README.md

---
dataset_info:
  features:
  - name: Image
    dtype: image
  - name: Decision
    dtype: string
  splits:
  - name: train
    num_bytes: 2506019114
    num_examples: 30622
  download_size: 2508835535
  dataset_size: 2506019114
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: apache-2.0
task_categories:
- image-to-text
- image-text-to-text
language:
- en
tags:
- classification
- image
size_categories:
- 10K<n<100K
---

> [!IMPORTANT]
> This dataset contains sensitive visual content intended for training Image Guard, Image Filter, and Image Shield models. Due to the sensitive nature of the data, access is gated and requires manual review. Access may be approved for users who provide a valid research, safety, or development purpose.

# **ImageShield-OneDecision-Classification**

**ImageShield-OneDecision-Classification** is a refined multimodal classification dataset derived directly from [`prithivMLmods/ImageShield-Guardrail-Pro`](https://huggingface.co/datasets/prithivMLmods/ImageShield-Guardrail-Pro)[cite: 1]. It converts extensive multimodal guardrail samples into a structured, single-turn format targeted at **one-pass decision-making for unsafe, sensitive, and policy-violating imagery**. Each entry pairs a visual input with a serialized JSON string in the `Decision` column, delivering scene narration, classification rationale, boolean flags, and numeric class indicators[cite: 1].

## Dataset Overview

* **Repository:** `prithivMLmods/ImageShield-OneDecision-Classification`[cite: 1]
* **Parent Dataset:** [ImageShield-Guardrail-Pro](https://huggingface.co/datasets/prithivMLmods/ImageShield-Guardrail-Pro)[cite: 1]
* **Collection:** ImageShield-Guardrail Datasets[cite: 1]
* **Dataset Size:** 10K < n < 100K samples (~2.51 GB)[cite: 1]
* **Modalities:** Image, Text[cite: 1]
* **Split:** `train`[cite: 1]
* **License:** Apache-2.0[cite: 1]

## Dataset Structure

The dataset contains two primary fields[cite: 1]:

| Field | Type | Description |
| :--- | :--- | :--- |
| `image`[cite: 1] | `Image`[cite: 1] | High-resolution input image across diverse visual domains (real-world photos, digital art, illustrations, indoor/outdoor scenes). |
| `Decision`[cite: 1] | `string`[cite: 1] | A serialized JSON object containing scene descriptions, content safety rationales, boolean indicators, and binary moderation labels. |

### Decision Schema

The serialized `Decision` JSON string follows this structure:

```json
{
  "scene_description": "A close-up shot captures an adult male and a young boy working together at a craft table. The man, wearing a plaid shirt, leans over to assist the boy, who is focused on assembling small objects using scissors and glue. Both individuals appear engaged in a collaborative activity, likely building or crafting something together.",
  "is_nsfw": false,
  "classification_reason": "The image depicts a wholesome scene of two children and an adult collaborating on a creative project. There is no evidence of nudity, revealing swimwear, suggestive poses, or any sexually explicit content. The focus remains entirely on the educational and social interaction between the subjects.",
  "nsfw": 0,
  "safe": 1
}
```

#### Field Explanations

* **`scene_description`** (`string`): Comprehensive, objective description of the visual contents, entities, activities, and visual context.
* **`is_nsfw`** (`bool`): Boolean flag indicating whether the image violates safety guidelines (`true`) or is benign (`false`).
* **`classification_reason`** (`string`): Grounded natural language justification explaining why the sample is classified as safe or unsafe.
* **`nsfw`** (`int`): Binary indicator (`1` for unsafe/NSFW, `0` for safe).
* **`safe`** (`int`): Binary indicator (`1` for safe/benign, `0` for unsafe/NSFW).

## Quickstart & Usage

### Loading with Hugging Face Datasets

```python
import json
from datasets import load_dataset

# Load the train split
dataset = load_dataset("prithivMLmods/ImageShield-OneDecision-Classification", split="train")

# Retrieve and parse a sample
sample = dataset[0]
image = sample["image"]
decision_meta = json.loads(sample["Decision"])

print("Scene:", decision_meta["scene_description"])
print("Is Safe:", bool(decision_meta["safe"]))
print("Reason:", decision_meta["classification_reason"])
```

### Formatting for Vision-Language SFT (e.g., Qwen-VL, LLaVA)

```json
{
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "image"},
        {"type": "text", "text": "Analyze the visual content of this image, evaluate its safety status, and return your moderation decision as structured JSON."}
      ]
    },
    {
      "role": "assistant",
      "content": "{\"scene_description\": \"...\", \"is_nsfw\": false, \"classification_reason\": \"...\", \"nsfw\": 0, \"safe\": 1}"
    }
  ]
}
```

## Use Cases

1. **Multimodal Content Moderation:** Train vision-language models and classifier backbones to filter unsafe visual streams before passing them downstream.
2. **Explainable Guardrails:** Enable safety agents to output step-by-step reasoning (`classification_reason`) alongside binary labels to reduce false positive flags.
3. **Automated Auditing:** Benchmark existing multimodal systems against borderline edge cases, stylized imagery, and contextual safety scenarios.

## Lineage & Citation

This dataset is refined from **ImageShield-Guardrail-Pro** to provide unified, single-step decision targets for safety classification pipelines.

```bibtex
@dataset{imageshield_onedecision_classification,
  author    = {PrithivMLmods},
  title     = {ImageShield-OneDecision-Classification},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {[https://huggingface.co/datasets/prithivMLmods/ImageShield-OneDecision-Classification](https://huggingface.co/datasets/prithivMLmods/ImageShield-OneDecision-Classification)}
}
```