NMT Transformer (Dutch → English) on TED Talks (IWSLT)

Dataset

• Dataset: ted_talks_iwslt
• Config: nl_en_2016
• Original dataset provides a single split; it was split into 80/10/10 (train/valid/test) with seed=42.

Tokenizer

• SentencePiece BPE
• vocab_size=8000
• PAD=0, UNK=1, BOS=2, EOS=3

Model

• Encoder layers: 4
• Decoder layers: 4
• d_model: 256
• heads: 8
• FFN: 1024
• dropout: 0.1
• max_length: 96

Training

• AdamW lr=3e-4
• CosineAnnealingLR
• CrossEntropy with label_smoothing=0.1
• Trained on Colab GPU (T4)

Evaluation

Files

• nmt_transformer.pt: model checkpoint
• spm_nl_en.model / spm_nl_en.vocab: SentencePiece tokenizer
• inference.py: simple interactive translation script

Run inference