Character-level Swedish → English translator (EDAN20, lab 6)
Transformer encoder–decoder (3+3 layers, d_model 512, 8 heads) with a shared character vocabulary and tied input/output embeddings. Trained on the Tatoeba/Anki Swedish–English pairs (swe.txt). Based on Pierre Nugues, Python for Natural Language Processing, ch. 17.
Files: pico_model.pth (weights), vocab.json (char → index),
config.json (model size), pico_model.npz (same weights in NumPy).
To use: build Translator(**config) from the notebook, then
model.load_state_dict(torch.load("pico_model.pth")).