SalamandraTA 7B v2
SalamandraTA 7B v2 expands BSC's translation model beyond its original European-language coverage.
Model and training
The v2.0 card lists 40 languages and three varieties, adding Arabic, Hindi, Japanese, Korean and Simplified Chinese. It focuses on sentence, paragraph and document translation. BSC describes continued training with monolingual and parallel data before instruction tuning. This version remains available on its named branch under Apache 2.0.SalamandraTA 7B v2: official model card
Practical use
Practical guidance. This release is relevant when comparing the effect of added languages before v3's task and license changes. Preserve the repository revision in a deployment record. Include non-Latin scripts and long paragraphs in evaluation, and check whether translated passages preserve all source information rather than only sounding natural.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | salamandraTA-7b-instruct@v2.0 |
| License | Apache License 2.0.SalamandraTA 7B v2: official model card |
| Model size | 7B class.SalamandraTA 7B v2: official model card |
Related models
- SalamandraTA 7B v1
Barcelona Supercomputing Center · SalamandraTA
SalamandraTA 7B v1 is the first versioned 7B translation release based on BSC's multilingual Salamandra model.
Launch date not confirmedDownloadable - SalamandraTA 7B v3
Barcelona Supercomputing Center · SalamandraTA
SalamandraTA 7B v3 is BSC's expanded translation model with terminology and structured-document support.
Launch date not confirmedDownloadable