This work describes the Grial-TA team submission to the constrained track of WMT2026 General Machine Translation Task for the following language pairs: English to-Icelandic, English-to-Ladin and English-to-Ligurian. Our submission leverages full fine-tuning of theSalamandraTA-7b-instruct model. Two separated models were fine-tuned: one for Icelandic and one for Ligurian and Ladin.

This submission tackles both tracks (in English-Polish) of the shared task using an instruction-aligned Gemma 4 model fine-tuned on highly curated, domain-specific corpora. For Track 2, we overcome the lack of explicit dictionaries by employing a combined BERT and POS-based automatic term extraction process.