FLOR 760M
FLOR 760M is BSC's compact language adaptation of BigScience BLOOM 1.1B for Catalan, Spanish and English.
Research release
Model and training
BSC trained a target-language tokenizer and reduced the embedding vocabulary, then continued pretraining on 26 billion tokens. Matching token embeddings were reused; new entries received an average-based initialisation. The smaller parameter count therefore comes partly from vocabulary changes, not only fewer transformer layers. The card marks the release as deprecated and no longer maintained, although weights remain published.FLOR 760M: official model card
Practical use
Practical guidance. This model provides a small historical baseline for language adaptation. It can help separate the effect of vocabulary design from later increases in model size. Use it for reproducible research with its matching tokenizer. Its discontinued support and base-model training should remain clear when comparing it with current assistants.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | FLOR-760M |
| License | Apache License 2.0.FLOR 760M: official model card |
| Model size | 760 million parameter class.FLOR 760M: official model card |
Related models
FLOR 1.3B adapts BigScience BLOOM 1.7B to Catalan, Spanish and English through vocabulary replacement and continued pretraining.
Launch date not confirmedRestricted accessDownloadableFLOR 6.3B is BSC's larger BLOOM-derived language model for Catalan, Spanish and English.
Launch date not confirmedRestricted accessDownloadable