Dataset Card for VelumBench

VelumBench is a test set for anonymising Swiss court decisions in German, French and Italian. Each case is a published decision in which the court's placeholders (A.________) have been filled with invented Swiss identities, so a system sees text that reads like a decision before anonymisation, and the inserted spans are the exact answer key for what it must remove. Names the courts publish on purpose (judges, clerks, lawyers) are the answer key for what must stay.

It was built for Velum, a Hack Apertus 2026 Track 2B project that anonymises decisions with Apertus and cites the rule behind every edit. The runs/ folder holds the raw outputs of every benchmark run reported in the technical report.

Dataset Details

• Curated by: Calder Luo
• Shared by: Calder Luo, for the Online Hack Apertus 2026 (Track 2B)
• Language(s) (NLP): German, French, Italian
• License: CDLA-Permissive-2.0 for the invented identities, the annotations and the split. Decision texts are official decisions and are not protected by copyright (Art. 5 para. 1 let. c URG, SR 231.1).
• Git repository: https://github.com/calderbuild/velum (generator in track_2b/src/velum/bench.py)

	dev	test
Decisions	79 (de 18, fr 36, it 25)	537 (de 198, fr 180, it 159)
Courts	BGer 56, BVGer 9, GE 8, VD 2, TI 2, ZH 1, BE 1	BGer 356, BVGer 80, ZH Obergericht 24, BE 24, VD 23, GE 17, GR 9, TI 4
Gold mentions to remove	1,459 (person 1,187, company 187, address 52, place 33)	8,390 (person 6,713, company 951, address 374, place 352)
Distinct gold entities	376	2,471
Published names that must stay	450 mentions (298 names)	3,168 mentions (2,072 names)
Decision dates	2008 to 2026	2008 to 2026
Median length	13,685 characters	13,088 characters

Uses

Direct Use

Scoring systems that anonymise Swiss court decisions: how many identities they remove (recall), how many decisions come out with nothing leaked, how many published names they wrongly remove (over-redaction), whether one party gets one placeholder throughout (consistency), and how much text they change outside the names. The scorer is velum bench in the Git repository; it needs no model and no LLM judge.

Out-of-Scope Use

• Training or evaluating re-identification. The inserted identities are invented and say nothing about the real parties.
• Using the invented names, companies or addresses as facts about real people or businesses.
• Treating the gold as a full anonymisation standard. It covers only what the court itself replaced with a placeholder (see Risks, Bias, and Limitations).

Specifications

• Source rows: voilaj/swiss-caselaw (CC0-1.0) at revision 90d5cd212466d4286ed574f825ae5b3df8cb3645.
• Stratified sample of 750 decisions with seed 20261001: Federal Supreme Court 420, Federal Administrative Court 90, Federal Criminal Court 90, six cantonal courts 150. 616 of them yielded at least one gold mention.
• Split: sha256(decision_id) mod 7 == 0 goes to dev, the rest to test.

Dataset Structure

Component	Files
Evaluation dataset (test cases)	data/dev.jsonl, data/test.jsonl
Model responses	runs/*.jsonl.gz, one per benchmark run, with summaries
Metadata with instance-level licences	metadata/metadata.jsonl: one line per case with split, court, language, docket, date, source URL, gold and keep counts, license and provenance

One JSON object per line in data/dev.jsonl and data/test.jsonl.

Field	Meaning
id, court, language, docket, decision_date, legal_area, source_url	From the source row
text	The decision with invented identities in place of the placeholders: the system input
gold	{start, end, text, entity, kind} per inserted mention; kind is person, company, place or address; entity groups the mentions of one party (A, A#address)
keep	{text, kind, count} for each published name that must stay (court_official or lawyer)
published_text	The decision exactly as the court published it
license	Instance-level licence note
provenance	{hf_dataset, revision, file, row_group, file_row_number, content_hash}, enough to re-fetch the source row

runs/ holds one gzipped JSONL file per benchmark run on the test split (1 October 2026, CSCS endpoint, temperature 0, unless noted) and its summary:

Run	Decisions	Recall	Over-redaction	Changes outside the names
rewrite_Apertus-v1.5-70B (per-language-20): Apertus 70B rewrites the whole decision	60	70.8%	4.3%	215
velum_Apertus-v1.5-70B, scored on the same 60 (per-language-20)	60	96.7%	2.8%	0
velum_Apertus-v1.5-70B	537	89.4%	1.2%	0
velum_Apertus-v1.5-8B	537	76.4%	0.5%	0
velum_Apertus-v1.5-8B, scored on the same 60 (per-language-20)	60	92.6%	3.7%	0
velum_Apertus-v1.5-8B-Q4_K_M-llamacpp (per-language-20): the air-gapped setup, 8B Q4_K_M in llama.cpp on CPU, added 4 October 2026	60	86.0%	0.3%	0
patterns_none: regular expressions only	537	4.6%	0.0%	0

Each line of a run file is one decision: the output text, every edit with the article that requires it, the risks the audit raised, the score and the token usage. Every summary recomputes from its run file without a model call (track_2b/results/README.md in the Git repository).

Dataset Creation

Curation Rationale

Swiss courts must publish their decisions in anonymised form, and each court has precise rules for what to replace and what to keep. Published decisions show only the result, so there was no open, exact test set for this task in German, French and Italian. VelumBench turns the courts' own placeholders into one.

Source Data

Decisions of the Federal Supreme Court, the Federal Administrative Court and six cantonal courts as published by the courts and collected in voilaj/swiss-caselaw.

Data Collection

I read only the row groups needed for the sample from the pinned revision of the source dataset and kept the source identifiers in provenance.

Source Data Producers and Data Subjects

The courts wrote and anonymised the decisions. The parties appear only as the courts' placeholders, which VelumBench replaces with invented identities. Judges, clerks and lawyers appear as the courts published them; Swiss courts publish these names on purpose.

Ownership and Consent Management

Official decisions are not protected by copyright in Switzerland (Art. 5 para. 1 let. c URG), and the source collection is CC0-1.0. The courts' publications stay authoritative. If a court withdraws or re-anonymises a decision, I will remove the case; open an issue in the Git repository.

Data Processing and Quality Control

python -m velum build finds the placeholders, decides from the surrounding words whether a letter stands for a person, a company or a place, and inserts an invented identity from per-language pools, the same one for every occurrence of the letter. When a party in the header is followed by an address line, it also inserts an invented street address. No word of an inserted name occurs in the source decision. In all 616 cases, each of the 9,849 gold spans points at exactly its text, and a unit test in the repository generates a case and checks its kinds, its keep list and its scoring.

Data Instance Licenses

Decision texts: official decisions, not protected by copyright. Invented identities, annotations, split and run outputs: CDLA-Permissive-2.0.

Annotations

Annotation Process

The gold is generated, not hand-labelled: every inserted span is recorded with its offsets as it is written. The keep list is read from the composition block and from "vertreten durch / représenté par / patrocinato da".

Annotator Details

No human annotators. The generator is deterministic code in the Git repository. No language model was used to build the gold.

Risks, Bias, and Limitations

• The gold is only as complete as the court's own anonymisation: if a court left a name in the published text, VelumBench does not count it as a leak, and over-redaction is scored only on judges, clerks and lawyers.
• Only "letter + period + underscores" placeholders are recognised, so the Federal Criminal Court (bare A.) and most Ticino decisions are missing, and Romansh is not covered.
• Courts that anonymise lawyers keep their lawyer placeholders, which are not scored.
• Addresses are inserted only for the first party in the header.
• The kind of a placeholder (person, company, place) comes from the surrounding words, so a few companies are scored as persons and the reverse. Recall does not depend on the kind.
• Quasi-identifiers (occupations, small places with dates, unusual events) have no gold.

Personally Identifiable Information

No real party is re-identified. Every person, company and place in gold is invented, and none of its words occurs in the source decision. Addresses combine common Swiss street names with real postcodes and towns. The source dataset warns that some cantonal decisions may still contain names the court missed; VelumBench adds none, but keeps whatever the court published.

Sensitive Information

Court decisions can describe health, criminal proceedings, migration status or family matters of the parties. Those passages are as the courts published them, about parties the courts anonymised.

Recommendations

Report results per language and per court, and read recall together with over-redaction and the changes outside the names: a system can reach high recall by removing too much.

Citations

APA:

Luo, C. (2026). VelumBench: an open test set for anonymising Swiss court decisions [Data set]. Hack Apertus 2026. https://github.com/calderbuild/velum

Dataset Card Authors

Calder Luo

Dataset Card Contact

Open an issue at https://github.com/calderbuild/velum