Source-linked AI summary
Gurukul AI: An Interactive AI-Driven Educational Platform for Indian Education System
Isha Narang, Sneh Gosai, Mayank Singh
TL;DR
Existing educational LLMs and QA resources are not well aligned with India's linguistically diverse, exam-oriented, textbook-based school context. GurukulAI addresses this gap with an NCERT-derived dataset, a fine-tuned LLaMA 3.1 model, and retrieval-augmented generation, with evaluations confirming educational alignment and usability and the system positioned as scalable, culturally grounded generative AI for India.
Problem
Existing LLMs and QA datasets are predominantly Western-centric, English-only, domain-generic, or not textbook-aligned, limiting support for India's linguistically diverse, exam-oriented school education.
Method
The system uses 18,720 NCERT-derived question–answer pairs for Classes 9–12, fine-tunes LLaMA 3.1, and combines it with retrieval-augmented generation for bilingual, syllabus-aligned QA.
Results
Quantitative evaluation on 700 test questions and qualitative feedback from 40 students confirm educational alignment and usability; the full RAG model achieves the highest scores.
Takeaways & Limitations
GurukulAI provides a public, NCERT-grounded web platform for chapter-specific answers, adaptive learning support, and MCQ and theory evaluation in English and Hindi.
Takeaways & Limitations
GurukulAI currently supports only Classes 9–12 in the science stream and aligns solely with NCERT textbooks, limiting applicability to ICSE and state curricula.
Abstract
from arXiv · showhide
Recent advances in large language models (LLMs) like ChatGPT and LLaMA have transformed AI-driven education, but these systems are predominantly trained on Western-centric data, making them ill-suited for regional curricula like India's. The Indian education system is linguistically diverse, exam-oriented, and structured around standardized syllabi, not addressed by existing datasets or tools. In this work, we curate a syllabus-aligned QA dataset based on NCERT (National Council of Educational Research and Training) textbooks for classes 9-12, capturing the content, context, and teaching style of Indian curricula. The final dataset, comprising 18,720 question-answer pairs across five subjects, is publicly available at https://huggingface.co/datasets/LingoIITGN/Gurukul. We fine-tune the LLaMA 3.1 8B model using this dataset and deploy it in a Retrieval-Augmented Generation (RAG) framework tailored to educational needs. We introduce GurukulAI, an open-access platform that enables Indian students to chat with the model, get doubts cleared, practice exam-style questions, receive contextual answers, and interact in both English and Hindi. By localizing AI for Indian classrooms, our work bridges the gap between global LLM capabilities and regional educational demands. The code is available at https://github.com/lingo-iitgn/GurukulAI.
1 Introduction
GurukulAI addresses the mismatch between Western-centric educational LLMs and India’s linguistically diverse, syllabus-driven school system. It contributes an NCERT-aligned dataset and an open bilingual platform combining generative tutoring, retrieval, and practice features.
- Western-trained LLMs are limited for India’s linguistically diverse, exam-driven educational contexts.
- Existing QA datasets are generally English-only and domain-generic, while Indian alternatives lack textbook alignment and open textbook-grounded generation.
- 18,720 bilingual Q&A pairs across five subjects and Classes 9–12 form GurukulAI’s NCERT-aligned dataset.The dataset includes chapter-specific metadata.
- The open platform combines generative tutoring, RAG-based inference, image-based question handling, and MCQ/theory modes.
- Quantitative evaluation on 700 test questions and qualitative feedback from 40 students assess educational alignment and usability.
2 Related Work
Related work demonstrates the educational potential of LLMs and multilingual benchmarks but leaves gaps in curriculum alignment, textbook grounding, and classroom-oriented QA. GurukulAI addresses these gaps with a chapter-tagged, NCERT-grounded generative system.
- Western-centric training limits LLM generalization and cultural relevance in India’s linguistically diverse, exam-oriented education system.
- Global platforms provide capabilities such as conversational feedback, emotional support, OCR, and retrieval, but lack NCERT alignment and often bilingual, textbook-grounded interaction.
- Indian initiatives primarily target engagement, evaluation, or regional-language learning rather than open-domain generative, curriculum-grounded question answering.
- Global QA benchmarks are largely domain-general, English-only, and not aligned with school-level syllabi.
- Indian multilingual benchmarks still lack explicit NCERT chapter alignment and subject-aware QA that mirrors classroom learning.
- GurukulAI fills the identified gap by combining fine-tuned generation with textbook retrieval for syllabus-aware tutoring in Classes 9–12.
3 Data Curation
The authors curate a bilingual, chapter-tagged NCERT QA dataset for Classes 9–12 from textbooks, exam papers, and educational websites. Automated refinement and manual review produce a text-based resource released publicly.
- 18,720 Q&A pairs across five subjects form a bilingual, chapter-annotated dataset for Classes 9–12.The subjects are Science, Mathematics, Social Science, English, and Hindi.
- Content was collected from English and Hindi NCERT PDFs, CBSE question papers, and educational websites.
- Each item was tagged by class, subject, and chapter, then filtered and refined using the Claude Sonnet API.
- The refinement process removed incomplete or visually dependent items and rewrote ambiguous or compound questions into self-contained forms.
- Questions were also completed and aligned with textbook tone before manual review for fluency, syllabus relevance, and coherence.
- The dataset is publicly available through Hugging Face.
4 GurukulAI: System Design and Implementation
GurukulAI uses a modular fine-tuned LLaMA model, NCERT retrieval pipeline, and bilingual web interface. The platform supports grounded interaction through chat, question handling, follow-up learning, MCQs, and theory practice.
- System architecture: The system comprises a fine-tuned generative model, RAG module, and user-facing web interface for grounded bilingual QA.
- Model adaptation: LLaMA 3.1 8B was fine-tuned with LoRA on the NCERT-aligned dataset for textbook-aware answering across subjects.
- Retrieval pipeline: NCERT chapters are chunked, embedded with BGE-m3, indexed in FAISS, and queried by retrieving the top five relevant passages.
- User interface: The web interface supports English and Hindi input, chat, MCQ solving, and theory-question practice.
- Retrieval pipeline: Inference combines the user query with retrieved NCERT passages in a context-rich prompt passed to the fine-tuned LLaMA model.
- Interactive features: Users can upload scanned or handwritten questions and receive three contextually related follow-up questions after interactions.
- Interactive features: The platform links answers to relevant NCERT chapters and provides MCQ scoring and free-form theory-response feedback.
5 Evaluation
GurukulAI was evaluated through automatic metrics, a student feedback study, and ablations of fine-tuning and retrieval. Results indicate strong curriculum-grounded performance and user acceptance, while harder questions and removing retrieval reduce performance.
- Automatic Evaluation: The evaluation used 700 held-out questions across Science, Social Science, English, and Hindi, with manually curated NCERT answers and easy, medium, and hard labels.
- Automatic Evaluation: ROUGE-L measured token overlap for factual coverage, while Semantic Similarity measured contextual alignment using Sentence-BERT cosine similarity.
- Automatic Evaluation: GurukulAI performed best on Social Science and Science, while English and Hindi achieved slightly lower ROUGE-L and semantic similarity scores.
- Automatic Evaluation: Performance remained relatively stable across Classes 9–12, with a minor dip in Class 11 associated with greater abstraction and complexity.
- Automatic Evaluation: Hard questions reduced performance, affecting ROUGE-L more sharply while semantic similarity remained relatively robust.
- Qualitative User Study: In a 43-participant study, users rated the system factually reliable at 86%, helpful at 84%, and easy to understand at 81%.Response time satisfaction was 88%, and 89% reported improved subject understanding.
- Qualitative User Study: Most participants preferred GurukulAI for syllabus coverage and NCERT grounding, although suggestions included simpler explanations, more consistent UI/UX, and interactive follow-up features.The platform was preferred over existing methods by 72% of users and considered less distracting than conventional online search by 77%.
- Ablation Study: Among four configurations, the full RAG model achieved the highest scores; removing retrieval reduced Semantic similarity by 13% and ROUGE-L by 6–8%.The configurations included pretrained, fine-tuned-only, full RAG, and noisy-query robustness settings.
6 Conclusion and Future Work
GurukulAI combines a curriculum-aligned dataset, a fine-tuned generative model, and retrieval-augmented inference for Indian school education. The authors present it as feasible scalable, culturally grounded generative AI for inclusive education in India, while planning broader coverage and multimodal improvements.
- Conclusion: GurukulAI integrates a fine-tuned generative model, retrieval-augmented inference, and syllabus-grounded QA data to provide chapter-specific answers and adaptive learning support.Its public interface supports MCQ and theory evaluation grounded in NCERT content.
- Conclusion: The platform is designed for Indian school education and provides a public web interface for NCERT-grounded learning interactions.
- Future Work: Future work targets coverage beyond Grades 9–12, regional board syllabi, multimodal inputs such as diagrams and equations, adaptive learning paths, and lightweight user modeling.
- Conclusion: The authors describe their evaluations as demonstrating the feasibility of scalable, culturally grounded generative AI for inclusive education in India.
7 Limitations
GurukulAI remains limited in curriculum scope, reasoning ability, multimodal handling, and reliability. These boundaries motivate broader curriculum support, improved reasoning, multimodal input handling, and more interpretable behavior.
- Scope: GurukulAI supports only Classes 9–12 in the science stream and aligns solely with NCERT textbooks, limiting applicability to ICSE and state curricula.
- Reasoning: The model struggles with high-cognitive-load questions requiring multi-step reasoning, particularly in Mathematics and Physics.
- Multimodality: GurukulAI lacks multimodal capabilities and cannot process visual inputs such as diagrams or charts.
- Reliability: Despite contextual retrieval, the black-box generative model may produce factually incorrect or misleading answers, so users are advised to cross-verify critical responses.
- Future Directions: The identified future-work directions are broader curriculum support, improved reasoning, multimodal input handling, and more interpretable model behavior.
8 Ethics
The authors report ethical handling of publicly available educational sources and state that no personal or user data was collected for training or evaluation. Dataset filtering and manual review were used for quality, but residual bias and language-access limits remain.
- Data Practices: The QA dataset was curated from publicly available NCERT textbooks and educational websites under fair-use for non-commercial academic purposes.
- Data Practices: No personal or user data was collected during training or evaluation.
- Quality Control: Dataset entries were filtered with the Claude Sonnet API and manually reviewed to ensure quality.
- Caveats: The authors acknowledge that some biases may persist because of automated processing.
- Caveats: Current support for only English and Hindi may limit accessibility for users of other languages.
- Use Guidance: GurukulAI is intended as a supplementary learning tool rather than a substitute for formal education, and users are advised to verify critical responses.