Dataset · PKU-Alignment
QA pairs with harm annotations
PKU-Alignment/BeaverTails
100K+ human-annotated QA pairs across 14 harm categories for AI safety.
What's inside
BeaverTails is an AI safety dataset collection. The repository features human annotations of question-answer pairs with harm category labels; a single pair may belong to multiple categories. It covers 14 categories, including animal abuse. Language is English, modality is text.
What you can do with it
- Train a harmful response classifier
- Test LLM safety filters
- Create LoRA for refusing harmful prompts
- Tasks
- Text classification
- Languages
- English
- Size
- 100K–1M rows
- Downloads this month
- 21.8K
- Likes
- 119
Models trained on it
- OpenSafetyLab/MD-Judge-v0_2-internlm2_7bText generation 7.1K
- PKU-Alignment/beaver-dam-7b 4.3K
- mradermacher/MD-Judge-v0_2-internlm2_7b-i1-GGUF 1.3K
- mradermacher/torchcast-decision-12b-GGUF 543
- OpenSafetyLab/MD-Judge-v0.1Text generation 526
- mradermacher/Stentor-30M-Instruct-heretic-safety-defiltered-i1-GGUFText generation 499
- mradermacher/MD-Judge-v0_2-internlm2_7b-GGUF 325
- mradermacher/Stentor-12M-Instruct-GGUFText generation 252
How to load it
from datasets import load_dataset
ds = load_dataset("PKU-Alignment/BeaverTails")