bergamot-mt-cy-en
Welsh → English translation model for on-device deployment.
Built with Marian NMT using the Bergamot tiny architecture, trained on 1.67M Welsh-English sentence pairs.
Architecture
| Parameter | Value |
|---|---|
| Encoder layers | 4 |
| Decoder layers | 1 |
| Hidden dim | 256 |
| FFN dim | 1024 |
| Vocabulary | 16k joint SentencePiece |
| Quantization | intgemm8 (int8) |
| Model size | ~15–20MB |
| Peak RAM (inference) | ~50–80MB |
Files
| File | Description |
|---|---|
model.intgemm8.bin |
Quantized Marian model (deploy this) |
vocab.cy.spm |
Welsh SentencePiece vocabulary |
vocab.en.spm |
English SentencePiece vocabulary |
Inference
Use with bergamot-translator or Marian decoder:
marian-decoder \
-m model.intgemm8.bin \
-v vocab.cy.spm vocab.en.spm \
--cpu-threads 4 \
--beam-size 1