ShreyashDhoot/KTO_trial

---
language: en
license: mit
task_categories:
- image-classification
tags:
- image
- safety
- adversarial
- inpainting
- kto
dataset_info:
features:
- name: "image"
dtype: image
- name: "prompt"
dtype: string
- name: "id"
dtype: string
- name: "disturbing"
dtype: int8
- name: "hate"
dtype: int8
- name: "illegal activity" # Keep the quotes here
dtype: int8
- name: "safe"
dtype: int8
- name: "sexual"
dtype: int8
- name: "violence"
dtype: int8
- name: "feathered_mask"
dtype: image
- name: "image_masked_removed"
dtype: image
---

# KTO Training Dataset

Processed from [kricko/cleaned_auditor](https://huggingface.co/datasets/kricko/cleaned_auditor) using the Auditor model.

## Description
Each example contains the original image alongside adversarial heatmaps, feathered masks,
and masked images with detected unsafe regions blacked out.

## Features
| Column | Type | Description |
|---|---|---|
| `image` | Image | Original input image |
| `prompt` | string | Text prompt associated with the image |
| `id` | string | Unique identifier |
| `disturbing` | int8 | Disturbing content score |
| `hate` | int8 | Hate content score |
| `illegal activity` | int8 | Illegal activity score |
| `safe` | int8 | Safe content score |
| `sexual` | int8 | Sexual content score |
| `violence` | int8 | Violence content score |
| `feathered_mask` | Image | Feathered adversarial mask (sigma=5, 75th percentile threshold) |
| `image_masked_removed` | Image | Original image with adversarial regions blacked out |

## Processing Details
- **Threshold**: 75th percentile of adversarial heatmap
- **Morphological dilation**: 5×5 kernel, 2 iterations
- **Feathering**: Gaussian blur sigma=5.0
- **Total examples**: 27553