# captcha数据使用说明 ## 数据概览 本数据集包含8类专用验证码数据,每类任务包含2k条数据,用于训练多模态验证码识别模型。 | 任务类型 | SFT数据文件 | 图像数据目录 | | -------- | ------------------------------ | ------------------------ | | 图片连接 | `captcha_connect_icon.jsonl` | `captcha/connect_icon` | | 按序点击 | `click_order_dataset.jsonl` | `captcha/click_order` | | 坐标定位 | `coordinates_dataset.jsonl` | `captcha/coordinates` | | 飞镖计数 | `dart_counting_dataset.jsonl` | `captcha/dart_counting` | | 骰子计数 | `dice_counting_dataset.jsonl` | `captcha/dice_counting` | | 图像选择 | `image_select_dataset.jsonl` | `captcha/image_select` | | 区域选择 | `patch_select_dataset.jsonl` | `captcha/patch_select` | | 旋转匹配 | `rotation_match_dataset.jsonl` | `captcha/rotation_match` | ## 数据结构说明 ### 文件组织形式 - **图像文件夹**:每个样本使用独立文件夹管理 - **样本文件夹内容**: - SFT训练所需的图片文件 - 一个JSON配置文件,包含: - 各按钮的位置信息 - 训练标签值 - 测试时的辅助信息 > **注意**:如果仅用于模型训练,可以忽略JSON配置文件。该文件主要用于SFT数据制作和测试集信息提供。 ## TuningFactory配置示例 在TuningFactory中使用该数据集进行SFT训练时,`dataset_info.json` 的配置方式如下: ```python "captcha_image_select": { "file_name": "/data/cpfs_0/wangpengcheng/data/image_select_dataset.jsonl", "formatting": "sharegpt", "columns": { "messages": "conversations", "system": "system", "image": "image" }, "tags": { "role_tag": "role", "content_tag": "content", "user_tag": "user", "assistant_tag": "assistant" } } ``` ## 使用建议 1. **数据路径**:请根据实际存储位置修改配置文件中的 `file_name` 路径 2. **批处理**:可以根据计算资源情况调整批处理大小和训练策略