ChatRoman:Regional languages to Roman Script is a fine-tuned version of Google’s open-weight Gemma 4 model that converts English sentences into casual romanized Kannada and Hindi, the way people type in chat messages. Many Indian-language speakers write in roman letters on messaging apps, but general-purpose language models often answer with extra commentary, such as lists of options and explanations, instead of the single natural reply the user needs.

The project addresses this with parameter-efficient fine-tuning. Using Unsloth and LoRA adapters, the model learns the output style from example pairs drawn from the AI4Bharat IndicCMix dataset, and it trains only about 0.25% of its weights. This makes training feasible on a free Colab T4 GPU. The bilingual model is trained on 3,000 sentence pairs, 1,500 in Kannada and 1,500 in Hindi, and is evaluated on 200 held-out test sentences, 100 per language.

Performance is measured with chrF, a character-level similarity score that suits romanized text, where spellings vary. In a Kannada pilot on 30 test sentences, the plain model scored 33.1 chrF and the fine-tuned model scored 57.1, a gain of 24.0 points. The bilingual results will be reported alongside it.

The project includes an interactive Gradio app where users enter an English sentence, choose a language, and compare the plain and fine-tuned outputs side by side.

Limitations: The test set is small, chrF can undercount correct romanizations that use different spellings, and the model still keeps some English words in its output. The current work uses a subset of the full dataset and a single base model size.