NCERT Vidya
Two sizes (4B and 9B). One curriculum. CPT → SFT → DPO pipeline. 4.60 / 5.0 on Gemini-judge tutoring eval.
Run locally · Ollama 0.24.0+ · ollama.com/neosaket/vidya
$ ollama pull neosaket/vidya:9b
$ ollama run neosaket/vidya:9b "Explain Newton's second law with an example."Two sizes, one pipeline
Same CPT → SFT → DPO pipeline. Same NCERT / JEE / NEET corpora. Pick the size that fits your GPU.
NCERT Vidya 4B
4B parametersRuns on consumer GPUs
The lighter variant. Trained with the same CPT → SFT → DPO pipeline as 9B, optimised to fit on a 24 GB consumer GPU.
- Base model
- Qwen3.5-4B
- Max sequence length
- 2048 tokens
- VRAM (fine-tune)
- ~24 GB
- VRAM (q4_k_m inference)
- ~3 GB
- Ollama tag
- neosaket/vidya:4b
Best for
- •Personal tutoring laptops with a single 24 GB GPU
- •Lab and classroom deployments on commodity hardware
- •Rapid iteration before scaling up to 9B
NCERT Vidya 9B
9B parametersFlagship tutor
Our primary trained model. CPT → SFT → DPO on an RTX 5090, with an optional GRPO stage for JEE / NEET reasoning alignment. Scored 4.60 / 5.0 on Gemini-judge tutoring evaluation.
- Base model
- Qwen3.5-9B
- Max sequence length
- 2048 tokens
- Trained on
- RTX 5090 (32 GB VRAM)
- Tutoring eval
- 4.60 / 5.0 (Gemini judge)
- Ollama tag
- neosaket/vidya:9b (Ollama 0.24.0+)
Best for
- •End-to-end student tutoring across NCERT, JEE and NEET
- •Step-by-step problem solving in Physics, Chemistry, Math, Biology
- •Self-hosted deployments where data must stay on-prem
The training pipeline
Four stages. Each stage’s scripts and configs are open in the repo, so you can reproduce, audit or fork them.
1. CPT
Continued pre-training
Adapt the Qwen3.5 base on NCERT textbook corpora so the model learns the language, terminology and structure of Indian K-12 curricula.
2. SFT
Supervised fine-tuning
Instruction-tune on tutoring Q&A — NCERT chapter exercises, JEE/NEET past papers and synthetic Q&A generated with a local judge model.
3. DPO
Direct preference optimisation
Preference pairs steer the model toward step-by-step explanations, accurate maths and pedagogically helpful tone.
4. GRPO (optional)
Reasoning alignment
Optional fourth stage that uses reward-shaped rollouts to tighten JEE / NEET multi-step reasoning where it matters most.
What makes Vidya different
Built for Indian students
Trained on NCERT Classes 6-12, IIT-JEE and NEET content. Speaks the curriculum, not generic web text.
Step-by-step tutoring
Explains concepts, derives equations and solves problems the way a patient tutor does — with examples, intuition and checks.
Built on Qwen3.5
Inherits Qwen3.5’s multilingual base. Fine-tuned on English NCERT, JEE and NEET content — with the Qwen base intact for code-switching when students need it.
Local-first deployment
Exports to GGUF q4_k_m and runs via Ollama on your own machine. No data leaves your school, lab or home.
Open data, open weights
Trained on publicly licensed sources — NCERT textbooks, open exam-paper datasets, JEE/NEET benchmarks — with provenance documented.
Honest about limits
Calibrated to say “I don’t know” rather than hallucinate. Evaluated with an LLM-as-judge tutoring rubric, not just MCQs.
4.60 / 5.0
Tutoring quality on the Gemini-judge evaluation set — covering NCERT explanations, JEE problem solving and NEET conceptual questions.
Build with NCERT Vidya
Open source, open weights, locally deployable. Use it in your school, study app, or research project — or fine-tune it further for your subject of choice.