IKS-Instruct: A 24,000-Example Multilingual Dataset for Teaching Language Models Indian Knowledge Systems

Synthesis: Presents a 24,795-example multilingual instruction dataset for teaching LLMs to deliver educational content grounded in Indian Knowledge Systems. Spans seven languages and bridges a gap in non-Western pedagogical content for instruction tuning. Demonstrates that domain-specific educational datasets improve Large Language Models (LLMs) performance on culturally grounded knowledge tasks.

Key Findings

• IKS-Instruct contains 24,795 instruction-response pairs for teaching language models to deliver educational content grounded in Indian Knowledge Systems (IKS), spanning seven languages (English, Hindi, Sanskrit, Tamil, Telugu, Kannada, Malayalam).
• The dataset covers 41 pedagogical techniques from Vedic oral and mathematical traditions and is aligned with the Central Board of Secondary Education (CBSE) curriculum for classes 6 through 12.
• Pairs derive 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.
• The base model without IKS fine-tuning scores near zero on IKS-specific dimensions, and model quality does not increase monotonically with data curation — a result the authors report alongside data-quality gains.

Dataset Design & Method

IKS-Instruct targets a gap in instruction tuning: existing datasets such as Alpaca, FLAN, and Dolly are dominated by English-language general-knowledge tasks and lack coverage of specialized pedagogical domains. The six source types are engineered so that each pair carries structured metadata — source provenance, technique classification, language, subject, class level, and quality scores — released in JSONL format. This design lets the dataset support both instruction-following and culturally grounded pedagogy rather than generic fact recall.

What this means for practice

• Software developers. Fine-tune compact models on domain instruction data instead of renting general-purpose scale: the strongest 7B IKS fine-tune scored 6.39 under the five-judge panel, within 0.15 of Nemotron-Nano's 6.54 at a fraction of the deployment cost.
• Software developers. Anchor every pair in identifiable source material — classical texts, curriculum templates, manually verified sutra demonstrations — rather than free generation, since 67% of auto-generated Vedic mathematics content was wrong before filtering.
• Designers. Budget for bilingual pairs when targeting Language Learning and Culturally Relevant Pedagogy in under-resourced languages: Marathi and Bengali were absent from training, yet fine-tuned models demonstrated Vedic sutras there at a technique-fidelity score of 5.8 against 0.3 for the base model.
• Researchers. Score with a fixed external judge panel and read absolute values as comparative within that panel only, because version gaps such as 6.39 versus 6.54 are small relative to judge noise.

Limitations

• The language distribution is imbalanced — English at 36.1% against Malayalam at 6.3% — and fine-tuned models show weaker IKS capability in Telugu and Malayalam than in English and Hindi.
• Quality rests on LLM judges, which favor longer responses, formal academic language, and their own generation patterns; the panel includes Llama-family judges while the fine-tunes are LoRA adapters on the Llama-based Airavata backbone, so residual self-preference cannot be excluded and human expert calibration is left to future work.
• 67% of auto-generated Vedic mathematics pairs were erroneous before filtering, subtle errors may persist in retained pairs, and the manually verified v2.1 reduced coverage of mathematical techniques.
• Coverage stops at the CBSE classes 6-12 alignment — no primary (classes 1-5) or higher education — and oral guru-shishya traditions and manuscript-only texts are structurally absent because the dataset is built from digitally available sources.

Citation

Shwetha Singaravelu, Gayathri Muruganantham, Lakshmi Rajendran, Santhosh Sivasubramani (2026). IKS-Instruct: A 24,000-Example Multilingual Dataset for Teaching Language Models Indian Knowledge Systems.