Spaces:
Running
Running
metadata
title: README
emoji: 🪔
colorFrom: indigo
colorTo: yellow
sdk: static
pinned: false
Nalandadata
Verified, curriculum-aligned Indian STEM data for frontier AI labs
Training · Post-training · Evaluation — across reasoning, multimodal understanding, and document intelligence.
Nalandadata builds high-quality, curriculum-aligned data sourced from S. Chand — India's largest academic textbook publisher — spanning all subjects, grade levels, and major Indic languages alongside English. Textbook content is structured, expert-authored, and verified, which makes it valuable far beyond education: reasoning chains, scientific diagrams, structured tables, and multilingual content that transfer directly to general-purpose model training and evaluation.
📦 Products
Datasets
| Dataset | What it is |
|---|---|
| NalandaJEENEETBench | 116,831 JEE & NEET questions with verified answers + worked solutions. RLVR-ready ground truth. |
| nalanda-image-qa | 22,000+ scientific image Q&A pairs from NCERT diagrams (physics, chemistry, biology). |
| DrishtiTable | 1,421 annotated tables for document AI / table structure recognition — with a full TEDS benchmark + leaderboard. |
Models
| Model | Result |
|---|---|
| nalanda-qwen-7b-grpo | Qwen-7B + GRPO on NalandaJEENEETBench: +6.3pp (vs −16pp for naive SFT) — verified answers make RLVR work. |
| nalanda-image-vl | Multimodal diagram understanding: +9.3pp over zero-shot. |
| DrishtiTable-Qwen2.5-VL-7B | Table recognition at 83.2% TEDS — beats GPT-4o on our benchmark. |
Benchmarks & demos
- 🏆 DrishtiTable Leaderboard — live TSR leaderboard ranked by TEDS.
- 🔬 Nalanda Live Demos — try our models on STEM text & images.
✅ Why it works
- Verified ground truth → every JEE/NEET item has a checkable answer, enabling RLVR / GRPO pipelines that actually improve capability.
- Expert-authored, structured source → reasoning chains, diagrams, and tables, not scraped web noise.
- Multilingual, curriculum-aligned → English + major Indic languages across all grade levels.
📜 Licensing & access
We license datasets for AI training, post-training, and evaluation.