ParamBench: A Graduate-Level Benchmark for Evaluating LLM Understanding on Indic Subjects

ParamBench: A Graduate-Level Benchmark for Evaluating LLM Understanding on Indic Subjects
Ayush Maheshwari, Kaushal Sharma, Vivek Patel, Aditya Maheshwari

📋 Overview

ParamBench is a comprehensive graduate-level benchmark in Hindi designed to evaluate Large Language Models (LLMs) on their understanding of Indic subjects. The benchmark contains 17,275 multiple-choice questions across 21 subjects, covering a wide range of topics from Indian competitive examinations.

This benchmark is specifically designed to:

• Assess LLM performance on culturally and linguistically diverse content
• Evaluate understanding of India-specific knowledge domains
• Support the development of more culturally aware AI systems

On Huggingface 🤗 🚀

We're excited to announce that the ParamBench dataset is now publicly available on the Hugging Face Hub!

You can find the dataset here: https://huggingface.co/datasets/bharatgenai/ParamBench

🎯 Key Features

• 17,275 Questions: Extensive collection of graduate-level MCQs in Hindi
• 21 Subjects: Comprehensive coverage of diverse academic domains
• Standardized Format: Consistent question structure for reliable evaluation
• Automated Evaluation: Scripts for benchmarking and analysis
• Detailed Metrics: Subject-wise and question-type-wise performance analysis

📊 Dataset Structure

Question Format

Each question in the dataset includes:

• unique_question_id: Unique identifier for each question
• question_text: The question text
• option_a, option_b, option_c, option_d: Four multiple choice options
• correct_answer: The correct option (A, B, C, or D)
• subject: Subject category
• exam_name: Source examination
• paper_number: Paper/section identifier
• question_type: Type of question (MCQ, Blank-filling, assertion/reasoning, etc.)

Subject Distribution

The benchmark covers 21 subjects including but not limited to:

• Music
• History
• Drama and Theatre
• Economics
• Anthropology
• Current Affairs
• Indian Culture
• And more...

🏗️ Repository Structure

ParamBench/
├── data/
│   └── full-data.csv          # Main dataset file
├── checkpoints/                    # Model evaluation checkpoints
├── results/                        # Analysis results and visualizations
├── benchmark_script.py             # Main benchmarking script
├── analysis_models.py              # Analysis and visualization script
├── requirements.txt                # Python dependencies
└── README.md                       # This file

🚀 Quick Start

Requirements

pip install -r requirements.txt

Basic Requirements

• Python 3.8+
• PyTorch 2.0+
• Transformers 4.45+
• Pandas
• NumPy
• Plotly (for visualization)

Running Benchmarks

1. Clone the repository

git clone https://github.com/yourusername/ParamBench.git
cd ParamBench

2. Run the benchmark

python benchmark_script.py

Configuration Options

The benchmark script supports various configuration options:

# In benchmark_script.py
group_to_run = "small"  # Options: "small", "medium", "large", or "all"
batch_size = 16         # Adjust based on GPU memory

📊 Running Analysis

After running benchmarks, generate comprehensive analysis reports:

python analysis_models.py

This will generate:

• Model performance summary CSV
• Subject-wise accuracy charts
• Question type analysis
• Combined report with all metrics

Citation

@article{maheshwari2025parambench,
  title={ParamBench: A Graduate-Level Benchmark for Evaluating LLM Understanding on Indic Subjects},
  author={Maheshwari, Ayush and Sharma, Kaushal and Patel, Vivek and Maheshwari, Aditya},
  journal={arXiv preprint arXiv:2508.16185},
  year={2025}
}