Sanskrit-to-English Neural Machine Translation

Assignment 2 implementation for Sanskrit-to-English neural machine translation.

Repository Contents (for GitHub submission)

This repository contains everything required for assignment submission and evaluation:

File / Folder	Purpose
Sanskrit_English_NMT_Assignment.ipynb	Training and inference code (.ipynb only, as required)
Datasets/	Provided Sanskrit-English train/dev/test CSV files
submission.csv	Final test-set predictions (Source_id, Sentence_en)
outputs/metrics.json	Final BLEU, BERTScore, inference time, and parameter count
outputs/training_history.csv	Training loss per epoch
outputs/training_curve.png	Training curve plot
outputs/translation_examples.json	Example translations with references
Sanskrit_English_NMT_Report.docx	Assignment report
requirements.txt	Python dependencies with install instructions

Pretrained Models and Methods Used

Component	Pretrained?	Details
Translation model (encoder-decoder GRU + attention)	No	Trained from scratch on the provided dataset only
BERTScore evaluation (bert-score)	Yes	Uses roberta-large downloaded locally for metric computation only
External translation APIs	No	Not used

The translation model does not use any pretrained weights. The only pretrained model is roberta-large, used exclusively by the bert-score library to compute BERTScore F1 during evaluation. It is not part of the translation pipeline.

How to Run

Use WSL and the local virtual environment:

source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook Sanskrit_English_NMT_Assignment.ipynb

Run all cells in order. The notebook will:

1. Load and preprocess the provided datasets
2. Train the custom seq2seq model with attention
3. Evaluate on dev and test sets (BLEU + BERTScore)
4. Report inference time and parameter count
5. Write submission.csv

Final Scores (Provided Test Set)

Metric	Value
Dev BLEU	0.0235
Test BLEU	0.0253
Test BERTScore F1	-0.0406
Test inference time	57.34 seconds
Trainable parameters	6,629,824

Notes

• Custom PyTorch encoder-decoder with additive attention and beam search
• No external APIs are used
• Only the provided assignment dataset is used for training
• BERTScore uses rescale_with_baseline=True as required