Specialised domain translation using Large Language Models remains challenging due to terminology constraints, scarce in-domain parallel data, and the tendency of fine-tuning to degrade in-context learning performance. We present our submission for the English–Polish and Spanish–Basque translation in the Medical and Engineering & Technology domains. Our method follows three stages: i) We extract in-domain terms from a terminology database and prompt a multilingual LLM to generate contextual sentences for each term. ii) We build an efficient search index over large-scale, unconstrained web-crawled and medical parallel corpora. Using the synthetic sentences as queries, we retrieve and filter the most semantically relevant pairs to build an in-domain fine-tuning dataset. iii) We structure our training instances with a mixture of 0-shot and semantically retrieved 5-shot prompts, and fine-tune lightweight multilingual LLMs (Gemma-3-4B, and Tiny-Aya-Global) via QLoRA. Fine-tuned Gemma-3 achieves the best results on held-out data across all metrics on English–Polish (Medical) and on Spanish–Basque (Engineering & Technology), showing that few-shot fine-tuning enhances specialised domain translation quality and preserves in-context adaptation.

In-context learning is a method for improving machine translation in Large Language Models, but its performance is sensitive to the quality of the few-shot example selection. Current retrieval strategies use semantic similarity by computing sentence embeddings, and these methods often require significant computational overhead and specialised expertise. We evaluate the impact of retrieval strategies on translation performance in a specialised domain, comparing traditional, token-based fuzzy matching against semantic sentence embeddings. We use a medical corpus from the European Medicines Agency (EMEA) for the English-Romanian and English-German language pairs, and we evaluate translation quality with automatic metrics and manual evaluation. Our results show that 1-shot and 5-shot prompting significantly outperforms the 0-shot baselines for quality in automatic evaluations for both language pairs, and in manual evaluation for English-German. For the English-Romanian pair, the average scores of the manual evaluation for both quality and ranking follow the same trend, but with no statistical significance. Token-based fuzzy matching overwhelmingly has higher automatic quality scores than embedding-based retrieval.

Multilingual Neural Machine Translation (MNMT) models allow to translate across multiple languages based on only one system. We study the quality of a domain-adapted MNMT model in the medical domain for English-Romanian with automatic metrics and a human error typology annotation based on the Multidimensional Quality Metrics (MQM). We further expand the MQM typology to include terminology-specific error categories. We compare the out-of-domain MNMT with the in-domain adapted MNMT on a standard test dataset of abstracts from medical publications. The in-domain MNMT model outperforms the out-of-domain MNMT in all measured automatic metrics and produces fewer errors. In addition, we perform the manual annotation over the reference test dataset to study the quality of the reference translations. We identify a high number of omissions, additions, and mistranslations in the reference dataset, and comment on the assumed accuracy of existing datasets. Finally, we compare the correlation between the COMET, BERTScore, and chrF automatic metrics with the MQM annotated translations. COMET shows a better correlation with the MQM scores compared to the other metrics.