README.md

---
widget:
- text: "KOMMISSIONENS BESLUTNING\naf 6. marts 2006\nom klassificering af visse byggevarers ydeevne med hensyn til reaktion ved brand for så vidt angår trægulve samt vægpaneler og vægbeklædning i massivt træ\n(meddelt under nummer K(2006) 655"
datasets:
- multi_eurlex
metrics:
- f1
model-index:
- name: coastalcph/danish-legal-longformer-eurlex-sd
  results:
  - task:
      type: text-classification
      name: Danish EURLEX (Level 2)
    dataset:
      name: multi_eurlex
      type: multi_eurlex
      config: multi_eurlex
      split: validation
    metrics:
    - name: Micro-F1
      type: micro-f1
      value: 0.76144
    - name: Macro-F1
      type: macro-f1
      value: 0.52878
---

# Model description

This model is a fine-tuned version of [coastalcph/danish-legal-longformer-base](https://huggingface.co/coastalcph/danish-legal-longformer-base) on the Danish part of [MultiEURLEX](https://huggingface.co/datasets/multi_eurlex) dataset using an additional Spectral Decoupling penalty ([Pezeshki et al., 2020](https://arxiv.org/abs/2011.09468).

## Training and evaluation data

The Danish part of [MultiEURLEX](https://huggingface.co/datasets/multi_eurlex) dataset.

## Use of Model

### As a text classifier:

```python
from transformers import pipeline
import numpy as np

# Init text classification pipeline
text_cls_pipe = pipeline(task="text-classification",
                         model="coastalcph/danish-legal-longformer-eurlex-sd",
                         use_auth_token='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')

# Encode and Classify document
predictions = text_cls_pipe("KOMMISSIONENS BESLUTNING\naf 6. marts 2006\nom klassificering af visse byggevarers "
                            "ydeevne med hensyn til reaktion ved brand for så vidt angår trægulve samt vægpaneler "
                            "og vægbeklædning i massivt træ\n(meddelt under nummer K(2006) 655")

# Print prediction
print(predictions)
# [{'label': 'building and public works', 'score': 0.9626012444496155}]
```

### As a feature extractor (document embedder):

```python
from transformers import pipeline
import numpy as np

# Init feature extraction pipeline
feature_extraction_pipe = pipeline(task="feature-extraction",
                                   model="coastalcph/danish-legal-longformer-eurlex-sd",
                                   use_auth_token='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')

# Encode document
predictions = feature_extraction_pipe("KOMMISSIONENS BESLUTNING\naf 6. marts 2006\nom klassificering af visse byggevarers "
                                      "ydeevne med hensyn til reaktion ved brand for så vidt angår trægulve samt vægpaneler "
                                      "og vægbeklædning i massivt træ\n(meddelt under nummer K(2006) 655")

# Use CLS token representation as document embedding
document_features = token_wise_features[0][0]

print(document_features.shape)
# (768,)
```

## Framework versions

- Transformers 4.18.0
- Pytorch 1.12.0+cu113
- Datasets 2.0.0
- Tokenizers 0.12.1