Dataset · allenai
Prompts for evaluating toxic generation
allenai/real-toxicity-prompts
100K English web snippets with metadata for checking neural network toxic degeneration.
What's inside
RealToxicityPrompts is a set of 100K sentence fragments from the web. Each example contains a prompt and metadata: file name, begin/end positions, and more. Language is English; data is presented in tabular and text format. The dataset is intended for studying the risk of neural toxic degeneration in models.
What you can do with it
- Evaluate LLM toxicity
- Filter training data
- Test safety filters
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 19.6K
- Likes
- 125
Models trained on it
- abacusai/Smaug-Toxicity-PerspectiveText classification 214
- INC4AI/neural-chat-7b-v1-1Text generation 142
- sarimahsan101/qwen2.5-0.5b-detox-ppoText generation 38
- opus-research/opus-moderation-3Text classification 26
- opus-research/opus-moderation-4-largeText classification 23
- opus-research/opus-moderation-4-fastText classification 22
- ZetangForward/SegmentCNNText classification 16
- sarimahsan101/Qwen2.5-0.5B-DPO-detox 14
How to load it
from datasets import load_dataset
ds = load_dataset("allenai/real-toxicity-prompts")