Nalandadata/nalanda-qwen-7b-grpo
Text Generation • 8B • Updated • 2
Request access to Nalandadata
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access requires manual review. Please tell us who you are and how you intend to use this resource so we can support your work and share licensing options. By requesting access you agree not to redistribute the data and to contact info@nalandadata.ai for commercial licensing.
Log in or Sign Up to review the conditions and access this dataset content.
NalandaJEENEETBench: JEE & NEET STEM Evaluation Benchmark
The first open benchmark for evaluating LLMs on Indian competitive exam questions (JEE Mains, JEE Advanced, NEET UG).
NalandaJEENEETBench is a curated sample from Nalanda Data's proprietary dataset of 116,000+ expert-curated JEE and NEET examination questions with verified correct answers and step-by-step solutions.
📦 Public sample (no login required): Nalandadata/NalandaJEENEETBench-sample — 10 benchmark questions, free, no sign-up required.
🔬 Fine-tuned model: Nalandadata/nalanda-qwen-7b-grpo — Qwen 2.5 7B fine-tuned on the full 116K dataset.
Live demo: Nalanda Live Demo — test JEE & NEET questions against our fine-tuned models.
What's Inside
| Split | Questions | Content | Purpose |
|---|---|---|---|
benchmark |
785 | MCQs with correct answers (no solutions) | Evaluate your model on JEE/NEET |
train_sample |
500 | MCQs with correct answers + full step-by-step solutions | Preview our training data quality |
Both splits cover 5 subjects: Physics, Chemistry, Mathematics, Biology, and English (reading comprehension). Subject distribution reflects the natural mix of JEE/NEET question pools rather than enforced per-subject quotas.
Why This Benchmark Matters
There is no standard evaluation suite for Indian competitive exams. MMLU covers Western curricula. NalandaJEENEETBench fills this gap:
- 785 held-out MCQs across the full JEE/NEET syllabus
- Verified correct answers — every question has been expert-validated
- LaTeX math notation — proper representation of equations and formulas
- Multi-subject coverage — Physics, Chemistry, Mathematics, and Biology (each 150+ questions), plus a small English reading-comprehension slice
Proven Results
Note (2026-06-01): The accuracy figures in the table below were computed against an earlier version of this benchmark (800 rows across 4 subjects, before content-label corrections). The current 785-row, 5-subject benchmark (Physics 167 · Chemistry 167 · Mathematics 207 · Biology 223 · English 21) may yield different per-subject numbers. Re-evaluation against the corrected split is planned for a future release.
We used the full 116K dataset (not this sample) to fine-tune Qwen 2.5 7B using a two-stage GRPO pipeline.
Results on NalandaJEENEETBench (this benchmark)
| Subject | Qwen 7B Baseline | Llama 70B Baseline | + Our Data (7B) | Improvement |
|---|---|---|---|---|
| Physics | 51.0% | 59.0% | 65.0% | +14.0pp |
| Chemistry | 61.5% | 70.0% | 71.5% | +10.0pp |
| Mathematics | 56.0% | 57.5% | 64.5% | +8.5pp |
| Biology | 73.5% | 66.5% | 77.5% | +4.0pp |
| Overall | 60.5% | 63.2% | 69.6% | +9.1pp |
A 7B model fine-tuned on our data outperforms a 70B model across all STEM subjects. The data is the differentiator.
Quick Start
Evaluate your model
from datasets import load_dataset
# Load the benchmark
bench = load_dataset("Nalandadata/NalandaJEENEETBench", split="benchmark")
for q in bench:
prompt = (
f"Question: {q['question']}\n\n"
f"(A) {q['option_a']}\n"
f"(B) {q['option_b']}\n"
f"(C) {q['option_c']}\n"
f"(D) {q['option_d']}\n\n"
f"Think step by step. State your final answer as A, B, C, or D."
)
# Run your model on `prompt`
# Compare output to q['correct_answer']
Preview training data quality
# Load the training sample (includes step-by-step solutions)
train = load_dataset("Nalandadata/NalandaJEENEETBench", split="train_sample")
for q in train.select(range(5)):
print(f"[{q['subject']}] {q['question'][:100]}...")
print(f"Answer: ({q['correct_answer']})")
print(f"Solution: {q['solution'][:200]}...")
print("---")
Data Schema
benchmark split
| Column | Type | Description |
|---|---|---|
subject |
string | Physics, Chemistry, Mathematics, Biology, or English |
question |
string | Question text (may contain LaTeX) |
option_a |
string | Option (A) |
option_b |
string | Option (B) |
option_c |
string | Option (C) |
option_d |
string | Option (D) |
correct_answer |
string | Correct option letter (A, B, C, or D) |
train_sample split
Same as benchmark, plus:
| Column | Type | Description |
|---|---|---|
solution |
string | Full step-by-step solution with reasoning |
The Full Dataset
This is a sample. The full Nalanda Data question bank contains:
| Property | Value |
|---|---|
| Total questions | 116,000+ |
| Physics | 27,800+ |
| Chemistry | 52,500+ |
| Mathematics | 25,700+ |
| Biology | 9,900+ |
| Format | MCQ with verified correct answers + step-by-step solutions |
| Notation | LaTeX mathematical notation throughout |
| Sources | JEE Mains, JEE Advanced, NEET UG, Board Exams |
| Quality | Expert-curated, subject-label verified, deduplicated |
What makes the full dataset valuable for AI training
- Verified correct answers — enables reinforcement learning (GRPO/DPO/PPO), not just SFT
- Step-by-step solutions — teaches reasoning chains, not just answer selection
- LaTeX math notation — proper STEM representation that models can learn from
- Proven results — demonstrated +9.1pp improvement on a 7B model, outperforming 70B baselines
- Competitive moat — proprietary data that cannot be scraped or replicated
Licensing
The full dataset is available for licensing. Use cases include:
- Foundation model pre-training — enrich STEM reasoning in base models
- Supervised fine-tuning — domain adaptation for education and tutoring
- Reinforcement learning — GRPO/DPO with verified correct answers as reward signal
- Evaluation — benchmark your models on Indian competitive exam performance
- EdTech products — power AI tutoring, question generation, and adaptive learning
Contact
For commercial licensing, full dataset access, custom data work, or partnerships: 📧 info@nalandadata.ai
For technical questions, integration help, or fine-tuning support: 📧 tech@nalandadata.ai
Citation
If you use NalandaBench in your research, please cite:
@misc{nalandabench2026,
title={NalandaJEENEETBench: A JEE and NEET STEM Evaluation Benchmark for Large Language Models},
author={Nalanda Data},
year={2026},
url={https://huggingface.co/datasets/Nalandadata/NalandaJEENEETBench},
note={Sample from 116K+ expert-curated Indian competitive exam questions. Part of the NalandaBench suite.}
}
Related
- Fine-tuned model: Nalandadata/nalanda-qwen-7b-grpo — Qwen 2.5 7B fine-tuned on the full dataset
- Paper: Coming soon — "Domain Data Drives LLM Fine-Tuning Performance: Lessons from 116K JEE/NEET Questions"
License
This sample dataset is released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial). You may use it for research and evaluation. Commercial use of the full dataset requires a license from Nalanda Data.
- Downloads last month
- 15
Models trained or fine-tuned on Nalandadata/NalandaJEENEETBench