Swedish University Course Classification Models and Corpus (SU Admin 2026)

This repository contains trained models and data for automatic classification of Swedish university courses into disciplinary domains (utbildningsområden, UO), developed as part of a master's project at Stockholm University.

Contents:

• bert_binary_model.zip  Fine-tuned KB-BERT model for multi-label classification (90.9% subset accuracy, 93.6% micro-F1)
• bert_distributional_model.zip  KB-BERT model for direct percentage distribution prediction (92.4% top-1 accuracy, 1.44 MAE)
• tfidf_baseline.zip  TF-IDF + Linear SVC baseline pipeline (87.9% subset accuracy)
• corpus_raw.csv  Original course plan corpus (9,770 course versions, 4,880 unique courses)
• corpus_preprocessed.csv  Cleaned and deduplicated training data with train/validation splits
• boglind_2026_su_admin_course_classification.pdf  Report

Data Source:
2023 extract from Ladok (Swedish national student information system) provided by Stockholm University Administration. Contains course titles, descriptions, learning objectives, and UO classifications. No personal data about students or instructors.

Models:
All transformer models use KB-BERT (KB/sentence-bert-swedish-cased). The distributional model was trained with KL-divergence loss to directly predict percentage allocations across 10 disciplinary domains.

Implementation:

https://github.com/fboglind/lis070-su-admin-project