Swedish → English pico translator
A character-level encoder-decoder transformer trained on about 20,000 Tatoeba sentence pairs (EDAN20, lab 6). Two builds of the same model run side by side, in your browser. Short, everyday sentences work best.
Loading Python…
GPU build
Downloading weights (89 MB)…
Laptop build
Downloading weights (15 MB)…
About the models
Both builds come from the notebook 17_02_translation_transformer_shared_embs (Nugues, Python for NLP). They read and write one character at a time, with greedy decoding. The browser runs a NumPy re-implementation of the trained PyTorch model via Pyodide. It gives exactly the same translations as PyTorch.
| GPU build | Laptop build | |
|---|---|---|
| d_model / layers (enc+dec) | 512 / 3+3 | 256 / 2+2 |
| Training pairs / epochs | 20,521 / 80 | 12,000 / 5 |
| Val. char accuracy | 0.913 | 0.637 |
| BLEU (validation set) | 0.505 | 0.043 |