Plume 32K
Plume 32K is BSC's translation research model with a 32,000-token vocabulary.
Research release
Model and training
This version uses the smallest vocabulary in the Plume study. BSC trains the Gemma 2B architecture from scratch on parallel text centred on Catalan. Google Gemma weights are not the starting point. The research covers 16 supervised translation directions and 56 directions without direct training examples. The 32k suffix is vocabulary size, not context length.Plume 32K: official model card
Practical use
Practical guidance. Use this model to investigate tokenisation in sentence translation. A smaller vocabulary changes how a sentence is split into tokens, which affects sequence lengths and embedding storage. Compare all three Plume releases on identical language pairs and preserve their different tokenizers.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | Plume32k |
| Context window | 8,192 tokens Configured maximum positions; the documented training task is sentence translation, not long-document translation. Plume32k: released model configuration |
| License | Apache License 2.0.Plume 32K: official model card |
| Model size | Gemma 2B architecture; total size varies with the vocabulary.Plume 32K: official model card |
| Architecture configuration | 18 decoder layers, 2,048 hidden width and 32,000 vocabulary entries.Plume32k: released model configuration |
Related models
Plume 128K is BSC's translation research model with a 128,000-token vocabulary.
Launch date not confirmedResearch releaseDownloadablePlume 256K is BSC's translation research model with a 256,000-token vocabulary.
Launch date not confirmedResearch releaseDownloadable