In development
Balagh بلاغ
Balagh is a text model we're fine-tuning to read and write Urdu, Punjabi, Sindhi and Pashto. The name comes from the Urdu word for conveying a message.
Why we're building it
Most language models learn from text gathered online, and far more of that text is in English than in the languages people speak across Pakistan. The result is familiar to anyone who has tried: Urdu answers that read like a translation, and Punjabi, Sindhi or Pashto that comes out shaky or switches back to English halfway through.
We want a model that a shopkeeper in Hyderabad or a clinic in Peshawar can use in the language they actually work in.
The four languages
Urdu
اردوPakistan's national language, written in the Perso-Arabic script.
Punjabi
پنجابیThe most widely spoken language in Pakistan, written here in Shahmukhi.
Sindhi
سنڌيWritten in its own Arabic-based alphabet, with letters Urdu doesn't use.
Pashto
پښتوSpoken across Khyber Pakhtunkhwa and Balochistan, with letters of its own.
How we're making it
We aren't training a model from scratch. We're taking an existing model that already understands language well and fine-tuning it on text in these four languages, so it learns their vocabulary, grammar and scripts properly.
Balagh works with text: reading, writing and answering in these languages. The things we're aiming for are answering questions, summarising documents, and translating between each of them and English.
Where it stands
Balagh is still being trained and isn't available to use yet. We'll update this page as the work moves along.