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 adapted before inclusion.

Dataset at a glance

Rows	~405K
Language (released text)	Turkish
Task	Prompt safety classification (binary + multi-label)
Split	train / validation / test ≈ 80% / 10% / 10%
License	apache-2.0

Dataset structure

Column	Type	Description
prompt	string	The input text (Turkish).
safety	string	"safe" or "unsafe".
category	list[string]	Zero or more of the 10 hazard labels below. Multi-label is allowed (e.g. hate and harassment together). Safe prompts typically have an empty list.
language	string	"tr" or "en" — source-language origin of the row (native Turkish corpus vs. English-origin row that was translated/adapted). Released prompt text is Turkish in either case.
source	string	Which upstream dataset the row came from.

Hazard taxonomy

The taxonomy is inspired by the MLCommons hazards taxonomy, adapted for this release.

Label	Description
VIOLENT_CRIMES	Physical violence, weapons, and controlled substances: threats, incitement, acquiring/using weapons, animal abuse, terrorism, and drug-related crime.
NON_VIOLENT_CRIMES	Non-violent illegal or unethical activity: fraud, theft, financial crime, cybercrime, and general “help me do something illegal/unethical” requests without physical harm.
HATE_DISCRIMINATION	Content targeting people based on protected characteristics with hateful, stereotyping, or discriminatory language, including identity-based threats.
HARASSMENT_OFFENSIVE	Generic toxicity, profanity, insults, and harassment that is not specifically identity-based.
SEXUAL_CONTENT_ADULT	Sexual or adult content involving consenting adults, including explicit descriptions and erotica requests.
CSAE	Child sexual abuse/exploitation material or requests. Highest-severity category.
SELF_HARM_SUICIDE	Suicidal ideation, self-harm intent, or requests for self-harm/suicide methods.
INJECTION_JAILBREAK	Prompt injection and jailbreak attempts: overriding instructions, extracting system prompts, or bypassing safety via adversarial framings.
MISINFORMATION_POLITICAL	Disinformation, conspiracy content, political manipulation, and requests to fabricate or spread misleading information.
PRIVACY_VIOLATION	Requests involving private/personal information: PII exposure, doxxing, surveillance, or other privacy-invasive content.

Benign prompts are labeled safety = "safe" (including hard negatives that look risky but are not harmful).

Label counts

Individual tag appearances (multi-label rows contribute to more than one category; SAFE is counted separately):

Category	Count
SAFE	206,556
HARASSMENT_OFFENSIVE	81,802
HATE_DISCRIMINATION	45,050
SEXUAL_CONTENT_ADULT	26,612
VIOLENT_CRIMES	24,922
INJECTION_JAILBREAK	24,604
NON_VIOLENT_CRIMES	18,242
SELF_HARM_SUICIDE	17,063
PRIVACY_VIOLATION	10,437
MISINFORMATION_POLITICAL	7,480
CSAE	7,318

Source datasets

Turkish sources

Source	Notes
Overfit-GM/turkish-toxic-language	5-way Turkish toxicity labels (OTHER / PROFANITY / INSULT / RACIST / SEXIST)
Toraman et al. (hate speech dataset)	Turkish tweets labeled Normal / Offensive / Hate
coltekin/offenseval2020_tr	Turkish labeled Normal / Offensive

Turkish sources were deduplicated and passed through heuristic quality filters.

English sources

Source	Notes
lmsys/toxic-chat	Real user↔AI conversations with toxicity + jailbreaking flags
neuralchemy/Prompt-injection-dataset	Prompt injection attempts across many attack categories
qualifire/prompt-injections-benchmark	Benign vs. jailbreak benchmark prompts
OR-Bench (Over-Refusal Benchmark)	Benign prompts that look harmful (SAFE hard negatives)
PKU-Alignment/BeaverTails	QA prompts with multi-label harm annotations
allenai/WildGuardMix	Harmful/benign prompts with fine-grained subcategories
nvidia/Aegis-AI-Content-Safety-Dataset-2.0	Prompts with multi-label violated-category annotations
lmsys/lmsys-chat-1m	In-the-wild conversations with OpenAI Moderation signals
OnAY self-harm dataset (self_harm_12.5K_samples)	Self-harm-related prompts
xTRam1/safe-guard-prompt-injection	Prompt-injection prompts (label=1 rows only)
allenai/WildChat-4.8M-Full	In-the-wild conversations with OpenAI Moderation signals

From chat-style English corpora, only single-turn prompts were used. Undersampled / rare classes were additionally sampled from moderation-flagged rows in LMSYS and WildChat.

Curation pipeline

English-origin and Turkish-origin streams were processed separately, then merged.

Turkish stream

1. Deduplication
2. Heuristic quality filtering

English stream

1. Preprocessing, filtering, and deduplication (MinHash)
2. Unredacting (~28K rows): redacted spans in English sources were filled from a pool of Turkish given names, place names, and related entities (English-origin rows only)
3. Resampling underrepresented classes from LMSYS / WildChat (single-turn, moderation-flagged)
4. Translation with google/translategemma-27b-it
5. Cultural adaptation (~26K rows) with google/gemma-4-31B-it, in a CultureGuard-style process: detect culturally grounded content, then adapt it

Merge, labeling, and synthesis

1. Merge streams → ~401K rows
2. LLM-as-judge labeling (multi-label, 10 categories) with a three-model jury:
   • google/gemma-4-26B-A4B-it
   • Qwen/Qwen3.6-35B-A3B
   • openai/gpt-oss-120b
3. Label acceptance if either:
   • at least 2 of 3 judges agree on the label, or
   • at least 1 judge agrees with the original source label
4. Synthetic augmentation (~4K rows) for underrepresented classes, generated with google/gemma-4-31B-it
5. Final dataset → ~405K rows, split 80 / 10 / 10 (train / validation / test)

Intended use

• Training and evaluating Turkish prompt safety / guardrail classifiers
• Research on multi-label hazard taxonomies, over-refusal (hard negatives), injection/jailbreak detection, and related moderation topics

Out-of-scope use

• Training dialogue agents to produce harmful content
• Treating labels as legal, clinical, or policy judgments without human review
• Assuming coverage of assistant responses as this release is prompt-oriented

Bias, risks, and limitations

• Harmful content: The dataset contains offensive, violent, sexual, self-harm-related, and otherwise distressing text. Engage only according to your own risk tolerance and institutional guidelines.
• Synthetic and translated text: Large portions are machine-translated and/or culturally adapted; judge labels are model-based. Residual translation artifacts, cultural mismatches, and judge errors are possible.
• Taxonomy & policy dependence: Category boundaries are design choices. Other policies may disagree on borderline cases (especially harassment vs. hate, or political misinformation).
• Class imbalance: SAFE and a few unsafe categories dominate; rarer tags (e.g. CSAE, MISINFORMATION_POLITICAL) remain smaller even after synthetic augmentation.
• Source bias: Upstream English and Turkish corpora reflect their collection channels (social media, chat logs, red-teaming sets) and may under- or over-represent certain dialects, demographics, or harm types.

The views expressed in the prompts do not reflect the views of YTU CE COSMOS or the dataset curators.

License

apache-2.0