Fine-Tuning Small Language Models for Urdu Customer-Support Automation

Abstract

We evaluate parameter-efficient fine-tuning (LoRA) of small open-weight language models on a curated Urdu customer-support corpus, achieving competitive intent accuracy at a fraction of the inference cost of hosted APIs.

Methodology

The study follows a reproducible pipeline with publicly documented datasets and open-source tooling. Full experimental details, code and evaluation notebooks are available on request from the authors.

Key Findings

• Practical, low-cost approaches can match or exceed commercial baselines in the local context.
• Reproducibility and documentation significantly improve adoption by practitioners.
• Results are integrated into the ITA Academia curriculum as case studies.