From SALAMANDRA to SALAMANDRATA: BSC Submission for WMT25 General Machine Translation Shared Task
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866908493068369920 |
|---|---|
| author | Gilabert, Javier Garcia Liao, Xixian Da Dalt, Severino Bohman, Ella Mash, Audrey Fornaciari, Francesca De Luca Baucells, Irene Llop, Joan Argote, Miguel Claramunt Escolano, Carlos Melero, Maite |
| author_facet | Gilabert, Javier Garcia Liao, Xixian Da Dalt, Severino Bohman, Ella Mash, Audrey Fornaciari, Francesca De Luca Baucells, Irene Llop, Joan Argote, Miguel Claramunt Escolano, Carlos Melero, Maite |
| contents | 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 |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2508_12774 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | From SALAMANDRA to SALAMANDRATA: BSC Submission for WMT25 General Machine Translation Shared Task Gilabert, Javier Garcia Liao, Xixian Da Dalt, Severino Bohman, Ella Mash, Audrey Fornaciari, Francesca De Luca Baucells, Irene Llop, Joan Argote, Miguel Claramunt Escolano, Carlos Melero, Maite Computation and Language 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 |
| title | From SALAMANDRA to SALAMANDRATA: BSC Submission for WMT25 General Machine Translation Shared Task |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2508.12774 |