source
"The cats
are running."
eng_Latn
tokenize
subword ids
[1201, 412,
8834, 92]
SentencePiece
encoder-decoder
NLLB, mBART,
M2M, Marian
forced_bos = fra_Latn
beam search
target ids
[831, 5102,
2918, 4]
fra_Latn
detok
output
"Les chats
courent."
watch for: hallucinations, off-target language, terminology drift, formality mismatch
score with sacrebleu (BLEU + chrF); always reference-matched