EuroSafeGuard v1
Collection
Open constitutional classifiers: a probe-escalated guardrail for LLMs, its CC++ recreation and training prompts. • 3 items • Updated • 2
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains harmful requests, including jailbreak-wrapped requests, collected to train and evaluate safeguards. Access is reviewed by hand by the EuroSafeAI team. Please describe who you are and, in at least 50 words, the research or safety purpose you will use the data for; requests without a clear purpose, without an institutional affiliation or with fewer than 50 words of purpose are declined. By requesting access you agree to use the data only to build, evaluate or study safety classifiers and defences, never to train models to comply with harmful requests, not to redistribute it, and to respect the license of each source.
Log in or Sign Up to review the conditions and access this dataset content.
EuroSafeGuard v1: training prompts
Prompts used to train the input classifier and the activation probe of EuroSafeGuard v1.
Content warning: contains harmful requests, many wrapped in jailbreak templates and obfuscations (no model replies).
Paper · Code · Collection
Contents
| Config | Trains | Rows | Withheld text |
|---|---|---|---|
gate (default) |
input classifier (Qwen3.5-9B, 11 heads) | 540,536 | 129,478 |
probe |
activation probe (Llama-3.1-8B-Instruct) | 116,524 | 32,747 |
from datasets import load_dataset
ds = load_dataset("EuroSafeAI/EuroSafeGuard-v1-prompts", "gate", split="train")
Access
Request access on this page. Requests are reviewed by hand: give your name, affiliation, institutional email and at least 50 words on the safety research you will use the data for. Use is limited to building, evaluating or studying safeguards; no training models to comply with harmful requests, no redistribution.
Columns
| Column | Meaning |
|---|---|
row_id, input_id |
row id; input_id groups a prompt with its rewrites |
text |
the prompt (empty for withheld rows) |
text_sha256 |
sha256(text.strip()) |
category |
one of 11 harm categories, or Benign |
entry_type |
harmful, dual_use_harmful, dual_use_benign, benign, general_benign |
source, attack |
origin of the row; jailbreak or obfuscation applied (none if plain) |
license |
license of the row's source |
restricted, reconstructable, text_source |
text withheld / rebuildable by hash / dataset to rebuild from |
label (probe only) |
binary training label |
Gate targets: harmful 1 on its category's head, dual_use_harmful 0.8, dual_use_benign 0.5, benign and
general_benign 0.
Withheld rows
Rows copied verbatim from gated or non-redistributable sources (CeSIA constitution datasets, WildGuardMix,
WildJailbreak, WildChat-1M, SOSBench) ship without text. Accept each source's terms, huggingface-cli login, then:
python reconstruct.py --in gate_train.parquet --out gate_train_full.parquet
License
Each row carries its source's license. Several sources are non-commercial or share-alike, so the dataset as a whole is for non-commercial research use.
| License | Sources |
|---|---|
| CC-BY-NC-SA-4.0 (ours, from the CeSIA constitution) | redact_*, plain, plain_rescued, paraphrase, jf, dan_*, bon_*, obf_* |
| CC-BY-NC-SA-4.0, gated (withheld) | cesia_aug |
| ODC-BY (withheld) | wildguardmix_train, wjb_* (WildJailbreak), wildchat_train |
| CC-BY-NC-SA + Responsible Use Agreement (withheld) | sosbench_* |
| CC-BY-NC-4.0 | expertbio_*, falsereject, camel_*, alpaca_train, toxicchat_train |
| CC-BY-4.0 | aegis2_train, trigger_benign (ours), gib_aug (ours) |
| CC-BY-SA | dolly (3.0), histmisinfo_* (4.0), arc_* (4.0) |
| MIT | ultrachat_train, mmlu_auxtrain, gsm8k_train, pubmedqa, phtest |
| Apache-2.0 | openbookqa, oasst_train |
| CC0-1.0 | civil_* |
| Mixed (MIT, CC-BY-SA-3.0, CC-BY-NC-4.0, CC-BY-4.0) | benign_ext |
Citation
@misc{waibl2026eurosafeguard,
title = {EuroSafeGuard: Open Constitutional Classifiers},
author = {Waibl, Leonhard and Chen, Yen-Shan and Simko, Samuel and Thomas, Aditya and
Singh, Arth and Diwan, Nirav and Sch{\"o}lkopf, Bernhard and Jin, Zhijing},
url = {https://guard.safe.eu/paper},
year = {2026}
}
- Downloads last month
- 64