Dataset · OpenSafetyLab
LLM safety evaluation benchmark and dataset
OpenSafetyLab/Salad-Data
A collection of 21,318 prompts and attacks (jailbreaks, toxicity) for testing and improving neural network safety.
What's inside
The base set contains 21,318 text and tabular entries, predominantly in English (with a small multilingual portion), compiled to evaluate model safety. The corpus combines GPT-3.5-generated queries (15,433) with data from open sources such as HH-harmless, HH-red-team, Advbench, ToxicChat, and Do-Not-Answer.
What you can do with it
- Testing LLM resilience against jailbreaks and harmful content generation
- Training moderation classifiers to filter neural network prompts and responses
- Tasks
- Text classification, Text generation
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 1.9K
- Likes
- 37
Models trained on it
- protectai/deberta-v3-base-prompt-injection-v2Text classification 380.7K
- protectai/deberta-v3-small-prompt-injection-v2Text classification 6.9K
- vllm-sr/mmbert32k-jailbreak-detector-mergedText classification 2.8K
- vllm-sr/Vela-2.0-0.8BZero shot classification 2K
- vllm-sr/Vela-2.0-0.3BZero shot classification 2K
- Horizon-Labs/content-safety-guard-smallText classification 1.1K
- vllm-sr/Vela-1.0-Encoder-307M-ShieldText classification 762
- Horizon-Labs/content-safety-guard-baseText classification 341
How to load it
from datasets import load_dataset
ds = load_dataset("OpenSafetyLab/Salad-Data")