An end-to-end English–French neural machine translation project, built in two stages: first with a custom LSTM encoder-decoder architecture, then rebuilt using a fine-tuned Hugging Face transformer for significantly stronger results.

The project demonstrates a full machine translation workflow — data preprocessing, model training, evaluation against a baseline, and an interactive Streamlit inference app.

Evaluated on 500 held-out sentences from the Tatoeba corpus:

The fine-tuned transformer matches the Helsinki-NLP/opus-mt-en-fr base model on BLEU while outperforming it on chrF — the more morphologically sensitive metric. See the transformer README for full evaluation details and example translations.

seq2seq-en-fr/
├── transformer_version/    # Hugging Face fine-tuned transformer
│   └── README.md           # Full details, setup and results
├── lstm_version/           # Custom LSTM encoder-decoder (v1)
│   └── README.md           # Architecture details and learnings
└── README.md               # This file

This project grew out of a deliberate progression — starting with a hand-built LSTM to understand the mechanics of sequence-to-sequence learning, then moving to a transformer architecture to see the improvement modern pretraining brings. The LSTM version is preserved in this repo as a record of that learning process.

• Python, PyTorch, Hugging Face Transformers
• Seq2SeqTrainer, AutoTokenizer
• Streamlit (inference UI)
• sacrebleu, chrF (evaluation)

• v1 — LSTM: Custom encoder-decoder with GRU/LSTM cells, embedding layers, trained from scratch on Tatoeba pairs
• v2 — Transformer: Fine-tuned Helsinki-NLP/opus-mt-en-fr on 50,000 sentence pairs, evaluated with BLEU and chrF