Nalandadata is a research team building verified reasoning data for frontier models, and proving what it does to them.

Our data produces measurable lift across model families, modalities and training methods, all in the open. On document understanding, a fine-tuned 8B model reads tables from images more accurately than every frontier cloud model we tested, including GPT-5.5 and Gemini 3 Pro. On step-by-step reasoning, verified-reward reinforcement learning on our solutions took a 7B model from 60.5% to 66.8%. On multimodal understanding, fine-tuning a vision model on our diagram data raised held-out accuracy by 9.3 points. Different base models, different methods, one common source of the gain: the data. Every model, sample and held-out test set is public at huggingface.co/Nalandadata, so each result can be reproduced.

The data behind these results is human-authored and rights-cleared, drawn from S Chand Group's 86-year archive of 11,000+ expert-authored titles, 100% owned, with full provenance. No web scraping. No synthetic generation. Data written to teach, not scraped to exist.

We work with teams across the pipeline, not only at licensing: custom dataset curation for a target capability, fine-tuning and verified-reward RL (RLVR), evaluation, and alignment data — the same approach that produced the results above, applied to your domain.

We build for frontier AI labs, foundation-model teams, enterprise AI teams and sovereign AI programmes. See the research and request a sample at nalandadata.ai.

Nalandadata

Org chart

Teams

Offices

Nalandadata

Org chart

Teams

Offices

Related companies