IKS-Instruct: A 24,000-Example Multilingual Dataset for Teaching Language Models Indian Knowledge Systems
IKS-Instruct creates a 24,795-pair multilingual dataset covering 41 Indian Knowledge System techniques, significantly enhancing model cultural depth and pedagogical ability.
Key Findings
Methodology
The dataset integrates six source types: classical texts, curriculum templates, Vedic mathematical sutras, bilingual instruction pairs, technique-based multi-turn dialogues, and cross-tradition analyses. Data generation employs three strategies: hybrid teacher models, mechanism-first approaches, and task abstraction, combined with multi-stage quality filtering—deduplication via MinHash, language/source verification, teacher fidelity scoring, and multi-judge evaluation. Data is structured in JSONL format with metadata on source, technique, subject, grade, tradition, and cognitive load. Fine-tuning uses a 7B Transformer model, evaluated through a multi-dimensional scoring system assessing fidelity, pedagogical quality, factual accuracy, and cultural depth.
Key Results
- The final dataset contains 24,795 pairs across seven languages and 41 techniques. The fine-tuned 7B model achieves a median judge score of 6.39, close to the 6.54 score of the strong general model Nemotron-Nano, vastly outperforming the base model's near-zero score on IKS-specific metrics.
- Model performance improvements are evident across cultural, technical, and pedagogical dimensions, validating the effectiveness of targeted instruction data. Data quality enhancements do not always linearly translate to performance gains, highlighting the importance of diversity and relevance.
- Coverage of multiple languages and techniques enables the model to perform cross-cultural, interdisciplinary teaching, providing a foundation for AI-assisted Indian education.
Significance
This work addresses the critical gap in multilingual, culturally rich instruction datasets for Indian education, enabling AI models to deliver culturally contextualized pedagogical content aligned with CBSE standards. It advances the integration of Indian Knowledge Systems into AI education, fostering cultural preservation and educational innovation. The methodology demonstrates how multi-source, multi-strategy data construction coupled with multi-dimensional evaluation can produce high-quality, culturally deep models, setting a new benchmark for domain-specific instruction tuning in underrepresented languages and traditions.
Technical Contribution
The study introduces a comprehensive pipeline combining multi-source data collection, three generation strategies, and rigorous multi-stage quality filtering. It innovates with a multi-dimensional scoring framework that evaluates fidelity, pedagogical quality, and cultural depth, ensuring high data integrity. Model fine-tuning on a 7B Transformer architecture with multi-turn dialogues and cross-tradition analysis significantly enhances the model’s pedagogical and cultural capabilities, demonstrating scalability and adaptability for complex, domain-specific tasks.
Novelty
This is the first systematic effort to construct a large-scale, multi-language, multi-technique instruction dataset grounded in Indian Knowledge Systems, with explicit curriculum alignment and multi-dimensional quality assessment. The integration of diverse pedagogical techniques, especially classical recitations and Vedic sutras, distinguishes it from existing general-purpose datasets, offering a new paradigm for culturally-aware AI education.
Limitations
- Despite extensive source coverage, some rare or highly specialized techniques and texts remain underrepresented, which may limit model generalization in niche areas.
- The multi-judge evaluation, while rigorous, involves subjective bias and may benefit from automation or larger reviewer pools in future iterations.
- Model scale is limited to 7B parameters; larger models might further improve performance but entail higher computational costs and complexity, requiring future optimization.
Future Work
Future directions include expanding data sources to incorporate more regional and lesser-known texts, refining generation and filtering techniques for higher fidelity, and integrating reinforcement learning with human feedback to further improve pedagogical interactions. Additionally, deploying models in real-world educational settings and establishing continuous update mechanisms will be crucial for long-term impact and relevance.
AI Executive Summary
Deep Dive
Plain Language Accessible to non-experts
想象你在一家传统的印度工坊学习古老技艺。老师用手势和口述，逐步教你如何编织、雕刻或绘画，每个步骤都非常细致。这个工坊里，老师会用古老的书籍和口头传授的知识，结合现代的教学方法，帮助你理解复杂的技艺。AI模型就像这个老师，通过大量的文化和教学内容学习，变得能用多种语言、讲解不同的技艺，还能根据你的理解水平调整难度。它的目标，是让每个学习者都能在自己的语言和文化背景下，轻松掌握传统技艺，同时传承宝贵的文化遗产。
ELI14 Explained like you're 14
想象你在学校里学一门特别的课程，老师用不同的语言讲解古老的印度数学和哲学，比如用梵语、泰米尔语和英语。老师会用很多例子，教你如何用古老的方法解题，还会和你用对话的方式，帮你理解复杂的概念。这个AI就像那个老师，它通过学习很多印度传统知识和教学技巧，变得非常聪明，能用你的语言讲解这些内容，还能帮你理解深奥的文化故事。它不仅能回答你的问题，还能引导你一步步学习，像个真正的老师一样。这样，你就可以在学习中既了解古老的文化，又掌握现代的知识，变得更聪明、更有文化底蕴！
Abstract
Instruction tuning has become the standard method for adapting large language models to follow human intent, yet existing instruction datasets are dominated by English-language general-knowledge tasks and lack coverage of specialized pedagogical domains. This paper presents IKS-Instruct, a dataset of 24,795 instruction-response pairs for teaching language models to deliver educational content grounded in Indian Knowledge Systems (IKS). The dataset spans seven languages (English, Hindi, Sanskrit, Tamil, Telugu, Kannada, and Malayalam), covers 41 pedagogical techniques from the Vedic oral and mathematical traditions, and is aligned with the Central Board of Secondary Education (CBSE) curriculum for classes 6 through 12. The pairs are derived from six source types: classical text corpora (Bhagavad Gita, Thirukkural, Sangam literature, Vedic texts), curriculum-aligned pedagogical templates, Vedic mathematical sutra demonstrations, bilingual instruction pairs, technique-grounded multi-turn dialogues, and cross-tradition comparative analyses. Quality is assessed through a multi-judge evaluation framework in which independent language models score responses on 12 dimensions including technique fidelity, pedagogical quality, factual accuracy, and IKS cultural depth. Under a uniform five-judge external panel (median aggregation over 1,201 stratified items), the strongest IKS-Instruct fine-tune of a compact 7B model reaches a median judge score of 6.39, within 0.15 of a strong general-purpose reference model (Nemotron-Nano at 6.54) at a fraction of its deployment cost, while the base model without IKS fine-tuning scores near zero on the IKS-specific dimensions. Model quality does not increase monotonically with data curation, a result we report together with the corresponding data-quality gains.