dataset_info:
features:
- name: text
dtype: string - name: label
dtype: int64
splits: - name: train
num_bytes: 108963370
num_examples: 77696 - name: validation
num_bytes: 13531170
num_examples: 9712 - name: test
num_bytes: 13931350
num_examples: 9713
download_size: 75438943
dataset_size: 136425890
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: validation
path: data/validation-* - split: test
path: data/test-*
license: cc-by-4.0
task_categories:
- split: train
- text-classification
language: - en
pretty_name: Harmfull and Harmless prompts library
size_categories: - 10K<n<100K
Harmful vs Harmless Prompts Dataset
This dataset aggregates multiple sources of jailbreak prompts, harmful queries, and benign prompts.
Labels
- 0 = harmless / regular prompt
- 2 = harmful or jailbreak-related prompt
Sources
Includes datasets from:
- TrustAIRLab in-the-wild jailbreak prompts (MIT License)
- DiegoAI597 harmful actions [Apache 2.0]
- djapp18 JailbreaksOverTime [CC-BY-4.0]
- Bravansky compact jailbreaks [MIT]
Splits
- train
- validation
- test
Some original splits were randomly partitioned to ensure balanced validation and test sets.
License: CC BY 4.0 (dataset compilation and annotations)
Underlying texts remain under their original licenses:
- MIT
- Apache-2.0 (with NOTICE preserved)
- CC BY 4.0
Intended Use
- AI safety research
- adversarial robustness evaluation
- prompt classification tasks