---
license: cc-by-nc-sa-4.0
language:
- en
size_categories:
- 100K<n<1M
---

# Dataset Card for **SafeVid‑350K**

## Dataset Summary
`SafeVid‑350K` is a large‑scale, **video‑centric safety alignment dataset** designed to instill *Helpful, Honest, Harmless (3H)* principles in Video Large Multimodal Models (VLMMs).  It contains **350 000 preference pairs** (chosen vs. rejected responses) covering **30 scene categories** and **29 fine‑grained safety sub‑dimensions**.  Each example links to a YouTube video segment via `YoutubeID + start/end timestamp`, includes a rich textual description, and provides adversarial questions explicitly crafted to probe safety vulnerabilities.

## Supported Tasks
| Task | Description | Metrics |
|------|-------------|---------|
| **Safety‑aligned response generation** | Given a video segment + adversarial question, produce a safe, helpful answer | Safety Rate|

## Dataset Structure

### Data Fields
| Field | Type | Description |
|-------|------|-------------|
| `YoutubeID` | `string` | YouTube video ID |
| `Start_timestamp` / `End_timestamp` | `string` | `"HH:MM:SS.mmm"` segment boundaries |
| `Caption` | `string` | Video auto‑caption (original language) |
| `Aesthetic_Score` | `float` | Aesthetic quality from CLIP‑based predictor |
| `UMT_Score` | `float` | Temporal saliency score from UMT |
| `Language` | `string` | ISO‑639‑1 code of caption language |
| `Type` | `string` | Original InternVid top‑level category |
| `description` | `string` | **GPT‑4–refined** dense narration (≈100 words) |
| `⟨3H‑subcategory⟩` | `list[string]` | Adversarial questions for that safety dimension |
| `Response_pairs` | `dict` | “rejected_response vs. chosen_response” for each question |

> ⚠️ **Video bytes are *not* redistributed**; only metadata + IDs are included to respect YouTube TOS.

### File Layout
```text
data/                 # 30 JSON files, one per scene
  ├─ man-made environment_building_business_filtered.json
  ├─ natural environment_land_beach_filtered.json
  ⋯