aliasit-pii-dataset-v3-small

Half the training split of aliasit-pii-dataset-v3, with validation and test kept whole. For cheap training runs that still evaluate on the real thing.

This is a subsample of mapo80/aliasit-pii-dataset-v3. The training split is halved by uniform sampling on document id; validation and test are the full splits of the parent dataset, unchanged. It exists to make a training run cheap without making the evaluation a different question. Every category present in the parent is present here.

Documents	101,665
Entities	1,778,896
Entity types	79 (46 marked critical)
Languages	ar, de, el, en, es, fr, it, nl, pt, sl, tr
Taxonomy	v3.0.0 — taxonomy.yaml ships in this repo
Revision	1.0.0
Representation	text + character spans, end exclusive, non-overlapping
Distinct sources	27, each pinned to a commit SHA or a file checksum

Splits

Split	Documents	Entities	Languages	Training
train	69,338	1,047,508	11	allowed
validation	16,261	368,663	10	allowed
test	16,066	362,725	9	FORBIDDEN

test must not be used for training, nor for threshold calibration, model selection, distillation or hyper-parameter tuning. Anything that needs data for those uses validation. This is not a formality: these splits are how the project measures whether a model works, and an evaluation that cannot fail is not an evaluation.

Sources and licences

Every source is pinned: a 40-character commit SHA for Hugging Face repositories, a SHA-256 of the downloaded file for public web files. A branch is not a pin.

Source	Generation	Licence	Attribution	Docs
nvidia/Nemotron-PII	delta1	CC-BY-4.0	required	1,690
E3-JSI/synthetic-multi-pii-ner-v1	delta1 + delta3	MIT	—	433
abhinavdread/msme-document-presence-dataset	delta2	MIT	—	300
AIFA — farmaci di classe A per principio attivo	delta2	CC-BY-4.0	required	128
albertobarnabo/synthetic-receipts-ocr (IT locale)	delta2	APACHE-2.0	—	511
ANPR — archivio dei comuni (Ministero dell'Interno)	delta2	CC-BY-4.0	required	290
ANPR — tabella stati esteri (Ministero dell'Interno)	delta2	CC-BY-4.0	required	195
arnaudstiegler/synthetic_us_passports_easy	delta2	APACHE-2.0	—	294
computed in-repo (statutory codice fiscale algorithm)	delta2	APACHE-2.0	—	247
DataDock/geonames	delta2	CC-BY-4.0	required	1
dossier-legal/italian-legal-corpus	delta2	CC-BY-4.0	required	500
Ethosoft/TR-DocVQA-Synth	delta2	CC-BY-4.0	required	902
huseyinatahaninan/ContextualIntegritySyntheticDataset	delta2	APACHE-2.0	—	14
istat-ai/court-rulings-coi	delta2	APACHE-2.0	—	243
kurkowski/synthetic-contextual-anonymizer-dataset	delta2	MIT	—	373
mik3ml/personas-italian	delta2	APACHE-2.0	—	255
Ministero della Salute — ICD-9-CM, versione italiana 2007	delta2	IODL-2.0	required	520
Toridion/lindisfarne-m1	delta2	CC-BY-4.0	required	586
albertobarnabo/synthetic-receipts-ocr (UK + DE locales)	delta3	APACHE-2.0	—	344
IVASS — Lista imprese assicurative vigilate	delta3	CC-BY-4.0	required	190
lucianfialho/privacy-filter-br-dataset	delta3	APACHE-2.0	—	698
PICO2/tau2-bench-data	delta3	MIT	—	167
subhash-holla/pii-anon	delta3	CC0-1.0	—	176
ai4privacy/pii-masking-openpii-1m	v2	CC-BY-4.0	required	59,949
gretelai/synthetic_pii_finance_multilingual	v2	APACHE-2.0	—	19,258
rizzoaiacademy/anonimizzazione-testi-italiano-clean	v2	MIT	—	72,313
urchade/synthetic-pii-ner-mistral-v1	v2	APACHE-2.0	—	21,905

Audited and mapped but contributing no documents, because a higher-priority source had already covered the need: richardyoung/synthea-575k-patients, itamarcohen/insurance_car_Qfiles, strova-ai/financial_credit_dataset.

Document counts above are for the parent dataset; this subsample halves the training split and keeps validation and test whole.

Attribution notices

These sources are redistributed under licences that require attribution. If you redistribute this dataset, these notices must travel with it.

• NVIDIA, Nemotron-PII (CC-BY-4.0)
• Agenzia Italiana del Farmaco (AIFA), liste dei farmaci di classe A (CC-BY-4.0)
• Ministero dell'Interno - ANPR, archivio dei comuni (CC-BY-4.0)
• Ministero dell'Interno - ANPR, tabella 2 stati esteri (CC-BY-4.0)
• GeoNames (CC-BY-4.0), mirror DataDock/geonames
• dossier-legal, italian-legal-corpus (CC-BY-4.0)
• Ethosoft, TR-DocVQA-Synth (CC-BY-4.0)
• Ministero della Salute, Manuale ICD-9-CM versione italiana 2007 (IODL 2.0)
• Toridion, Lindisfarne M1 (CC-BY-4.0)
• IVASS - Istituto per la Vigilanza sulle Assicurazioni, Lista imprese assicurative vigilate (CC-BY-4.0)
• Ai4Privacy / Ai Suisse SA (CC-BY-4.0)

Licence mix: APACHE-2.0 (10), CC-BY-4.0 (10), CC0-1.0 (1), IODL-2.0 (1), MIT (5). The strictest terms are CC-BY-4.0 and IODL-2.0 — both allow commercial use and derivative works and require attribution — so the dataset as a whole travels as CC-BY-4.0 with attribution.

Record schema

id                  stable document id
text                the document
language            ISO 639-1, or `und`
split               train | validation | test | benchmark
ent_start[]         character offsets, `end` = start + len, exclusive
ent_len[]  ent_label[]  ent_source_label[]  ent_source_dataset[]  ent_mapping[]
unm_*[]             source annotations with no canonical counterpart, kept on the record
prov_*              dataset, repo_id, revision, original_id, licence, register, ...
fp_*                exact, normalized, entity_set, template, template_shape
cont_extra_json     origin (v2 / v3_delta…), source_family, annotation_origin, group_id

Invariants checked at build time and by the gate:

1. text[start:end] == entity text for every entity and every unmapped span;
2. 0 <= start < end <= len(text);
3. entities never overlap (the model is single-label BIO);
4. every label exists in the shipped taxonomy;
5. spans survive the round trip through the tokenizer, BIO encoding and decoding.

Known data gaps

2 categories sit below a hard support threshold, 1 of them critical. They are listed rather than hidden, and each says which source was examined and why it was not usable.

Category	Critical	Train	Val	Test	Reason
prescription_number	yes	238	23	21	INSUFFICIENT_EXACT_RECORDS
train_ticket_number	—	84	30	34	ONLY_EXISTING_SOURCE_AVAILABLE

Three entity types were removed from the taxonomy rather than filled with the wrong data: serial_number and vehicle_registration_number (no public source annotates them as the taxonomy defines them), and digital_signature (39% of its existing gold was the phrase "digital signature" instead of a value). Their spans were migrated or kept as unmapped spans, never silently dropped.

Limitations

1. Synthetic. Effectively all of it. No real person's data is in here, which is the point, but synthetic-to-real transfer is not measured by anything in this repository.
2. Derived frames. Part of the support comes from documents written around structured records rather than natural text. Every record declares which it is in cont_extra_json.annotation_origin.
3. Partial annotation from one source. E3-JSI/synthetic-multi-pii-ner-v1 has an open type vocabulary and only a verified allow-list is mapped, so its documents carry annotations for some of their entities and not all. For that reason every such document sits in train only.
4. Reference-model contamination. urchade/synthetic-pii-ner-mistral-v1 is the training set of urchade/gliner_multi_pii-v1, and rizzoaiacademy/anonimizzazione-testi-italiano-clean is declared as the training corpus of rizzo-pii-0.3B. Source validation and benchmark are different questions and must not be reported as one.
5. Two known id collisions. nvidia/Nemotron-PII reuses the same uid across rows with different text, so two document ids sit on more than one document, and one value of medical_record_number crosses the train/validation boundary. It is measured, named, and cannot be fixed without rewriting a frozen document.

Verifying what you downloaded

manifest.json ships with the files and carries the SHA-256 of each parquet. A dataset without a committed checksum is not a version, it is a copy — if a hash differs, this card does not describe what you have.

import hashlib, json, pathlib
m = json.load(open('manifest.json'))
for f in m['files']:
    h = hashlib.sha256(pathlib.Path('data', f['path'].split('.')[-2] + '.parquet').read_bytes()).hexdigest()
    print(f['path'], h == f['sha256'])

Citation

@dataset{aliasit_pii_aliasit_pii_dataset_v3_small,
  title   = {aliasit-pii-dataset-v3-small},
  version = {1.0.0},
  note    = {79 PII entity types, multilingual, Italian-first},
}

Derived from the sources listed above, under their respective licences. The attribution notices must travel with it.