Chinorab/beneficia-cases

Dataset Card for Beneficia evaluation set

The set has 14 fictitious company folders (20 documents in German, French and Italian). Each folder has the expected beneficial owners under the Swiss transparency law (TJPG, SR 955.3; TJPV, AS 2026 364), and the dataset also includes the outputs of Apertus v1.5 8B on every case. It was built to evaluate Beneficia, an offline assistant that drafts beneficial owner declarations for fiduciaries (Hack Apertus 2026, track 2B).

Dataset Details

The cases cover:

• direct and indirect holdings, including a chain with two intermediaries;

• the "multiplication trap": 40 % of a 100 % holding is not control;

• a voting agreement and fiduciary holdings;

• nobody above 25 %, so the chair of the board applies;

• inconsistent figures and an undocumented foreign holder;

• a DOCX register, a percentage written in prose, a wrapped sentence and a degraded scan.

• Curated by: Anas (team Beneficia)

• Funded by [optional]: none

• Shared by [optional]: Anas (Hugging Face: Chinorab)

• Language(s) (NLP): German, French, Italian

• License: CC BY 4.0, also stated for every instance in metadata.jsonl

• Git repository: https://github.com/Chinorab/beneficia (code under Apache 2.0, track_2b/data/cases holds the source PDF, DOCX, TXT and JPG files)

Uses

Direct Use

Use it to evaluate structured extraction from Swiss corporate documents (share registers, shareholder agreements, board extracts), and the application of the TJPG/TJPV beneficial owner rules: direct holdings, indirect control through intermediaries, control by other means, and the default manager rule.

Out-of-Scope Use

• It is not a source of legal advice and is not a statement of how the law applies to a real company.
• It is not representative of real registers in volume or messiness.
• It is too small to train models.

Specifications

Each case is one folder per target company. The expected answer gives:

• the holdings, as holder, company and units or written percentage;
• the beneficial owners, as name, basis (direct, indirect, other_means, default_manager), band (25_50, 50_75, 75_100) and status (settled or pending_review);
• the review reasons that must be raised.

Dataset Structure

• evaluation.jsonl: one line per case.
  • case_id, target.
  • documents: id, filename, kind, language, and pages with numbered lines. This is the exact text the system read; the scan was read by Tesseract inside the container.
  • expected.
• responses.jsonl: one line per case.
  • model, and raw_extraction_outputs: the JSON returned by Apertus for every pass and window.
  • findings, review, missing, holdings, dropped_items (facts rejected because they were not on the cited line), seconds, and score (deterministic comparison with expected).
• metadata.jsonl: one line per case, with case_id, language, fictitious, licence, source, personal_data, notes (what the case tests) and files.

There is a single split. The 14 cases served as both the development and the evaluation set (see Limitations).

Dataset Creation

Curation Rationale

Since 1 October 2026, Swiss companies must report their beneficial owners to the federal transparency register. No public test set existed for checking an assistant on this task in German, French and Italian. Real registers are confidential, so the set had to be synthetic.

Source Data

The source is synthetic company documents: share registers, a quota register, a shareholder agreement, board extracts, a fiduciary note and a scanned register.

Data Collection

The documents were written by the curator and generated with track_2b/data/tools/make_cases.py (reportlab for PDF, python-docx for DOCX, Pillow for the degraded scan, which has blur, rotation and noise). Every document carries a "fictitious" banner in its language.

Source Data Producers and Data Subjects

The curator produced all of it. There are no data subjects: every company, person, address, date and figure is invented.

Ownership and Consent Management

The curator owns all content. No third-party data was used, so no consent was needed.

Data Processing and Quality Control

Each expected.json was written by hand from the law (TJPV art. 1, 2, 3, 4, 13; TJPG art. 4) before the system was run. The model outputs come from the final pipeline running in a Docker container with no network interface, with Apertus v1.5 8B Q4_K_M GGUF on llama.cpp b10548 and temperature 0. The scores are computed deterministically by beneficia.evaluate.

Data Instance Licenses

All instances are under CC BY 4.0, as stated in metadata.jsonl.

Annotations

The annotations are the expected answers in expected (holdings, beneficial owners, review reasons).

Annotation Process

A single annotator applied the TJPV rules to each fictitious folder. Indirect control follows TJPV art. 2: more than 50 % of an intermediate entity that itself holds at least 25 %. Percentages are not multiplied.

Annotator Details

The developer of Beneficia, with no formal legal qualification. There is a single annotator, so no inter-annotator agreement is available.

Risks, Bias, and Limitations

• Small and synthetic. 14 cases and 20 short documents.
• Not held out. The pipeline was tuned on these same cases, so the reported score (14/14 cases fully correct, 0 silent errors, holdings precision and recall 98 %) is not a held-out measure.
• Narrow legal scope. The legal reading is the annotator's. It covers companies under TJPG art. 4 only. SICAV, trusts, foundations and associations are out of scope.
• Simple formats. Real documents have longer tables, handwriting, several share classes and inconsistent spelling.

Personally Identifiable Information

None. All names, companies and addresses are invented for this project.

Sensitive Information

None.

Recommendations

• Use the set as a smoke test and a regression test, not as evidence of legal accuracy.
• Add real anonymised folders, with consent, before relying on a system built on it.

Citations [optional]

BibTeX:

@misc{beneficia2026,
  title  = {Beneficia evaluation set: fictitious Swiss beneficial owner cases},
  author = {Anas},
  year   = {2026},
  note   = {Hack Apertus 2026, track 2B},
  url    = {https://github.com/Chinorab/beneficia}
}

APA:

Anas. (2026). Beneficia evaluation set: fictitious Swiss beneficial owner cases [Data set]. Hack Apertus 2026.

Glossary [optional]

• Beneficial owner (wirtschaftlich berechtigte Person, ayant droit économique): a natural person who ultimately controls a company. Control means at least 25 % of capital or votes, directly or indirectly, or control by other means (TJPG art. 4).
• Band: the extent of the holding that the register asks for: at least 25 and at most 50, over 50 and at most 75, or over 75 percent (TJPV art. 13).
• Silent error: a wrong beneficial owner marked as settled instead of "to be reviewed".

More Information [optional]

See track_2b/docs/evaluation.md and track_2b/technical_report.md in the Git repository.

Dataset Card Authors [optional]

Anas

Dataset Card Contact

Through the GitHub repository issues: https://github.com/Chinorab/beneficia/issues