Qwen 2.5 7B Urdu (v3 LoRA)

Fine-tuned Qwen 2.5 7B Instruct for Urdu (Urdu script + Roman Urdu + code-mixed). 79.5% win rate vs base Qwen across 2 LLM judges on a 100-prompt hand-curated set.

This Space is a Gradio frontend; the model runs on a private GPU endpoint (Modal H100). First call may take 30-90s while the container warms up. Subsequent calls are 2-8s.

Model card · https://huggingface.co/TayyabManan/qwen2.5-7b-urdu-v3 | Code & data · https://github.com/TayyabManan/Urdu-LLM

Sawal (سوال) — Urdu, Roman Urdu, or English

Temperature

Lower = more factual / shorter

0 1.5

Max new tokens

64 1024

(awaiting query)

Try these (factual QA, creative writing, code explanation, translation)

Sawal (سوال) — Urdu, Roman Urdu, or English	Temperature	Max new tokens

Built in public by Muhammad Tayyab. Apache-2.0. Code, eval data, and training pipeline at https://github.com/TayyabManan/Urdu-LLM.