Finnish Sentence-BERT Uncased
This Finnish sentence embedding model adapts the uncased FinBERT checkpoint for paraphrase similarity. It returns sentence representations rather than generating text.
Model design and training
Training uses the Finnish Paraphrase Corpus and automatically collected candidates, including 500,000 positive and five million negative pairs. Mean pooling combines token representations. The model learns whether two sentences are paraphrases. Its card does not publish an explicit model license; the reviewed repository contains no license file.sbert-uncased-finnish-paraphrase: official model card
Practical assessment
Practical guidance. The uncased tokenizer can reduce differences caused by inconsistent capitalization. Compare it on names, abbreviations and informal Finnish before choosing it for search. Do not mix its vectors with a cased model's index, even though both emit vectors of the same size.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | sbert-uncased-finnish-paraphrase |
| Context window | 128 tokens Default SentenceTransformer maximum sequence length; longer text is truncated under these packaged settings. sbert-uncased-finnish-paraphrase: sentence encoder settings |
| License | Not specified in the reviewed derivative card or repository files.sbert-uncased-finnish-paraphrase: official model card |
| Base model | TurkuNLP/bert-base-finnish-uncased-v1.sbert-uncased-finnish-paraphrase: official model card |
| Pooling | Mean pooling over token embeddings.sbert-uncased-finnish-paraphrase: official model card |
| Case handling | do_lower_case is true.sbert-uncased-finnish-paraphrase: sentence encoder settings |
Related models
- Finnish Sentence-BERT Cased
TurkuNLP ยท Finnish Sentence-BERT
This Finnish sentence embedding model adapts the cased FinBERT checkpoint for paraphrase similarity.
Launch date not confirmedDownloadable