worldboss/en-ee-from-scratch

worldboss/en-ee-from-scratch

From-scratch English → Ewe encoder–decoder (Pre-LN, 6×512, vocab 32000). This is not an NLLB / Transformers AutoModelForSeq2SeqLM checkpoint. Load best.pt with translate.py from the training repo.

Training data includes Ghana farmer Q&A, NLLB eng_Latn-ewe_Latn, and verse-aligned English–Ewe Bible from `ghananlpcommunity/ghana-corpus`. Farmer Q&A, NLLB, and the Bible corpus are CC-BY-NC.

Load and translate

huggingface-cli download worldboss/en-ee-from-scratch --local-dir en-ee-from-scratch
python translate.py --checkpoint en-ee-from-scratch/best.pt --text "How do I plant maize?"

from pathlib import Path
import torch
from translate import load_checkpoint, translate

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, tokenizer, cfg = load_checkpoint(Path("en-ee-from-scratch/best.pt"), device)
print(translate(model, tokenizer, cfg, "How do I plant maize?", device))

The Hub snapshot ships best.pt plus the BPE tokenizer (tokenizer.json). translate.py resolves tokenizer_dir="." next to the checkpoint.

Sample prompts:

• —How do I plant maize?
• —Keep the soil moist.
• —When should I harvest the crops?

Compare with the NLLB fine-tune using python compare_models.py in the training repo.