Industry · Content moderation, trust & safety, LLM safety
Safety data for the abuse your classifier cannot read
Sectarian slurs in Roman Urdu, coded harassment in Punjabi, threats in Pashto that pass every English-trained filter. Platforms with South Asian audiences moderate blind. We produce severity-graded lexicons, labelled examples, adversarial prompts and safety preference data written and rated by native speakers who understand the context.
Who this is for
For platforms and model builders with users in Pakistan and its diaspora
Pakistan has more than 110 million social-media users; their languages are almost absent from moderation training sets.
- Social platforms and messaging apps — toxicity, harassment and hate-speech classifiers that work in Urdu script, Roman Urdu, Punjabi, Pashto and Sindhi
- LLM safety and alignment teams — refusal training, red-team prompts and harm taxonomies grounded in local context
- Gaming, dating and marketplace platforms — chat moderation for South Asian user bases
- Trust & safety vendors — labelled data and native rater capacity for multilingual moderation products
- Research and policy groups — annotated corpora of sectarian and political abuse for measurement
What breaks
Why English-trained moderation misses South Asian abuse
The failures are lexical, orthographic and cultural.
Coded and sectarian language
Much of the most harmful speech is coded: sect, caste-adjacent and ethnic slurs that look innocuous out of context. Only native raters who know the code can label it.
Roman Urdu evasion
Abuse is typed in Roman Urdu with deliberate misspellings and mixed script to defeat filters. Lexicons need spelling variants and the normalisation rules to match them.
Severity is contextual
The same word is banter between friends and a threat between strangers. Severity grading with three raters and reported agreement is what makes the labels usable for thresholds.
Translated taxonomies miss categories
Harm taxonomies written for English miss honour-based threats, blasphemy accusations and sectarian incitement — categories that carry real-world violence risk in Pakistan.
What we supply
Lexicons, labelled examples, adversarial prompts and preference data
Rated by three calibrated native speakers per item, with agreement reported and a taxonomy adapted to the region.
Severity-graded lexicon
Slurs, hate speech, harassment, sectarian and coded terms in Urdu, Roman Urdu, Punjabi, Pashto and Sindhi, each with spelling variants, severity grade and context notes. See the Safety & Toxicity Lexicon.
Labelled examples
Real-register posts, comments and messages authored or curated to spec and labelled against your taxonomy, balanced across categories and severity, with hard negatives.
Red-team and refusal data
Adversarial prompts, jailbreak attempts and refusal examples written natively for LLM safety tuning; safety preference pairs ranked by calibrated raters. See RLHF Preference Pairs.
Moderation on your content
Native-speaker labelling of your own content stream against your policy, with agreement reporting, remotely or in your VPC. See Annotation & Evaluation.
Ready-made
Catalog datasets for trust and safety
Each ships with a datasheet, per-record consent artefacts and a free unwatermarked sample pack.
Safety & Toxicity Lexicon
Severity-graded, four languages
TextRLHF Preference Pairs
Safety preference rankings
TextNative Urdu Instruction & SFT Corpus
Includes refusal and safety examples
AnnotationAnnotation & evaluation
Labelling and red-teaming on your data
TextUrdu–English Parallel Corpus
For cross-lingual moderation
TrustConsent & fair pay
How raters are protected and paid
How a programme runs
From taxonomy to labelled corpus
Rater welfare is part of the process: exposure limits, rotation and support are built in.
Adapt the taxonomy
Your policy categories mapped to local harm types with native reviewers; you approve the final taxonomy.
Calibrate raters
Gold items and agreement thresholds; three raters per item in production.
Label and grade
Weekly batches with per-category counts, severity distribution and agreement statistics.
Adversarial refresh
Periodic updates as coded language evolves, because it does.
Answers
Questions trust and safety teams ask
Urdu in Nastaliq script and Roman Urdu, Punjabi in Shahmukhi and Roman, Pashto, and Sindhi, with spelling variants and mixed-script forms included because that is how abuse is actually typed.
Three calibrated native raters grade each item on a defined scale with context notes; disagreements are adjudicated and agreement is reported per category so you can set thresholds with known confidence.
Yes, against your policy, with rater exposure limits and welfare support, remotely or inside your perimeter, with weekly agreement reporting.
Yes — adversarial prompts and jailbreak attempts written natively, refusal examples and safety preference pairs, grounded in local harm categories such as sectarian incitement and honour-based threats.
Yes — most of our licensing goes to AI teams outside Pakistan. English-language MSA and DPA, USD invoicing by default, W-8BEN-E for US payers, delivery to your AWS, GCP or Azure region. See Where we deliver for compliance mapping by region.
What is your classifier missing in Urdu?
Send examples of what got through. We come back with a taxonomy mapping, a graded sample and the agreement numbers.
Average first response: under 6 business hours. NDAs signed same day.