Skip to main content

NLLB conversion and translation

Run from this recipe directory: cd recipes/nllb. This is an English-to-German inference example with facebook/nllb-200-1.3B; for translation training from scratch see WMT17.

Choose one tokenizer route. Use separate output directories so converting one route does not overwrite the other. Set HF_TOKEN if authentication is required.

Hugging Face tokenizer​

eole convert HF --model_dir facebook/nllb-200-1.3B \
  --output ./nllb-1.3b-hf --token "$HF_TOKEN"
printf '%s\n' 'What is the weather like in Tahiti?' > test.en
eole predict -c inference-hf.yaml

SentencePiece through OpenNMT tokenizer​

eole convert HF --model_dir facebook/nllb-200-1.3B \
  --output ./nllb-1.3b-onmt --token "$HF_TOKEN" --tokenizer onmt
printf '%s\n' 'What is the weather like in Tahiti?' > test.en
eole predict -c inference-pyonmttok.yaml

Both configurations use GPU 0 and write German translations to test.de. They add the source eng_Latn and target deu_Latn language prefixes. Change these prefixes together when changing the translation direction. Tokenizer settings are loaded from each converted checkpoint.