Nor-CaseHOLD: A Retrieval Benchmark for Norwegian Legal AI

Nor-CaseHOLD is an extractive legal retrieval benchmark built from 1,244 Norwegian legal documents — 627 Supreme Court (Høyesterett) decisions and 617 Skatteetaten bindende forhåndsuttalelser (BFU) — each paired with its official summary.

To the author's knowledge, this is the first open-source retrieval benchmark for Norwegian legal text.

Task

Given a full legal document, select the 5 sentences that best match the official summary. Extracted sentences are scored against the summary using ROUGE-1, ROUGE-2, and ROUGE-L.

Dataset

Split	Høyesterett	BFU	Total
Train	438	431	869
Validation	72	70	142
Test	117	116	233
Total	627	617	1,244

Splits are stratified by source for balanced representation.

Benchmark Results (v2)

Test set: n=233, top-5 extraction, 95% bootstrap CIs (10,000 samples).

Method	Type	ROUGE-1	95% CI
Oracle (sequential)	Oracle	55.87	[54.67, 57.12]
TF-IDF	Sparse	47.85	[46.50, 49.23]
BM25	Sparse	47.49	[46.16, 48.82]
Oracle (greedy)	Oracle	42.68	[41.52, 43.85]
Norwegian Legal BERT	Dense	38.40	[37.19, 39.64]
MiniLM	Dense (ST)	37.47	[36.28, 38.73]
mBERT	Dense	37.34	[36.28, 38.41]
NB-BERT-base	Dense	37.28	[36.18, 38.35]
Lead-5	Structural	25.07	[23.93, 26.21]

Pairwise Significance Tests (ROUGE-1)

Comparison	Δ ROUGE-1	p-value	Sig.
BM25 vs TF-IDF	-0.36	0.971	n.s.
Norwegian Legal BERT vs NB-BERT-base	+1.12	< 0.001	***
Norwegian Legal BERT vs mBERT	+1.06	0.001	**
Norwegian Legal BERT vs MiniLM	+0.92	0.055	n.s.

Sample Evaluation Output

Loading model: bendik-eeg-henriksen/norwegian-legal-bert
Evaluating 233 documents...
  20/233...
  ...
  220/233...

Results (top-5, n=233):
  ROUGE-1: 38.40
  ROUGE-2: 15.97
  ROUGE-L: 20.93

Usage

from datasets import load_dataset

dataset = load_dataset("bendik-eeg-henriksen/nor-casehold")
test = dataset["test"]

for example in test:
    print(example["source"])        # 'hoyesterett' or 'skatteetaten_bfu'
    print(example["sammendrag"])     # Official summary (reference)
    print(example["full_text"])      # Full document text
    print(len(example["sentences"])) # Pre-split sentences with scores

Acknowledgments

The NOR-CASEHOLD benchmark design was inspired by ITA-CASEHOLD (Licari et al., ICAIL 2023), who adapted CaseHOLD (Zheng et al., 2021) for civil law jurisdictions. Norwegian Legal BERT builds on NbAiLab/nb-bert-base.

Related

• Model: Norwegian Legal BERT
• Code: GitHub
• Technical paper SSRN: Nor-CaseHOLD: A Retrieval Benchmark for Norwegian Legal AI

Version History

• v2 (March 2026): Expanded to 1,244 documents (added 617 Skatteetaten BFU). Stratified splits. New baselines including hybrid and cross-encoder experiments.
• v1 (March 2026): Initial release with 627 Høyesterett decisions.

Citation

if using this work please consider citing

@misc{eeg-henriksen-2026-nor-casehold,
    title        = {NOR-CASEHOLD: A Retrieval Benchmark for Norwegian Legal AI},
    author       = {Eeg-Henriksen, Bendik},
    year         = {2026},
    howpublished = {HuggingFace Datasets},
    url          = {https://huggingface.co/datasets/bendik-eeg-henriksen/nor-casehold}
}