JEE-GRPO-v1: Visual Reasoning Dataset for VLMs

Dataset Summary

JEE-GRPO is a high-quality, multimodal dataset designed for training and benchmarking Vision Language Models (VLMs) on complex STEM problems.

Unlike traditional text-only datasets, this dataset renders JEE Main (Joint Entrance Examination) questions as high-resolution images. This approach bypasses OCR errors and perfectly preserves complex LaTeX equations, diagrams, and chemical structures, making it ideal for visual reasoning tasks.

This specific version is optimized for Group Relative Policy Optimization (GRPO) training, focusing strictly on Multiple Choice Questions (MCQs) with deterministic ground truth.

Credits & Acknowledgements

This dataset is built upon the incredible data collection work done by the open-source community.

• Base Data Source: The raw questions, options, and metadata were sourced from the jee_mains_pyqs_data_base.
• Contribution: We gratefully acknowledge HostServer001 for maintaining the comprehensive database of JEE Main previous year questions (PYQs) that made this visual benchmark possible.

Dataset Structure

Each sample in the dataset represents a single MCQ from the JEE Main archive.

Feature	Type	Description
id	string	Unique identifier for the question (e.g., q_12345).
image	image	A 1000x1600 (approx) rendered snapshot containing the question text, diagrams, and options grid.
subject	string	The domain of the question: Mathematics, Physics, or Chemistry.
difficulty	string	Metadata indicating difficulty: Easy, Medium, or Hard.
answer	string	The correct option label (A, B, C, or D). Used as the reward signal for GRPO.

Quick Start

You can load this dataset directly using the Hugging Face datasets library:

from datasets import load_from_disk, load_dataset

# If loading locally
dataset = load_from_disk("jee_grpo_dataset_final")

# Example: View the first item
print(dataset[0])
# Output: {'id': 'q_842', 'subject': 'Physics', 'answer': 'B', 'image': <PIL.PngImagePlugin...>}

Creation Process

1. Source: Questions were extracted from a structured JEE Main database (2025 and prior).
2. Rendering: Raw HTML/LaTeX was rendered using a headless browser (Playwright) to generate "visual snapshots."
3. Layout: Questions use a standardized card layout with a 2x2 grid for options to ensure visual consistency for the model.
4. Filtering: strictly filtered for MCQs with valid options and ground truth.

Intended Use

• Training: Fine-tuning VLMs (like Qwen2-VL, Pixtral) using Reinforcement Learning (GRPO/PPO).
• Benchmarking: Evaluating the visual reasoning capabilities of models on high-school/undergraduate level STEM problems.

Limitations

• Visual Only: The input is strictly an image. The raw text is not included in the input feature (by design, to force visual grounding).
• MCQ Only: Numerical value questions are excluded from this specific split (see JEE-Bench-SFT for numericals).