chenjigaram/kid-extract-synthetic

Dataset card — kid-extract synthetic corpus

Summary

Generated European fund disclosure documents (PRIIPs KID and UCITS KIID) paired with exact structured ground truth, for training and evaluating extraction models.

Documents are synthetic. They are rendered by this repository, not collected from fund providers. No real disclosure document is redistributed here.

Why synthetic

No public corpus of KIDs exists. ESMA states that where a member state does not require KID notification to the national competent authority, no central European database of these documents exists. Building a real corpus would mean scraping several dozen provider websites.

Generating documents instead gives exact labels at any volume with no annotation cost and no teacher model. The ground truth is produced by the same code that renders the document, so a label can never disagree with the text it describes.

Fields

Fifteen scalar fields plus four performance scenarios, each with a value and an annualised return, giving 23 scored fields. Every field is optional; a field absent from the document is null in the label.

fund_name, isin, currency, sri, srri, ongoing_charges_pct, entry_charge_pct, exit_charge_pct, transaction_costs_pct, performance_fee_pct, recommended_holding_period_years, investment_objective, benchmark, domicile, management_company, scenarios.{stress,unfavourable,moderate,favourable}.{value,return_pct}

PRIIPs KIDs carry sri and a scenario table. UCITS KIIDs carry srri and neither.

Variation

Dimension	Values
Layouts	12
Languages	English, German, French, Dutch
Document types	PRIIPs KID, UCITS KIID
Cost presentation	table, bullets, prose
Risk presentation	scale drawing, n / 7, sentence
Number notation	decimal point, decimal comma, basis points
Label wordings	3 to 4 per field per language
Field dropout	per field, 5% to 35%

Documents also carry simulated PDF extraction artefacts: ligatures, hyphenation across line breaks, irregular column spacing, and page headers and footers landing mid-content. Lines carrying a labelled value are exempt from character corruption, so every label stays recoverable from the text.

Practical-information sections contain unrelated percentages — depositary fees, switching charges — so that taking the first number on the page is not a winning strategy.

Splits

Split	Layouts	Vocabulary	Purpose
train	8	known	training
validation	8	known	model selection
test_seen	8	known	in-distribution score
test_unseen_layout	4 held out	reserved	generalisation

The last label wording of every field is reserved and appears only in test_unseen_layout. That split therefore combines layouts and vocabulary the system has never seen, which is the position a deployed extractor is in when a new provider appears.

Funds do not repeat across splits; ISIN overlap between training and every evaluation split is zero and asserted in the test suite.

Known limitations

• Generated documents are more regular than real ones. Absolute scores are optimistic.
• Twelve layouts stand in for hundreds of real providers. Structural variety is understated even though label variety is modelled explicitly.
• Objective text is drawn from templates, so investment_objective is easier to copy than free prose in a real document.
• Scenario values are modelled from risk level rather than from real fund performance.

Grounding in real fund attributes

Passing --reference-csv with the Morningstar European funds dataset from Kaggle substitutes real fund names, ISINs, ongoing charges and risk ratings for the synthesised ones. Only the attributes that dataset lacks are generated. Document text remains synthetic either way.

Reproduction

python -m kidextract.cli.build_dataset --out data/processed --train 6000 --validation 500 --test 500

Generation is deterministic given the split seeds, which are fixed in kidextract.dataset.build.default_plans.

Licence

MIT, matching the repository. Contains no proprietary or confidential material.