Pakistani developer, Junaid Ahmed, has created Pakistan’s first ever Pashto language model, Qehwa AI.
The model is specifically tailored for the Peshawari dialect of Pashto and seeks to overcome the shortcomings observed in current global AI systems when processing the Pashto language and its cultural nuances.
The model underwent evaluation through a custom benchmark comprising 150 tests across 15 distinct categories, achieving an overall accuracy rate of 85.3%.
This positions it as the first AI model explicitly designed for Pashto, complete with structured evaluation benchmarks. Qehwa was developed in two primary phases to enhance both language comprehension and task execution. Qwen2.5-7B serves as the foundational model for Qehwa.
Qwen represents a series of highly proficient, open-source large language models created by Alibaba Cloud. The designation “7B” indicates that it contains 7 billion parameters (the “brain cells” of the model).
The developers used this base model, which already possessed a profound understanding of general logic, coding, and multiple languages, and specifically trained it to excel in the Peshawari Pashto dialect.
During the initial phase, the model underwent extensive pre-training utilising 3.4 million documents in Pakistani Pashto. This stage concentrated on enhancing vocabulary, grammar, and cultural comprehension.
Qehwa AI responds in only Pashto
The model accommodates prompts in Pashto, English, and Urdu, while generating responses solely in Pashto. Additionally, it introduces the first dedicated benchmark for Pashto LLMs, featuring 150 evaluation tests.
The accuracy of English to Pashto translation reached 90%, while Urdu to Pashto translation achieved an accuracy of 84%. In specialised subject areas, the model scored 90% in culture and history, health and daily life, as well as geography and nature.
The cumulative score across all 15 categories is recorded at 85.3%.
Qehwa AI can be installed and run using the open-source tool, Unsloth or through a compression method using BitsAndBytes.
Read more: Pakistan launches first AI avatar ‘Laila’
