Alif: Advancing Urdu Large Language Models via Multilingual Synthetic Data Distillation
Paper • 2510.09051 • Published
output = llm(
"Once upon a time,",
max_tokens=512,
echo=True
)
print(output)Alif 1.0 3B Instruct
Alif 1.0 3B Instruct is an open-source instruction-tuned language model developed by traversaal.ai, focused on Urdu and English understanding and reasoning. Fine-tuned with high-quality, culturally aware synthetic data, Alif delivers strong multilingual performance with a particular emphasis on low-resource language alignment and nuanced instruction following.
Key Highlights
- Multilingual: Optimized for Urdu (primary) and English (secondary)
- Instruction-tuned: Trained on synthetic QA, chat, and reasoning tasks
- Lightweight: Only 3 billion parameters — fast, memory-efficient
- 4-bit Quantized Version: Available for on-device inference
- Training Framework: Fine-tuned 2× faster using Unsloth + TRL
Model Metadata
- Developer: large-traversaal
- License: Apache 2.0
- Base Model:
unsloth/Llama-3.2-3B - Model ID:
Alif-1.0-3B-Instruct - Parameters: 3 billion
Use Cases
Alif is ideal for:
- Urdu and English chatbots
- Question answering in low-resource languages
- Translation, summarization, and creative writing
- Running on edge devices with the 4-bit version
Citation
@article{ShafiqueAlif2025,
title = {Alif: Advancing Urdu Large Language Models via Multilingual Synthetic Data Distillation},
author = {Muhammad Ali Shafique and Kanwal Mehreen and Muhammad Arham and Maaz Amjad and Sabur Butt and Hamza Farooq},
journal = {arXiv preprint arXiv:2510.09051},
year = {2025},
url = {https://arxiv.org/abs/2510.09051}
}
- Downloads last month
- 51
Hardware compatibility
Log In to add your hardware
4-bit
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="large-traversaal/Alif-1.0-3B-Instruct", filename="Alif-1.0-3B-Instruct-Q4_K_M.gguf", )