DL Assignment 2 — Sequence Models & GPT-2 Fine-tuning

This repository contains code and documentation for Deep Learning Assignment 2 under the course 22CAC04 - Deep Learning at CBIT.

🔥 Assignment Goals

Question 1: Sequence-to-Sequence Transliteration

• Build a seq2seq model using RNNs to transliterate between Latin and Devanagari scripts.
• Flexible architecture supporting:
  • Character-level embeddings
  • Choice of RNN / GRU / LSTM cells
  • Variable embedding size, hidden size, and layers
• Evaluated on the Dakshina dataset.
• Computes:
  • Total number of operations
  • Total number of parameters
• Reports accuracy and outputs on test set.

Question 2: Fine-tuning GPT-2 for Lyrics Generation

• Cleaned and combined lyrics of Khalid and Lady Gaga
• Fine-tuned GPT-2 using Hugging Face Transformers and Datasets libraries
• Evaluated by generating sample lyrics from a custom prompt

📦 Requirements

• torch, transformers, datasets, pandas, sklearn
• Recommended: Google Colab or a machine with GPU

🚀 How to Run

1. Clone the repo
2. Install dependencies:

   pip install -r requirements.txt