Start with the language the tools overlook.

My role

In a two-person project, I owned model fine-tuning and active learning. Waleed Arshad owned data collection, augmentation and dataset creation.

Where it stands

Completed academic research project. Results depend on the corpus, augmentation and script conditions.

What changed.

An experiment in where to spend a limited annotation budget, with explicit comparisons between data conditions.

My part

I worked on the model and the labeling strategy for a four-topic classification task in Shahmukhi Punjabi. The team contributed a labeled dataset and a larger transliterated corpus; I do not claim the dataset work as mine alone.

What the headline leaves out

The full augmented pipeline reached 92.9% accuracy. A controlled comparison under different script conditions produced substantially different results. I keep the experimental setup attached to the number rather than treating it as the accuracy of an entire language.

The useful question

Active learning asks which examples are worth labeling next. I used maximum-entropy uncertainty sampling to concentrate the annotation effort on examples the model found difficult.

The stack

• PyTorch
• XLM-RoBERTa
• Hugging Face
• Active learning