toolathon123/eu-insurance-compliance

EU Insurance Compliance Audit Dataset (UK / FR / DE)

A fully synthetic, multilingual (English / French / German) dataset of insurance compliance review documents covering three European markets — United Kingdom, France and Germany — designed to support a small in-house financial-insurance compliance team (5–10 people) that must review more than 10,000 items per year (product documents, marketing & sales materials, regulatory filings and customer complaints).

The dataset supports text-classification / NLP-audit workflows such as fine-tuning classifiers to flag non-compliant documents before they reach the market, and can also be used for few-shot evaluation and RAG-style review tooling.

⚠️ Important: Synthetic & GDPR-safe data

ALL content in this dataset is fictional and synthetic. It contains no real natural persons, no real policyholders, no real companies, no real claims and no special-category data. Any personal-looking references are placeholder identifiers (e.g. CLM-2024-000123, GB-2025-0001234) or obviously fictional names. The dataset can therefore be shared openly and does not constitute processing of personal data under the GDPR / UK GDPR / BDSG / Loi Informatique et Libertés. It must not be used to infer anything about any real person.

Use this dataset for internal tooling, model development and demos. Do not treat the synthetic "violation" texts as legal advice; always route real compliance findings through qualified compliance/legal staff.

Dataset purpose

Each row is one document submitted to the annual compliance review, with a label used as the target for a text classifier:

• —Fields: id, country (UK/FR/DE), year (2020–2025), document_type, text, compliance_label.

document_type values

compliance_label values

Data statistics

• —Total rows: 12,000 (> 10,000 to support the annual compliance review volume).
• —Countries: UK 3,962 · FR 3,986 · DE 4,052 (roughly balanced).
• —Years: 2020–2025 (≈2,000 rows per year).
• —8 document types (≈1,400–1,600 rows each).
• —Label mix: ≈77% compliant, ≈23% non-compliant across the 6 violation categories — a realistic distribution for a mature review process.

Compliance context covered

The synthetic texts reference the regulatory frameworks the team works with:

• —UK: FCA (COBS/ICOBS), PRA, Financial Ombudsman Service, Consumer Duty, IDD, PRIIPs Regulation, UK GDPR.
• —France: ACPR, AMF, CNIL, Code des assurances, Directive sur la distribution d'assurances (DDA), RGPD, Code de la consommation.
• —Germany: BaFin, VAG, VersVermV, IDD, BGB, DSGVO.

How to use

from datasets import load_dataset

ds = load_dataset("project_20260813_012301_f820c644/eu-insurance-compliance", split="train")
print(ds)                      # Dataset({features: 6, num_rows: 12000})
print(ds.column_names)         # ['id','country','year','document_type','text','compliance_label']
print(ds[:5])                  # first 5 samples
print(ds.to_pandas().groupby("country")["compliance_label"].value_counts())

Intended uses & limitations

• —Intended: fine-tuning / evaluating multilingual text classifiers for insurance-compliance screening; internal demo & benchmarking of audit tooling.
• —Limitations: synthetic text; templates may not capture the full linguistic variety of real filings; labels are generated by construction, not by human legal review. Validate any model on a representative sample of real (anonymised) data before production use.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. See LICENSE.

Maintainer

Prepared by the EU insurance compliance NLP workstream. Synthetic data only — no personal data is processed.