TurkuNLP
TurkuNLP is an interdisciplinary language research group at the University of Turku. Its researchers develop Finnish language technology, large language models, text representations and methods for analysing large text collections. It is a university group, not a separate model-hosting company.
Company & leadership
- Institutional home
- University of Turku and its graduate school, Finland.
- Computer science research lead
- Professor Filip Ginter, identified as group leader in the HAIF programme.
- Digital linguistics research lead
- Professor Veronika Laippala, identified as group leader in the HAIF programme.
- FinGPT collaboration
- Developed with the National Library of Finland, with computing support from CSC.
An interdisciplinary university group
The group combines computer science and linguistics. Its work includes Finnish syntax and semantics, biomedical text analysis, language models, and the registers and genres found online. The public profile places researchers within the university and graduate school. It does not state a separate legal incorporation or formal group founding date.
The university's HAIF programme identifies separate computer science and digital linguistics research leads. Ginter's area covers computational NLP and large-scale model training. Laippala's area studies language use and large digital corpora. These are research leadership roles, not corporate chief-executive titles.
From language resources to generative models
The Finnish Paraphrase project developed manually classified pairs designed to vary in wording and structure. The work was supported by European Language Grid and Academy of Finland funding. Its releases include the corpus, candidate pairs and Finnish Sentence-BERT weights. These resources connect language annotation with representation learning.
The 2023 FinGPT study compared two model-development routes. It trained seven Finnish-only models from scratch and continued pretraining BLOOM to create BLUUMI. The authors also introduced FIN-bench for Finnish evaluation. The project released weights across several sizes, giving later studies a shared set of Finnish baselines.
Collaborative model development
In August 2023, the University of Turku and SiloGen announced a wider European-language model effort. The announcement named Ginter and Sampo Pyysalo as the university's project leads and described the use of HPLT language data and LUMI computing. Poro and Viking are covered once under the joint LumiOpen profile.
The catalog includes generative models, pretrained encoders and trained text classifiers. The WikiBERT article lists separate language weights together. The register-ablation article records the exact experiment branches. Continued-training checkpoints, file formats and pipeline components are identified within their related articles.
Encoder releases and training stages
Finnish ModernBERT uses six architecture sizes. Short releases share a 128,000-entry vocabulary; the other designs use vocabularies selected for their size. The paper separates context extension from the last training stage. EDU releases use a more balanced final data mixture, while the baseline mixture is mostly English. The CPT checkpoints precede learning-rate decay and support further training.