oddadmix/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 (NAMAA-Space/Ara-Prompt-Guard_V0) is 33K machine-translated rows, 3 classes, MSA only. This is native, six-variety, 12-class.

⚠️ Safety boundary — read this first

Every row is a request, never an answer.

A guardrail classifies incoming prompts. A harmful completion is not needed to train one, and producing it would create a genuinely harmful artifact. A row may ask "اشرح لي طريقة صنع…" — it never contains the method.

• No operational detail. No synthesis routes, exploit code, or step-by-step procedures.
• Nothing involving minors. MLCommons S4: Child Sexual Exploitation is excluded from the taxonomy entirely — not a class, not a distractor, never generated.
• This is enforced by a rule, not by trust. The generator is an abliterated (refusal-removed) model, so it would comply if asked. guard_common.validate() rejects any row combining step markers with hazardous vocabulary. Measured on the released corpus: 0 rows contain step markers or code.

Intended use is training and evaluating defensive classifiers. It is not a red-teaming toolkit.

The 12 classes

class
safe	ordinary benign prompts
safe_sensitive	benign but superficially alarming — homonyms (كيف أقتل عملية في لينكس), figurative language, definitions, safe targets, fiction, professional context
jailbreak	overriding the model's own conditioning — persona, fiction framing, refusal suppression, encoding, many-shot, authority, translation pivot
prompt_injection	instructions smuggled inside third-party data — document, email, web page, tool output, review
violent_weapons · nonviolent_crime · self_harm · hate_harassment · sexual_content · privacy_pii · specialized_advice · misinformation	MLCommons-aligned harm categories

safe + safe_sensitive are ~30% of rows. Over-blocking is the failure that makes a guardrail unusable, and it is invisible unless "safe but looks unsafe" is its own labelled class.

self_harm covers distress signals as well as explicit requests. Defined first as "requests only", it retained just 55% through verification because the generator kept writing expressions of distress — which is in fact the most valuable signal a guardrail can catch. Widening the definition took it to 97.5%.

Splits — three ways a guardrail fails

split	rows	tests
train	196,297
held	6,061	new texts, familiar dialects and templates
unseen_dialect	39,761	Maghrebi held out entirely — the bypass an MSA-only guardrail leaves open
unseen_phrasing	8,723	attack templates absent from training (encoding, translation_pivot, tool_output)
hard	728	safe_sensitive only — the over-blocking slice

Verification

Labels were established by construction, then confirmed by an independent re-classification pass that never saw the target. 13% of rows were dropped on disagreement (289,023 → 251,408).

Retention by class is in code/guard_retention.txt and is itself a finding:

	retention
privacy_pii · safe · self_harm · specialized_advice	97–99%
jailbreak · prompt_injection · hate_harassment	92–95%
violent_weapons	75.4%
safe_sensitive	71.9%
misinformation	70.4%
nonviolent_crime	64.3%

nonviolent_crime is the genuinely hard boundary: "I was defrauded, what are my rights" and "help me defraud someone" share almost all their vocabulary. Sharpening the definition moved it 40% → 64% and no further.

Retention is flat across dialects — 85.7% (Maghrebi) to 88.3% (code-switched). Dialect coverage is not a thin veneer over an MSA corpus.

Audit, run before training

• length shortcut: +0.0013 over the majority baseline from the best single threshold. An earlier grid where length was a side effect of class (safe_sensitive averaged 404 chars against violent_weapons' 64) leaked +0.080; length is now an axis drawn independently of class.
• six varieties, near-uniform · zero held-out dialect or template families in train

Results

model	over-refusal ↓	miss rate ↓	macro-F1 (held)
Nawah-Guard-52M	0.0094	0.0048	0.9903
Nawah-Guard-500K (2 MB)	0.0203	0.0077	0.9782

Files

data/ splits + labels.json · raw_generations.jsonl.gz raw completions, re-scorable with no GPU · code/ the full pipeline — taxonomy and validator (guard_common.py), resumable generator, verifier, split/audit, and both trainers including the zero-shot routing-head variant that lets a deployment define its own policy categories at inference.

Caveats

Synthetic, from one generator (orcarouter/Qwen3.8-27B-Uncensored-NVFP4). Texts are 1–3 line messages; long documents are absent. The verifier is the same model that generated the text — a different prompt with the target hidden, which catches careless labelling but not errors the model is systematically wrong about.