Does continued fine-tuning on corrected examples, new sentences, and dictionary entries improve the saved V1 translation system?
Data refinement
6,400 raw update records become 6,335 unique records, partitioned into 5,994 training records and 331 validation records after excluding 10 overlapping legacy records.
Training
NLLB-200-based seq2seq adaptation from V1: five epochs on one A100, with a 128-token source and target limit. The exact upstream checkpoint remains unspecified.
Evaluation
The final 149-pair benchmark reports BLEU 6.78 → 33.54. A separate 331-record development evaluation reports 7.27 → 22.18. Neither set is independently certified as unseen.
