Model Name | Layers / Units / Heads | Vocab. | Parameters | Legal |
---|---|---|---|---|
Maltehb/danish-bert-botxo |
12 / 768 / 12 | 32K | 110M | ❌ |
xlm-roberta-base |
12 / 768 / 12 | 256K | 278M | ❌ |
coastalcph/danish-legal-lm-base |
12 / 768 / 12 | 32K | 110M | ✅ |
coastalcph/danish-legal-bert-base |
12 / 768 / 12 | 32K | 110M | ✅ |
coastalcph/danish-legal-longformer-base |
12 / 768 / 12 | 32K | 134M | ✅ |
coastalcph/danish-legal-xlm-base |
12 / 768 / 12 | 32K | 110M | ✅ |
This model is pre-trained on a combination of the Danish part of the MultiEURLEX (Chalkidis et al., 2021) dataset comprising 65k EU laws and two subsets (retsinformationdk
, retspraksis
) of the Danish Gigaword Corpus (Derczynski et al., 2021) comprising legal proceedings. It achieves the following results on the evaluation set.
Model Name | Loss | Accuracy |
---|---|---|
Maltehb/danish-bert-botxo |
22.3 | 7.038 |
coastalcph/danish-legal-lm-base |
84.8 | 0.651 |
coastalcph/danish-legal-bert-base |
80.1 | 0.878 |
coastalcph/danish-legal-bert-base |
82.5 | 0.768 |
coastalcph/danish-legal-xlm-base |
83.1 | 0.727 |
Model Name | EURLEX Val. | EURLEX Test |
---|---|---|
Maltehb/danish-bert-botxo |
73.7 / 42.8 | 67.6 / 38.2 |
coastalcph/danish-legal-lm-base |
75.1 / 46.5 | 69.1 / 41.9 |
coastalcph/danish-legal-bert-base |
75.0 / 50.4 | 68.9 / 44.3 |
coastalcph/danish-legal-xlm-base |
TBA | TBA |
coastalcph/danish-legal-longformer-base |
75.7 / 52.9 | 69.6 / 47.0 |
coastalcph/danish-legal-longformer-base + SD Penalty (Pezeshki et al., 2020) |
76.1 / 52.9 | 69.9 / 47.0 |
The top-2 best models (coastalcph/danish-legal-longformer-base
, coastalcph/danish-legal-longformer-base-sd
) are available on HuggingFace Hub with instructions on how can be used as text classifier or feature extractor.
sh train_mlm_gpu.sh
export PYTHONPATH=.
python src/mod_teacher_model.py --teacher_model_path coastalcph/danish-legal-lm-base --student_model_path coastalcph/danish-legal-lm-base
export PYTHONPATH=.
python src/longformerize_model.py --roberta_model_path coastalcph/danish-legal-lm-base --max_length 2048 --attention_window 128