SLIDE

Data and code (still being added) for the paper Multi-label Scandinavian Language Identification (SLIDE) (presented at RESOURCEFUL-2025).

Models on HuggingFace and how to use them

reproduce metrics (table 4):

cd src/
./run_all.sh

reproduce evaluation on nordic_langid (table 5)

obtain data from nordic_langid on Huggingface and put *test.csv into src/evaluation

cd src/
python3 nordic_langid2jsonl.py 
python3 evaluate.py --method bert --model ltg/SLIDE-base --dataset nordic_dsl_test50k.jsonl
python3 evaluate.py --method bert --model ltg/SLIDE-base --dataset nordic_dsl_test10k.jsonl

The values that will be shown will be different from those in table 5 in the paper.

The values in table 5 were obtained with the understanding of loose accuracy as it is described in the paper.

The actual evaluate.py accepts a prediction if it is a subset of gold languages, not an intersection. (Values in table 4 were obtained with this understanding). However, while it influences exact values (less than 2%), the models' ranking remains the same.