Researchers curate Indian curriculum-aligned QA dataset and fine-tune LLaMA 3.1 8B for localized education

Researchers from IIT Gandhinagar and collaborators introduce GurukulAI, an interactive AI-driven educational platform designed for India’s education system. The team argues that mainstream LLMs like ChatGPT and LLaMA are trained on Western-centric data and are poorly suited to India’s linguistically diverse, exam-oriented, and syllabus-structured curriculum.

To bridge this gap, the authors curate a syllabus-aligned QA dataset based on NCERT textbooks for classes 9–12, covering five subjects. The dataset contains 18,720 question-answer pairs and is released publicly on Hugging Face.

The researchers fine-tune the LLaMA 3.1 8B model using this dataset and deploy it within a Retrieval-Augmented Generation (RAG) framework tailored to educational needs. The resulting system enables students to chat with the model, clarify doubts, practice exam-style questions, and receive contextual answers in both English and Hindi.

The authors release the dataset, fine-tuned model, and platform code as open-access resources, including the dataset on Hugging Face and the code on GitHub. They position the work as a localization effort that aligns global LLM capabilities with regional educational demands.

The preprint is dated July 18, 2026, and is available on arXiv under the Computation and Language category.