Finnish ModernBERT Large MS MARCO

Model design and training

It maps text to 1,024-dimensional vectors for cosine similarity. Training uses about 11.66 million query, positive-passage and negative-passage examples with CachedMultipleNegativesRankingLoss. The derivative card does not state a model license. Its Finnish model name does not establish that this retrieval-tuning dataset is Finnish.finnish-modernbert-large-short-5e-05-msmarco: official model card

The parent is a 475-million-parameter ModernBERT encoder trained on 362.2 billion tokens. Its corpus includes Finnish, Swedish, English, code, Latin and Northern Sámi. The parent card has Apache 2.0 terms; that fact is kept separate from the retrieval derivative's missing license statement.Finnish ModernBERT Large Short: parent architecture and training

Practical assessment

Practical guidance. Assess this encoder on Finnish queries before adopting it for Finnish search. Its base and retrieval-training languages differ. Preserve the exact model name and similarity settings in the index metadata so a later encoder change triggers a deliberate re-embedding process.