
|
||
## 概述
|
||
|
||
VisFactor 使用从 Factor-Referenced Cognitive Test (FRCT) 改编而来的 20 个以视觉为中心的子测试，评估多模态大语言模型的基础视觉认知能力。它聚焦于支撑高层视觉推理的基本能力，而非衡量单一下游任务的表现。
|
||
|
||
## 任务描述
|
||
|
||
- **任务类型**：包含二元判断和简短自由回答的视觉认知评估
|
||
- **输入**：1 至 4 张图像，与任务特定指令交错排列
|
||
- **输出**：一个 JSON 对象，包含布尔值、单词、数字、坐标对或字母形式的答案
|
||
- **领域**：可视化与空间处理、知觉闭合、视觉记忆及推理
|
||
|
||
## 主要特性
|
||
|
||
- 包含 3,046 行数据，对应 20 个 FRCT 子测试中的 808 个测试项
|
||
- 采用基于规则的变体和分组一致性检查，将平均随机猜测准确率降至约 2.9%
|
||
- 保留 VLMEvalKit 实现中的官方零样本提示及其图像顺序
|
||
- 覆盖隐藏图形识别、格式塔完形、视觉记忆、心理旋转、路径查找、折纸推理等相关能力
|
||
|
||
## 评估说明
|
||
|
||
- 使用官方 `VisFactor.tsv` 在 ModelScope 镜像中的 **test** 划分
|
||
- 提取最后一个 `{"answer": ...}` 对象，并应用官方针对不同类别的标准化规则
|
||
- 单个逻辑测试项可能包含多行数据，仅当所有行均正确时才计为正确
|
||
- 报告每个子测试在测试项级别的准确率；主得分是所涵盖子测试的未加权宏平均值
|
||
- 评分过程完全确定，无需依赖 LLM 评判器
|
||
|
||
## 属性
|
||
|
||
| 属性 | 值 |
|
||
|----------|-------|
|
||
| **基准测试名称** | `visfactor` |
|
||
| **数据集ID** | [lmms-lab-encoder/visfactor](https://modelscope.cn/datasets/lmms-lab-encoder/visfactor/summary) |
|
||
| **论文** | [Paper](https://arxiv.org/abs/2502.16435) |
|
||
| **标签** | `MultiModal`, `QA`, `Reasoning` |
|
||
| **指标** | `accuracy` |
|
||
| **默认示例数** | 0-shot |
|
||
| **评估划分** | `test` |
|
||
|
||
|
||
## 数据统计
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 总样本数 | 3,046 |
|
||
| 提示词长度（平均） | 463.45 字符 |
|
||
| 提示词长度（最小/最大） | 188 / 932 字符 |
|
||
|
||
**图像统计信息：**
|
||
|
||
| 指标 | 值 |
|
||
|--------|-------|
|
||
| 图像总数 | 6,048 |
|
||
| 每样本图像数 | 最小: 1, 最大: 4, 平均: 1.99 |
|
||
| 分辨率范围 | 100x100 - 668x911 |
|
||
| 格式 | jpeg |
|
||
|
||
|
||
## 样例示例
|
||
|
||
**子集**: `default`
|
||
|
||
```json
|
||
{
|
||
"input": [
|
||
{
|
||
"id": "1a4f53fe",
|
||
"content": [
|
||
{
|
||
"text": "Look at the two images:\n\nBelow is the first image, one simple shape:"
|
||
},
|
||
{
|
||
"image": "[BASE64_IMAGE: jpeg, ~2.6KB]"
|
||
},
|
||
{
|
||
"text": "Below is the second image, a larger, complex pattern:"
|
||
},
|
||
{
|
||
"image": "[BASE64_IMAGE: jpeg, ~8.5KB]"
|
||
},
|
||
{
|
||
"text": "Task: Decide whether the shape in the first image is hidden anywhere inside the second image. The shape will never be rotated, flipped, or resized. The shape will always be right-side-up and exactly the same size as in the first image.\n\nOutput: Respond with only one word: “TRUE” if it is present, “FALSE” if it is not, in JSON format as follows: {\"answer\": YOUR_ANSWER_HERE}."
|
||
}
|
||
]
|
||
}
|
||
],
|
||
"target": "T",
|
||
"id": 0,
|
||
"group_id": 0,
|
||
"metadata": {
|
||
"index": 0,
|
||
"category_id": "CF1",
|
||
"category_name": "Hidden Figures Test",
|
||
"eval_index": 0,
|
||
"additional": ""
|
||
}
|
||
}
|
||
```
|
||
|
||
## 提示模板
|
||
|
||
*未定义提示模板。*
|
||
|
||
## 使用方法
|
||
|
||
### 使用 CLI
|
||
|
||
```bash
|
||
evalscope eval \
|
||
--model YOUR_MODEL \
|
||
--api-url OPENAI_API_COMPAT_URL \
|
||
--api-key EMPTY_TOKEN \
|
||
--datasets visfactor \
|
||
--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=['visfactor'],
|
||
limit=10, # 正式评估时请删除此行
|
||
)
|
||
|
||
run_task(task_cfg=task_cfg)
|
||
```
|