FinBERT Finnish SQuAD 2.0
This release adapts FinBERT Base Cased for extractive question answering on Finnish SQuAD 2.0. It selects text spans rather than writing a new answer.
The trained task
Training used 384-token sequences and a document stride of 128. The BertForQuestionAnswering configuration has 512 position embeddings. The training length and architectural limit are different figures. The published weights include the trained question-answering head.FinBERT Finnish SQuAD 2.0: model card and trained artifactsFinBERT Finnish SQuAD 2.0: output labels and configuration
Practical assessment
Practical guidance. Supply the question and source passage together. Preserve passage offsets when splitting a long document, so the extracted span can be traced to the original text. An extracted sentence can still be irrelevant to the question.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | bert-base-finnish-cased-squad2 |
| Base model | FinBERT Base Cased.FinBERT Finnish SQuAD 2.0: model card and trained artifacts |
| Training task | Finnish SQuAD 2.0, including unanswerable questions.FinBERT Finnish SQuAD 2.0: model card and trained artifacts |
| Architecture | 12 layers and hidden width 768.FinBERT Finnish SQuAD 2.0: output labels and configuration |
| License | No explicit derivative weight license found in the model card or file list.FinBERT Finnish SQuAD 2.0: model card and trained artifacts |
Related models
- FinBERT Base Cased
TurkuNLP · FinBERT
FinBERT Base Cased is a Finnish masked-language encoder with a dedicated WordPiece vocabulary.
Launch date Downloadable - XLM-R Finnish QA Extractor
TurkuNLP · QA web mining
The Finnish QA Extractor labels question and answer spans within source documents.
Launch date not confirmedDownloadable