DFM Mimir v1 Release
Danish Foundation Models has released Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture and trained from scratch on permissi...
OdenseNLP
OdenseNLP
Safe, Efficient and Open Natural Language Processing @ University of Southern Denmark
Latest news
See all posts
Extensive open training-data collection on Hugging Face
We published an extensive mix of data with different objectives, ranging from English and Danish instruction and knowledge to mathematics and agentic-style post-training data.
FlexMoRE - Efficient modular language models
The Allen Institute for AI (Ai2) recently highlighted FlexMoRE, a new approach to building more efficient modular language models developed by researchers at OdenseNLP and collaborat...
OdenseNLP at LREC26
OdenseNLP was at LREC 2026 and authored/contributed in three papers:
Current focuses
Low-resource NLP
Language technologies for low-resource langauges, particularly Danish and neighboring Scandinavian languages.
Efficient NLP
Fast and efficient NLP architectures and methods.
AI Safety & Interpretability
Making AI systems more safe, trustworthy, and interpretable.
Data
Browse all dataDatasets
Post-Training Dataset Collection
An extensive collection for language-model training with objectives spanning English and Danish instruction and knowledge, mathematics, and agentic...
Danish Dynaword
A continually expanded corpus of openly licensed Danish free-form text from diverse domains. Version 1.2.20 contains 7.25 million documents and 6.9...
Benchmarks
DaLA
Danish linguistic acceptability benchmark with corrupted and non-corrupted sentences, published as ~8.68k examples with train/validation/test and f...
SDU-Daisy
Danish-culture benchmark based on the Danish Culture Canon, with 746 closed question-answer pairs for evaluating LLM cultural understanding.
GEC DaLA
Grammatical error correction version of DaLA, pairing original and corrupted Danish sentences with corruption types and affected-token annotations....
Models
See all modelsDFM Mimir
An instruction-tuned Danish-English foundation model based on HRM-Text, trained exclusively on permissible post-training data and designed for stro...
Danish Foundation Models
The complete, growing collection of Danish Foundation Models releases on Hugging Face, spanning text-generation and multimodal models, Danish-focus...
Qwen3.5-27B-psysafe
A supervised fine-tune of Qwen3.5-27B for psychologically informed refusals, providing structured, supportive responses to high-risk requests while...
Repositories
See all repositoriesPsychoSafe
Pipeline for building and evaluating psychologically informed refusal behavior in large language models through data creation, prompting, fine-tuni...
PropMe
Framework for evaluating memorization and propensity-aware memorization of training data in large language models.
DaLA
Source code for creating the Danish Corpus of Linguistic Acceptability, designed to evaluate Danish linguistic acceptability with real-world errors.
