rinnaBaku
Gemma 2 Baku 2B
Baku 2B continues pretraining Google's Gemma 2 2B on about 80B Japanese and English tokens.
Research release
A small base receives substantial language adaptation
rinna uses a LitGPT-based training implementation. The card warns of NaN values with padded input, bfloat16 and the default scaled-dot-product attention path in its documented software setup.Gemma 2 Baku 2B official model card
Check numerical behavior in the selected runtime
Practical guidance. A compact model can be attractive for local adaptation. Inspect padded batches as well as single inputs when evaluating a runtime; a successful single-prompt call does not cover the documented numerical issue.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | gemma-2-baku-2b |
| Architecture | 26 layers with hidden width 2,304.Gemma 2 Baku 2B official model card |
| License | Gemma terms.Gemma 2 Baku 2B official model card |
| Ancestry | google/gemma-2-2b.Gemma 2 Baku 2B official model card |
Related models
- Gemma 2 Baku 2B Instruct
rinna ยท Baku
Baku 2B Instruct adds Google's Gemma instruction chat vector to rinna's Japanese-adapted base, then applies odds-ratio preference optimization.
Launch date not confirmedDownloadable