NalandadataVerified reasoning data
Request access

Research

The experiments behind the claims.

Each case study shows what verified, curriculum-grade data does to a real model — the method, the held-out numbers, and a public mirror of the model and a data sample on Hugging Face. Everything here is reproducible.

84.9% TEDS
DrishtiTable
Table structure recognition
A fine-tuned 8B model tops every frontier cloud model on table extraction.
Read the case study →
+6.3 pts
NalandaBench
STEM reasoning · RLVR (GRPO)
Plain fine-tuning lost 16 points on the same data; verified-reward RL turned it into a 6.3-point gain.
Read the case study →
+9.3 pts
Nalanda Image VL
Multimodal science · vision-language
Fine-tuning a vision model on diagram data lifted held-out accuracy by 9.3 points.
Read the case study →

Reproduce it

The models and dataset samples are public.

Download the released checkpoints and held-out sets, run them on your own inputs, and check the numbers for yourself.

Hugging Face Browse our published models & dataset samples →