From SALAMANDRA to SALAMANDRATA: BSC Submission for WMT25 General Machine Translation Shared Task

TL;DR

This paper introduces SALAMANDRATA, an improved multilingual translation model trained on European languages, with two sizes, achieving strong performance in WMT25 shared task through specialized training and decoding strategies.

Contribution

We present SALAMANDRATA, a novel family of multilingual translation models with enhanced training procedures and adaptation for European and non-European languages, submitted to WMT25.

Findings

Strong translation performance across 38 European languages.

Effective use of quality-aware decoding strategies.

Public release of models and SALAMANDRATA-V2.

Abstract

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 adapted the model vocabulary to support the additional non-European languages included in the task. This was followed by a second phase of continual pre-training 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 Tuned Re-ranking using COMET and COMET-KIWI respectively. We publicly release both the 2B and 7B versions of SALAMANDRATA, along with the newer SALAMANDRATA-V2 model, on Hugging Face1