IndicNLP with Small Models

Experiments adapting compact multilingual models for Telugu and Indic-language tasks

Exploring how far small multilingual models can be pushed on Indic-language tasks with careful data design, parameter-efficient fine-tuning, and targeted evaluation.

Current focus:

  • Fine-tuning compact models such as Tiny Aya Fire with QLoRA
  • Building Telugu math and reasoning datasets from textbook-style material
  • Evaluating format-following, answer quality, and task transfer after small-model SFT
  • Studying when small models can replace larger hosted systems for focused Indic-language workflows

Tech: Python, Hugging Face Transformers, QLoRA, Tiny Aya, Telugu NLP, evaluation scripts