Arshavk/dl-genai-project

Deep Learning & Generative AI Project – Multiple Choice Question Answering

Overview

This repository contains a fine-tuned RoBERTa model for Multiple Choice Question Answering (MCQA). The model was developed as part of the IIT Madras BS Degree - Deep Learning and Generative AI Project.

The complete project combines three models:

• RoBERTa (Transformer)
• BiLSTM
• Support Vector Machine (SVM)

The final predictions are generated using a weighted ensemble of these models, while this repository contains the fine-tuned RoBERTa model used in the ensemble.

Dataset

• Multiple Choice Question Answering dataset
• Five answer options (A, B, C, D, E)
• Train/Validation split for model evaluation
• Duplicate text samples removed before training

Model

• Base Model: roberta-base
• Framework: Hugging Face Transformers
• Task: Text Classification
• Number of Classes: 5 (A-E)

Training Details

• Optimizer: AdamW
• Early Stopping enabled
• Weight Decay: 0.01
• Evaluation Strategy: Every Epoch
• Best model selected using Validation Macro F1

Validation Performance

Metric	Score
Validation Accuracy	1.0000
Validation Macro F1	1.0000

Ensemble

The complete project also evaluates:

• Support Vector Machine (SVM)
• BiLSTM
• Fine-tuned RoBERTa
• Weighted Ensemble

RoBERTa is the strongest individual deep learning model used in the ensemble.

Files

• model.safetensors
• config.json
• tokenizer.json
• tokenizer_config.json

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("Arshavk/dl-genai-project")
model = AutoModelForSequenceClassification.from_pretrained("Arshavk/dl-genai-project")

Author

Arsha Vinesh

IIT Madras BS Degree Programme

Deep Learning and Generative AI Project