Updesh is an Indian language synthetic text dataset released by Microsoft in 2025 to facilitate post-training of Large Language Models (LLMs) for Indian languages. The dataset contains 6,800,000 inference data and 2,100,000 generated data in the following languages: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu, and Urdu.

Citation

@misc{chitale2026updeshsynthesizinggroundedinstruction, title={UPDESH: Synthesizing Grounded Instruction Tuning Data for 13 Indic}, author={Pranjal A. Chitale and Varun Gumma and Sanchit Ahuja and Prashant Kodali and Manan Uppadhyay and Deepthi Sudharsan and Sunayana Sitaram}, year={2026}, eprint={2509.21294}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2509.21294}, }