The Gemma Sutras

Fine-tuning Gemma 3 for Sanskrit Sandhi Splitting

Overview

The Gemma Sutras is a research project focused on fine-tuning Google’s Gemma 3 language model for the specialized task of Sanskrit Sandhi Splitting. This work addresses one of the fundamental challenges in Sanskrit Natural Language Processing.

Background

What is Sandhi?

In Sanskrit, Sandhi (संधि) refers to the phonetic changes that occur when words are combined. When speaking or writing Sanskrit, adjacent sounds undergo transformations according to complex linguistic rules. For example:

  • रामः + अयम् → रामोऽयम्
  • सः + आगच्छति → स आगच्छति

The Challenge

Sandhi Splitting is the reverse process - taking a combined form and breaking it back into its constituent words. This is essential for:

  • Understanding Sanskrit texts
  • Building Sanskrit NLP systems
  • Automated translation and analysis
  • Making ancient texts more accessible

Technical Approach

Model Selection

  • Base Model: Gemma 3 by Google
  • Fine-tuning Strategy: Specialized training on Sanskrit Sandhi patterns
  • Dataset: Curated corpus of Sanskrit texts with annotated Sandhi splits

Innovation

  • Adapting modern large language models for classical language tasks
  • Handling complex linguistic rules through neural approaches
  • Creating a practical tool for Sanskrit scholars and researchers

Research Impact

Academic Recognition

  • ICML 2025 NewInML Workshop (Non-Archival)
  • EMNLP 2025 WiNLP Workshop (Archival)

Real-World Applications

  • Digital Humanities: Enabling computational analysis of Sanskrit literature
  • Education: Helping students learn Sanskrit grammar
  • Text Processing: Automated processing of historical texts
  • Preservation: Making ancient knowledge more accessible

Collaboration

Authors: Samarth P and Sanjay Mahalingam

This project demonstrates the power of combining modern AI techniques with traditional linguistic knowledge, creating tools that bridge ancient wisdom with contemporary technology.

Broader Significance

The Gemma Sutras project contributes to:

  • Low-resource Language NLP: Techniques applicable to other classical languages
  • Cultural Preservation: Using AI to maintain and promote linguistic heritage
  • Interdisciplinary Research: Uniting computational linguistics and Sanskrit studies

By making Sanskrit texts more computationally accessible, this work helps preserve and propagate one of humanity’s oldest and most sophisticated linguistic traditions.

References