license: mit
task_categories:
- text-classification
language: - en
tags: - safety
- guardrails
- jailbreak
- adversarial
- llm-safety
size_categories: - 10K<n<100K
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: calibration
path: data/calibration-* - split: test
path: data/test-* - split: steering_test
path: data/steering_test-*
dataset_info:
features: - name: prompt
dtype: string - name: adversarial
dtype: bool
splits: - name: train
num_bytes: 16959651
num_examples: 7740 - name: calibration
num_bytes: 2164382
num_examples: 1105 - name: test
num_bytes: 4571506
num_examples: 2211 - name: steering_test
num_bytes: 5023230
num_examples: 1772
download_size: 16711827
dataset_size: 28718769
- split: train
guard-glp-data
A prompt-level binary safety dataset for training and evaluating LLM guardrails,
with a focus on adversarial / jailbreak robustness.
Built as part of the Generative Latent Prior (GLP) guard project.
Dataset summary
Each row has two fields:
| Field | Type | Description |
|---|---|---|
prompt |
string |
The user-facing input text |
adversarial |
bool |
True = harmful / jailbreak; False = benign |
Approximate split sizes (randomly assigned per sample, seed not fixed):
| Split | Approx. size | Benign share | Malicious share |
|---|---|---|---|
| train | ~7 700 | ~3 500 | ~4 200 |
| calibration | ~1 100 | ~500 | ~600 |
| test | ~2 200 | ~1 000 | ~1 200 |
Total pool: 5 000 benign + 6 057 malicious ≈ 11 057 samples,
split 70 / 10 / 20 % (train / calibration / test) via per-sample random draw.
Sources
| Dataset | Role | Field used |
|---|---|---|
| HuggingFaceFW/fineweb | benign (5 000 samples) | text from sample-10BT, shuffled with seed 42 |
| centrepourlasecuriteia/jailbreak-dataset | malicious (5 507 samples) | prompt field, all non-Benign categories, shuffled with seed 42 |
| ddidacus/harmeval-gcg-llama3-1b | malicious (550 samples) | adv_prompt field (non-null entries) |
Split construction
Prompts and labels are collected from the three sources above.
Each sample is independently and randomly assigned to one of three splits:
- calibration with probability 0.10
- test with probability 0.20
- train with probability 0.70
All splits are shuffled after construction.
Intended use
This dataset is intended for:
- Training prompt-level safety classifiers (e.g. a GLP-based guard).
- Calibrating confidence thresholds for guardrail systems.
- Evaluating robustness against adversarial / jailbreak prompts.
Citation
If you use this dataset please cite the upstream sources:
@misc{fineweb,
title = {FineWeb},
author = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/HuggingFaceFW/fineweb}}
}
@misc{centrepourlasecuriteia_jailbreak,
title = {jailbreak-dataset},
author = {Centre pour la Sécurité de l'IA},
howpublished = {\url{https://huggingface.co/datasets/centrepourlasecuriteia/jailbreak-dataset}}
}