A combined, cleaned and deduplicated dataset for training lightweight detectors of jailbreak and prompt-injection attempts against LLMs, built from 13 public datasets. All credit goes to their authors (see below).

⚠️ Contains real adversarial prompts, some of which include harmful requests. Intended for building defensive classifiers, guardrails and red-team evaluation. Access is gated. Model responses from source datasets were deliberately not included.

Labels

label	id	meaning
`benign`	0	Normal prompts, including hard negatives: role-play ("act as…"), long system prompts, adversarially-phrased but harmless requests
`jailbreak`	1	Attempts to make the model ignore its safety rules (DAN, persona tricks, hypothetical framing, obfuscation…)
`injection`	2	Attempts to override the developer's or system instructions ("ignore previous instructions and…")

Direct harmful requests without any trick are labelled benign for attack type: that's a content-moderation problem, not a jailbreak. Where the source provides it, the harmful column flags them, so you can train both tasks.

Stats

156,802 unique prompts: benign 90,175 · jailbreak 1,766 · injection 64,861

split	rows	benign	jailbreak	injection
train	116,854	66,024	1,019	49,811
validation	14,335	8,168	111	6,056
test	14,592	8,266	134	6,192

How it was built

Loaded 13 source datasets and mapped their labels to the 3-class scheme above.
Dropped all model-response columns and kept prompts only.
Normalized whitespace and removed texts under 8 or over 20,000 characters.
Exact deduplication across all sources on normalized text; all_sources lists every dataset a prompt appeared in.
Resolved label conflicts by majority vote; ties dropped.
Split by template family (hash of the normalized first 200 characters), so variants of the same jailbreak template (e.g. DAN versions) never leak between train and test.

Columns

id, text, label, label_id, is_attack, harmful (bool or null), source_dataset, all_sources, source_config, source_split, source_license, source_url, n_chars, template_group

from datasets import load_dataset
ds = load_dataset("nuhmanpk/prompt-sentinel")

Credits & sources

This dataset is a compilation. All prompts come from the datasets below; please cite and credit the original authors.

Dataset	Author	License	Rows used	Papers	Contribution
TrustAIRLab/in-the-wild-jailbreak-prompts	TrustAIRLab	mit	14,040	train	arXiv:2308.03825
lmsys/toxic-chat	lmsys	cc-by-nc-4.0	9,516	train	arXiv:2310.17389
jackhhao/jailbreak-classification	jackhhao	apache-2.0	636	train	—
reshabhs/SPML_Chatbot_Prompt_Injection	reshabhs	mit	15,910	train	arXiv:2402.11755
jayavibhav/prompt-injection	jayavibhav	unspecified	79,673	train	—
S-Labs/prompt-injection-dataset	S-Labs	mit	15,095	train	—
neuralchemy/Prompt-injection-dataset	neuralchemy	apache-2.0	5,925	train	—
Lakera/gandalf_ignore_instructions	Lakera	mit	996	train	arXiv:2501.07927
fka/awesome-chatgpt-prompts	fka	cc0-1.0	1,970	train	—
OpenAssistant/oasst2	OpenAssistant	apache-2.0	5,255	train	arXiv:2304.07327
leolee99/NotInject	leolee99	mit	255	benchmark	arXiv:2410.22770
deepset/prompt-injections	deepset	apache-2.0	503	benchmark	—
xTRam1/safe-guard-prompt-injection	xTRam1	unspecified	7,028	benchmark	arXiv:2402.13064

Licenses were read from each dataset's Hub metadata. Some sources are non-commercial or share-alike, and each row's source_license tells you which terms apply. Filter by it if you need a commercially usable subset.

Sources attempted but not included in this build: allenai/wildjailbreak, allenai/wildguardmix, hackaprompt/hackaprompt-dataset, qualifire/prompt-injections-benchmark, allenai/wildguardmix

Limitations

Label schemes differ between sources; mapping to 3 classes involves judgment calls (documented in the build notebook).
Mostly English (some German via deepset/prompt-injections).
Near-duplicate variants beyond the template-family grouping may remain.
Jailbreak techniques evolve fast, so detectors trained on this will miss novel attacks. Refresh periodically.

Author

Compiled by Nuhman PK: 🤗 Hugging Face · 💻 GitHub · 📊 Kaggle · 💼 LinkedIn · 🐦 X · ✍️ Medium

v2

load_dataset("nuhmanpk/prompt-sentinel", "v2") adds, on top of the original rows: obfuscated attacks (base64, ROT13, leetspeak, homoglyphs, spacing, reversal, zero-width, Unicode-tag smuggling) with benign encoded controls, trigger-word benign prompts against over-defense, and indirect injections planted in documents (context column) with clean twins. augmentation names the transform. Training rows that collide with public benchmarks (deepset, xTRam1, jackhhao, NotInject, WildJailbreak eval, Qualifire) were removed. Counts: train 80,949 · validation 14,335 · test 18,592. Used to train Gatekeeper.