datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
detect-jailbreak
Content Warning:
This dataset contains unsafe model responses and user queries. Viewers may find the content disturbing.
Overview
Our evaluation dataset combines three existing datasets with custom augmentations to create a robust framework for assessing LLM vulnerabilities and defense effectiveness. The core components are the Verazuo dataset, the ZHX123 benchmark, and the Weapons of Mass Destruction Proxy (WMDP) dataset.
Credits and Citations
Our greatest… See the full description on the dataset page: https://huggingface.co/datasets/GuardrailsAI/detect-jailbreak.guardrail-tr
Guardrail-TR
Guardrail-TR (ytu-ce-cosmos/guardrail-tr) is a large-scale Turkish prompt-safety dataset for training and evaluating content-moderation / guardrail classifiers.
Dataset contains ~405K single-turn user prompts with:
a binary safe / unsafe label, and
multi-label hazard categories (a prompt may carry more than one category).
All released prompt text is Turkish. Rows that originated from English sources were processed, translated and, where relevant, culturally… See the full description on the dataset page: https://huggingface.co/datasets/ytu-ce-cosmos/guardrail-tr.arabic-guardrail
Arabic Guardrail — 250,842 rows, 12 classes
Defensive dataset for training Arabic prompt-safety classifiers. Each row is an incoming user
message and which of 12 safety classes it belongs to.
بالعربية: مجموعة بيانات عربية لتدريب نماذج تصنّف الرسائل الواردة قبل وصولها للمساعد الذكي.
Arabic guardrails were a gap. Hugging Face searches for Arabic jailbreak / safety /
prompt-injection datasets return zero results, and the one Arabic guardrail model… See the full description on the dataset page: https://huggingface.co/datasets/oddadmix/arabic-guardrail.guardrail-hard-negatives
Guardrail Hard Negatives (EN/TR)
A false-positive stress test for guardrails. A curated, paired benign/attack dataset for evaluating and training prompt-injection detectors and LLM guardrails. A bilingual false-positive challenge set: benign prompts that look like attacks (security researchers asking about injection, authorized admin actions, quoted payloads, legitimate roleplay) paired against real attacks, so you can measure the false-positive rate your users will actually… See the full description on the dataset page: https://huggingface.co/datasets/3nesdeniz/guardrail-hard-negatives.toxic-guardrail-id-en
Toxic Guardrail ID/EN
A normalized dataset for training a bilingual (Indonesian + English) toxicity
severity classifier intended for use as an application guardrail. Several source
datasets with incompatible label schemas are mapped onto a single ordinal
severity scale.
Curation runs through a deterministic TypeScript pipeline with a seeded PRNG, so
the splits are reproducible rather than the output of an ad-hoc notebook.
Rating scale
rating
meaning
default… See the full description on the dataset page: https://huggingface.co/datasets/laskar-ks/toxic-guardrail-id-en.Evaluation-Dataset-of-AI-Agent-Security-Guardrails
DKnownAI Agent Security Evaluation Dataset
Data Fields
Field
Type
Description
text
string
The adversarial input (prompt) to be evaluated by a security guardrail
action
string
Human-annotated label: blocked or allowed
Citation
@misc{li2026comparativeevaluationaiagent,
title={A Comparative Evaluation of AI Agent Security Guardrails},
author={Qi Li and Jiu Li and Pingtao Wei and Jianjun Xu and Xueyi Wei and Jiwei Shi and Xuan… See the full description on the dataset page: https://huggingface.co/datasets/CaiZhiTech/Evaluation-Dataset-of-AI-Agent-Security-Guardrails.context-adherence-guardrail-10k
Context-Adherence Guardrail — training data (10,710)
Training data for a single-token RAG-groundedness guardrail. Each item is a
(question, context, response) triple with a human- or construction-derived
PASS/FAIL label under one Behavior Spec:
FAIL iff the response makes at least one factual claim unsupported by or
contradicting the retrieved context — truth in the real world is irrelevant
(strict grounding). PASS otherwise, including responses that decline to
answer for lack… See the full description on the dataset page: https://huggingface.co/datasets/aaryand/context-adherence-guardrail-10k.production-ai-guardrail-evals
Production AI Guardrail Evals
Eighteen synthetic, assertion-bearing cases for testing whether a language model can stay inside an advisory role. The cases cover roadmap intake, release readiness, and catalog-change review—the same workload families I use in the Winwood AI Toolkit around IEM Rig.
This is a sanitized public derivative, not a dump of application logs or the private evaluation corpus.
What each row contains
case_id: stable public identifier;… See the full description on the dataset page: https://huggingface.co/datasets/mattwinwood/production-ai-guardrail-evals.guardrail4agent-sft
Guardrail4Agent SFT Dataset
AI 에이전트 Tool Call 데이터 유출 탐지를 위한 한국어 학습 데이터셋입니다.
카테고리
레이블
설명
건수
safe
정상 Tool Call
2,380
S1
PII 유출
1,190
S2
Credential 유출
1,190
S3
Prompt Injection
1,190
S4
내부 데이터 외부 전달
630
S5
시스템 프롬프트 추출
420
분할
data/sft_train.jsonl — 4,900건 (70%)
data/sft_val.jsonl — 1,050건 (15%)
data/sft_test.jsonl — 1,050건 (15%)
agent-symbolic-guardrailsThis dataset contains data associated with the paper Symbolic Guardrails for Domain-Specific Agents: Stronger Safety and Security Guarantees Without Sacrificing Utility.
Code: https://github.com/hyn0027/agent-symbolic-guardrails
Subsets
literature_review
This subset contains the metadata of the systematic literature review data. Details are discussed in Section 3 in the paper.
adversarial_MedAgentBench
This subset contains the adversarial tasks we… See the full description on the dataset page: https://huggingface.co/datasets/hyn0027D/agent-symbolic-guardrails.guardrails-datasetsynth-health-guardrails
Synth Health Guardrails — consensus-label release
Private release candidate. Licensed under CDLA-Permissive-2.0.
The repository remains private until League Labs publishes the release.
Sensitive-content notice. Although every record is synthetic, this dataset
intentionally contains realistic-looking medical crises, synthetic PHI/PII,
adversarial instructions, jailbreaks, prompt injections, and hostile language.
Synth Health Guardrails contains 13,713 synthetic member… See the full description on the dataset page: https://huggingface.co/datasets/LeagueLabs/synth-health-guardrails.korean_guardrail_testguardrail-arena
guardrail-arena — a two-axis, multilingual guardrail benchmark
from datasets import load_dataset
ds = load_dataset("fevziegeyurtsevenler/guardrail-arena")
Score any LLM prompt-injection guardrail on two axes at once — does it catch attacks
(miss-rate down) and does it avoid over-blocking benign traffic (over-refusal down) — in
English and Turkish. A guardrail is any callable text -> 0|1.
337 prompts: 217 injections (EN 110 + TR 107) · 80 plain benign · 40 security-adjacent… See the full description on the dataset page: https://huggingface.co/datasets/fevziegeyurtsevenler/guardrail-arena.agentic-execution-guardrail-eval
Agentic Execution Guardrail Eval
Agentic Execution Guardrail Eval is a lightweight evaluation dataset for testing safety guardrails in agentic execution environments.
The dataset focuses on risky patterns that may appear when AI agents generate prompts, shell commands, or code intended for execution.
Scope
The evaluation cases cover:
Shell execution and exfiltration patterns
Hidden instructions
Prompt injection attempts
Jailbreak-style reframing
Risky Python code… See the full description on the dataset page: https://huggingface.co/datasets/ole-tail/agentic-execution-guardrail-eval.harmful-prompts-gemini-pro-after-guardrail-evaluationImageShield-Guardrail-Pro
[!IMPORTANT]
This dataset contains sensitive visual content intended for training Image Guard, Image Filter, and Image Shield models. Due to the sensitive nature of the data, access is gated and requires manual review. Access may be approved for users who provide a valid research, safety, or development purpose.
ImageShield-Guardrail-Pro
ImageShield-Guardrail-Pro is an advanced multimodal dataset containing 30,630 high-resolution samples designed for training, supervised… See the full description on the dataset page: https://huggingface.co/datasets/prithivMLmods/ImageShield-Guardrail-Pro.ImageShield-Guardrail-Multimodal-20K
[!IMPORTANT]
This dataset contains sensitive visual content intended for training Image Guard, Image Filter, and Image Shield models. Due to the sensitive nature of the data, access is gated and requires manual review. Access may be approved for users who provide a valid research, safety, or development purpose.
ImageShield-Guardrail-Multimodal-20K
ImageShield-Guardrail-Multimodal-20K is a rich multimodal dataset consisting of 28,000 paired image-text instances designed for… See the full description on the dataset page: https://huggingface.co/datasets/prithivMLmods/ImageShield-Guardrail-Multimodal-20K.veto-guardrail-30k
Veto Guardrail 30K
A specialized dataset for training AI security guardrail models to evaluate tool calls against configurable security policies.
Model
This dataset was created to train Veto Warden 4B — a fast, specialized model for real-time tool call security validation.
Overview
Metric
Value
Examples
30,000
Format
Conversational (ShareGPT)
Task
Security policy evaluation
Domains
8 specialized categories
Task Description
The… See the full description on the dataset page: https://huggingface.co/datasets/ycaleb/veto-guardrail-30k.CONSCENDI-guardrail-benchmark
Dataset Card for "CONSCENDI-guardrail-benchmark"
More Information needed
qwen-0.5b_safety_guardrail_dataguardrail-training-data
Guardrail Training Data
A comprehensive collection of 3,978,555 labeled samples across 26 harm categories for training AI safety classifiers.
Dataset Description
This dataset contains both harmful and benign samples designed for training guardrail models that can detect and classify harmful content.
Dataset Structure
text: The text content to be classified
is_safe: Boolean indicating if the content is safe (False = harmful, True = safe)
category: Primary harm… See the full description on the dataset page: https://huggingface.co/datasets/budecosystem/guardrail-training-data.khp-youth-mental-health-guardrail
KHP Youth Mental Health Safety Guardrail Dataset
A synthetic multi-turn conversational dataset for training and evaluating input guardrails for AI assistants serving youth in mental distress. The dataset targets 9 distress signal categories and provides a binary high-risk label for classification, decoupled from the individual signal flags.
GitHub repository: Aser97/Guardrail-For-Agents
Dataset Summary
Split
Rows
High-risk
Low-risk
Train
~1578
~789
~789… See the full description on the dataset page: https://huggingface.co/datasets/AserLompo/khp-youth-mental-health-guardrail.PlaceboBench
Dataset Card
Dataset Description
PlaceboBench is a hallucination benchmark for retrieval-augmented generation (RAG) in the pharmaceutical domain. It is based on real clinical questions submitted by healthcare professionals to Swedish and Norwegian drug information centers (SVELIC/RELIS), answered by seven state-of-the-art LLMs using retrieved European Medicines Agency (EMA) product information documents as context.
The dataset contains 69 questions spanning 23 drugs, with… See the full description on the dataset page: https://huggingface.co/datasets/blue-guardrails/PlaceboBench.harmful-prompts-gemini-pro-before-guardrail-evaluationfinance-guardrail-v2hallucinationThis is a vendored reupload of the Benchmarking Unfaithful Minimal Pairs (BUMP) Dataset available at https://github.com/dataminr-ai/BUMP
The BUMP (Benchmark of Unfaithful Minimal Pairs) dataset stands out as a superior choice for evaluating hallucination detection systems due to its quality and realism. Unlike synthetic datasets such as TruthfulQA, HalluBench, or FaithDial that rely on LLMs to generate hallucinations, BUMP employs human annotators to manually introduce errors into summaries… See the full description on the dataset page: https://huggingface.co/datasets/GuardrailsAI/hallucination.clinical-tpib-pathway-stability-and-risk-guardrails-v0.1What this dataset tests
Given proposed next interventionsclassify stability in the response manifoldand add a guardrail that prevents known failure patterns.
Labels
stable_move
high_variance_move
risky_move
contraindicated_move
Typical failures
repeating tolerance loops
retrial after paradoxical worsening
allowing oscillation through exposure gaps
undertreating high-risk physiology
adding noise in flat nonresponse cases
Suggested prompt wrapper
System
You evaluate… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/clinical-tpib-pathway-stability-and-risk-guardrails-v0.1.guardrail_samples
Prem Studio Guardrail Datasets
This repo contains two closely related safety/guardrail datasets used in Prem Studio to train small safety models in the style of Llama Guard:
dataset_user_prompt_guardrail.jsonl→ Detect unsafe content in user messages.
dataset_system_response_guardrail.jsonl→ Detect unsafe content in agent/assistant messages (i.e. “did the model reply unsafely?”).
Both datasets follow the same pattern:
A system prompt that defines the task.
A user message that… See the full description on the dataset page: https://huggingface.co/datasets/prem-research/guardrail_samples.guardrail-training-dataset
Guardrail Training Dataset
Training dataset for fine-tuning language models to resist prompt injection attacks. Created for the Prompt Injection Challenge - an AI security challenge where users attempt to extract a hidden flag from a chatbot.
Dataset Description
Three main splits:
1. guardrail_v1 - Direct Attacks
Single-turn conversations with immediate prompt injection attempts and polite refusals. Examples include role-playing, system prompt overrides, and… See the full description on the dataset page: https://huggingface.co/datasets/Alindstroem89/guardrail-training-dataset.