andreagemelli/xfund-kie-it

xfund-kie — Italian XFUND → KIE (Key Information Extraction)

Original source & attribution

This dataset is derived from XFUND (https://github.com/doc-analysis/XFUND), the multilingual form-understanding dataset. The Italian (it) split — images (it.train/, it.val/) and underlying annotation files — was translated into a Key Information Extraction (KIE) task. The conversion pipeline and quality refinement were driven by Claude (Anthropic) in collaboration with manual analysis.

• Upstream dataset: XFUND (Document Analysis / Form Understanding)
• Upstream repo: https://github.com/doc-analysis/XFUND
• Language split used: Italian (it.train, it.val)
• Dataset folder: xfund-kie/

Methodology

1. Source extraction: Starting from 4882 source question→answer links from the Italian XFUND split.
2. Label-to-key mapping: Extracted field labels (e.g., Codice Fiscale → codice-fiscale) and mapped answers to structured keys. 2757 links (56%) mapped to a key; the rest are instruction paragraphs, list markers (il, n.), or del labels.
3. Quality pipeline (automated + manual audit):
   • Dropped one-label, several-answers: 96 cases (table columns, option lists) where a single label spans multiple answers — not representable as a unique key-value pair.
   • Dropped same-key, different-value (duplicate key conflict): 394 fields. A document asserting the same key twice with different values breaks the single-value assumption required for KIE ground truth. Top duplicate-dropped keys:
     • indirizzo 41
     • codice-fiscale 38
     • telefono 33
     • nome-completo 31
     • cap 30
     • firma 27
     • data-nascita 24
   • Fixed ordering defect: The initial pipeline performed the "several answers" drop before the duplicate check, so a repeated field could silently fail to invalidate its single-answer sibling. This leaked 15 fields (1%) asserting a unique value that the page contradicted. Fixed; leaks are now 0.
4. Final dataset: 1461 fields across 199 documents (0–16 keys per document). Reduced from 1476 fields after the fix.

Schema

xfund-kie/schema.json defines 37 extraction keys with natural-language descriptions. Examples:

• codice-fiscale: Italian tax identification code
• nome-completo: full name (given + surname)
• data-nascita: date of birth
• indirizzo: street address
• telefono: telephone / mobile number
• iban: International Bank Account Number
• partita-iva: Italian VAT identification number

Full list in xfund-kie/schema.json.

Data format

• xfund-kie/it.train.json — training split (lang: it, split: train)
• xfund-kie/it.val.json — validation split (lang: it, split: val)

Each document contains:

• id: document identifier (e.g., it_train_0)
• image: path to source image (it.train/it_train_...jpg or it.val/it_val_...jpg — symlinked to ../xfund/)
• text: extracted text from the form page
• annotation: key-value mapping (only mapped / validated keys)
• grounding: question (label text on page) and answer_id (position reference) for each annotated key

Residual ambiguity (transparent limitations)

Not all source content is cleanly extractable. Two categories remain after auditing:

1. Same type, different label — 151 / 517 type-checkable fields (95 docs)

Another value of the same shape exists on the page with a distinct label. Inspected cases: nearly all carry a clearly different label (e.g., Data rilascio 20/12/2019 vs Data inizio validità 30/12/2020). These are determinable from the label; the difficulty is KIE accuracy, not broken ground truth.

2. Same field, label rules miss — 68 flagged (49 docs)

Two sub-families:

• Bare synonyms not matched by current rules: Codice Fis .:, Cod.fiscale, Numero telefono fisso. Example: it_val_4 has three CF-like strings; only Codice fiscale/P.IVA: maps, so the annotation claims one CF out of three. If these mapped, the duplicate rule would drop the key instead.
• Qualified variants of a different entity: Tel. Referente, E-mail PEC, Codice Fiscale Azienda, Titolare di polizza: Cognome. These are currently ignored, so telefono = the subject's phone and the referent's is out of schema.

Users should treat annotations as validated single-value extractions, not exhaustive page transcriptions.

License / upstream terms

This derived dataset inherits the terms of the original XFUND dataset. Image links (xfund-kie/it.train/, xfund-kie/it.val/) are symlinks pointing to the upstream xfund/ folder to avoid duplication. The structured JSON files (it.train.json, it.val.json, schema.json) and this dataset card represent the refined KIE conversion.

Refinement process

• Pipeline built with Claude Opus 5 + manual audit rules.
• Duplicate-conflict and ordering-defect checks applied programmatically.
• Residual ambiguity manually inspected on flagged subsets; no automatic removal of ambiguous cases that remain determinable by label.
• Dataset version: post-fix (ordering leak resolved, 1461 fields retained).

Cite this project

@misc{gemelli2026xfundkieit
  title        = {XFUND-kie-it: A key information extraction italian subset of the original XFUND dataset},
  author       = {Gemelli, Andrea},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/andreagemelli/xfund-kie-it}}
}