KTO_trial
收藏资源简介:
KTO训练数据集是从'kricko/cleaned_auditor'数据集经过Auditor模型处理得到的。每个样本包含原始图像、对抗性热图、羽化掩码以及检测到不安全区域被涂黑的掩码图像。数据集包含以下字段:'image'(原始输入图像)、'prompt'(相关文本提示)、'id'(唯一标识符)、'disturbing'(令人不安内容评分)、'hate'(仇恨内容评分)、'illegal activity'(非法活动评分)、'safe'(安全内容评分)、'sexual'(性内容评分)、'violence'(暴力内容评分)、'feathered_mask'(羽化对抗掩码)和'image_masked_removed'(不安全区域被涂黑的图像)。处理参数包括:75百分位对抗热图阈值、5×5核的形态学膨胀(2次迭代)、sigma=5.0的高斯模糊羽化。数据集总样本量为27,553个,适用于图像分类、内容安全评估等任务。
The KTO training dataset is derived from the 'kricko/cleaned_auditor' dataset after being processed by the Auditor model. Each sample consists of the original image, adversarial heatmap, feathered mask, and the masked image where detected unsafe regions are blacked out. The dataset includes the following fields: 'image' (original input image), 'prompt' (relevant text prompt), 'id' (unique identifier), 'disturbing' (disturbing content score), 'hate' (hate content score), 'illegal activity' (illegal activity score), 'safe' (safe content score), 'sexual' (sexual content score), 'violence' (violence content score), 'feathered_mask' (feathered adversarial mask), and 'image_masked_removed' (image with detected unsafe regions blacked out). The processing parameters are as follows: 75th percentile adversarial heatmap threshold, morphological dilation with a 5×5 kernel (2 iterations), and Gaussian blur feathering with sigma=5.0. The dataset has a total of 27,553 samples and is applicable to tasks such as image classification and content safety assessment.
KTO Training Dataset 概述
数据集基本信息
- 语言:英语
- 许可证:MIT
- 任务类别:图像分类
- 标签:图像、安全、对抗性、修复、KTO
- 来源:数据集由 kricko/cleaned_auditor 经Auditor模型处理得到。
数据集内容与规模
- 总样本数:27553
- 核心内容:每个样本包含原始图像、对抗性热图、羽化掩码以及将检测到的不安全区域涂黑后的掩码图像。
数据特征(Features)
| 列名 | 数据类型 | 描述 |
|---|---|---|
image |
图像 | 原始输入图像 |
prompt |
字符串 | 与图像关联的文本提示 |
id |
字符串 | 唯一标识符 |
disturbing |
int8 | 令人不安内容评分 |
hate |
int8 | 仇恨内容评分 |
illegal activity |
int8 | 非法活动评分 |
safe |
int8 | 安全内容评分 |
sexual |
int8 | 性相关内容评分 |
violence |
int8 | 暴力内容评分 |
feathered_mask |
图像 | 羽化对抗掩码(sigma=5,75百分位阈值) |
image_masked_removed |
图像 | 对抗区域被涂黑后的原始图像 |
数据处理细节
- 阈值:对抗性热图的第75百分位
- 形态学膨胀:5×5内核,2次迭代
- 羽化处理:高斯模糊 sigma=5.0





