En → Fr Translator

Encoder-decoder transformer · trained from scratch on WMT14

Note

This is a proof-of-concept modal trained with limited compute. So, translations won’t be accurate and is intended for demonstration purposes only.

English

500 remaining

French

Translation will appear here

Encoder-decoder transformer · BPE tokeniser · beam search (width 5)

How it works

A 6-layer encoder-decoder transformer (512d, 8 heads, ~65M params) trained from scratch using PyTorch on the WMT14 English-French corpus. BPE tokenisation with a shared 32k vocabulary. Beam search decoding (width 5) with length normalisation. Served on Modal’s serverless infrastructure.