Author(s): Devavrat Samak

Originally published on Towards AI.

TL;DR: In this article, I share how I fine-tuned Google’s Gemma-3 on Peshwa-era Marathi texts using Unsloth and Lightning AI. You’ll learn about continual pretraining, low-cost LLM training workflows, and how this approach can inspire similar domain-specific models for heritage, culture, and education.

Fig1: Mighty Shaniwarwada (Image credit: Jam Travels/Shutterstock)

Introduction

It began with a question: Can a modern language model truly understand the elegance and complexity of historical texts? Not just contemporary dialects or everyday language, but the rich, layered prose from a bygone era — filled with political nuance, poetic rhythm, and cultural depth.

For me, this question struck close to home. I was born and raised in Pune, the seat of the Peshwas — the de facto leaders of the Maratha Empire in 18th-century India. One day, I was asked a seemingly simple question about them. Embarrassingly, I couldn’t answer it. Despite growing up amid this history, I realized I had very little connection to it. That moment stayed with me.

In today’s world, we consume history through screens, apps, and search engines. So, I thought: why not create a language model that could live and breathe that era’s knowledge? Something that wouldn’t just translate or summarize but genuinely understand the linguistic… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI