We present the BSC submission to the WMT26 General Machine Translation shared task: a 7B decoder-only model covering 23 directions, adapted from the SALAMANDRA (Gonzalez-Agirre et al., 2025) family through vocabulary replacement, two rounds of continued pre-training (monolingual, followed by instruction-formatted parallel data), supervised fine-tuning on a translation-only instruction mixture, and preference optimisation applied selectively to six weak directions via CPO-SimPO (Xu et al., 2024; Meng et al., 2024) with LoRA adapters (Hu et al., 2021). Submissions are decoded with Minimum Bayes Risk selection over a candidate pool filtered for truncation, degeneracy and off-target script. As official human evaluation was unavailable at the time of writing, we report a stage-wise internal evaluation on FLORES+ and WMT24++, together with automatic scores for the submitted system on the 12 test-set directions for which references have been released. Instruction tuning accounts for almost all measurable gain, but its size differs by an order of magnitude between the two evaluation sets; preference optimisation moves automatic metrics by less than the drift among untreated directions. The model is released under a Research-Only Licence and is available on request.

In this paper, we present the SalamandraTA family of models, an improved iteration of Salamandra LLMs (Gonzalez-Agirre et al., 2025) specifically trained to achieve strong performance in translation-related tasks for 38 European languages. SalamandraTA comes in two scales: 2B and 7B parameters. For both versions, we applied the same training recipe with a first step of continual pre-training on parallel data, and a second step of supervised fine-tuning on high-quality instructions.The BSC submission to the WMT25 General Machine Translation shared task is based on the 7B variant of SalamandraTA. We first extended the model vocabulary to support the additional non-European languages included in the task. This was followed by a second phase of continual pretraining and supervised fine-tuning, carefully designed to optimize performance across all translation directions for this year’s shared task. For decoding, we employed two quality-aware strategies: Minimum Bayes Risk Decoding and Translation Reranking using Comet and Comet-kiwi.We publicly release both the 2B and 7B versions of SalamandraTA, along with the newer SalamandraTA-v2 model, on Hugging Face.

High-quality machine translation requires datasets that not only ensure linguistic accuracy but also capture regional and cultural nuances. While many existing benchmarks, such as FLORES-200, rely on English as a pivot language, this approach can overlook the specificity of direct language pairs, particularly for underrepresented combinations like Catalan-Chinese. In this study, we demonstrate that even with a relatively small dataset of approximately 1,000 sentences, we can significantly improve MT localization. To this end, we introduce a dataset specifically designed to enhance Catalan-to-Chinese translation by prioritizing regionally and culturally specific topics. Unlike pivot-based datasets, our data source ensures a more faithful representation of Catalan linguistic and cultural elements, leading to more accurate translations of local terms and expressions. Using this dataset, we demonstrate better performance over the English-pivot FLORES-200 dev set and achieve competitive results on the FLORES-200 devtest set when evaluated with neural-based metrics. We release this dataset as both a human-preference resource and a benchmark for Catalan-Chinese translation. Additionally, we include Spanish translations for each sentence, facilitating extensions to Spanish-Chinese translation tasks.