EDAN20 lab 6: Swedish → English character-level transformer
These are the weights for an encoder-decoder transformer (nn.Transformer with shared, weight-tied character embeddings) from the notebook 17_02_translation_transformer_shared_embs.ipynb in Python for Natural Language Processing (P. Nugues). The model was trained on the Tatoeba swe-eng pairs from manythings.org/anki.
| File | Build | d_model | Layers (enc+dec) | FF | Heads | Max len | Params | Epochs | Val. acc. | BLEU |
|---|---|---|---|---|---|---|---|---|---|---|
gpu_model.pth + gpu_translator.vocab |
GPU (RTX 5070 Ti) | 512 | 3+3 | 2048 | 8 | 122 | 22.1 M | 80 | 0.913 | 0.505 |
pico_model.pth + pico_translator.vocab |
Laptop CPU | 256 | 2+2 | 1024 | 8 | 62 | 3.7 M | 5 | 0.637 | 0.043 |
The .vocab files are JSON token2idx maps: characters, plus <pad>=0, <unk>=1, <s>=2 and </s>=3. Decoding is greedy, one character at a time.
Demo: https://huggingface.co/spaces/IsacEP/swe-eng-pico-translator
Browser version
gpu.npz/gpu.json and laptop.npz/laptop.json contain the same weights as NumPy arrays (float32), plus a config with the vocabulary. The static demo loads them and runs a NumPy re-implementation of the model in the browser through Pyodide. It gives the same translations as the PyTorch model. fc.weight is left out because it is tied to embedding.input_embedding.weight.