CraneAILabs/afrimed-pii-teval-novel

AfriMed-PII teval-novel v2: unseen-name clinical PII test set

The same documents as AfriMed-PII teval v2 with every patient and third-party name replaced by a name whose tokens never occur in the training vocabulary (0.0% token overlap; surrogate names from Wikidata, CC0), plus a reshuffle control in which the teval v2 names are permuted across documents. This is the set behind the unseen-name results in the model cards. Private, for Foundation review.

Version 2 (14 September 2026). Every person name was regenerated under per-language naming conventions (for example a gendered Luganda name with a Christian or Muslim given name of the same gender), gender-matched to the source; documents and all other spans are unchanged. Changelog: CHANGELOG.md.

teval-novel v2 datasheet

What this is. The teval v2 documents (MEDDOCAN test, machine-translated into Luganda, Kiswahili, Kinyarwanda and Dholuo) with every PERSON and PERSON_THIRD_PARTY name replaced by a name whose tokens never occur in the training vocabulary, plus a reshuffle control. Evaluation only.

Version 2, 14 September 2026. Names now follow the per-language naming conventions used across the project (handoff/d5/src/name_rules.py) and are gender-matched to the Spanish source record: the patient's sex field or given name, a clinician's title or given name, a relative's kinship word. In v1 the novel names were drawn without gender or name type.

Name pool, source and licence

• Source: Wikidata (CC0-1.0), SPARQL harvest of 2026-09-11 by citizenship (P27) and ethnic group (P172), with given name (P735), family name (P734) and sex or gender (P21); the held-out eval side of opt2/names/split_manifest.json.
• Tagging (handoff/opt2/src/novel_pool_v2.py): Christian, Western or Muslim given names by a 3:1 P21 majority; language names by morphology checked against P21. Wikidata records citizenship, not language, so a few Luganda-slot names are other Ugandan Bantu names.
• Real-name rules, as defined in the release-compliance escalation of 11 September 2026 against the 20,043 Wikidata (given, family) pairs in opt2/names/forbidden_pairs.json: rule A (the figure of record), any consecutive token pair inside the name is a real person's pair; rule B, the whole two-token name is. Generation also applied a stricter screen: no two elements of a name co-occur, in any order, in one documented person's name in the 30,083-person Wikidata harvest. No emitted name hits either rule or the stricter screen.

Reshuffle control

The teval v2 names permuted across documents inside buckets of the same element count and gender, so every substituted name still conforms to the grammar and matches the document's cue. It separates "a different name here" from "a name never trained on".

Construction and proofs

Language	Documents	Name spans	Distinct novel names	Novel name tokens	Token overlap with D5 v2 names	with teval v2 names	with D5 v1 names	Grammar violations novel / reshuffle	Real-name hits
Luganda	231 (70 DEV)	999	758	543	0	0	0	0 / 0	0
Kiswahili	235 (72 DEV)	1011	759	307	0	0	0	0 / 0	0
Kinyarwanda	227 (69 DEV)	980	737	425	0	0	0	0 / 0	0
Dholuo	241 (73 DEV)	1043	785	274	0	0	0	0 / 0	0

Every document passes the span-integrity gate against its teval v2 source: text outside the name spans identical, non-name spans identical, offsets valid, every surrogate re-detected exactly where it is labelled. Split rule unchanged: orig_idx modulo 10 in 0, 1, 2 is DEV (30%), the rest TEST.

Record shape

{text, lang, spans:[{start, end, label}], n_spans, orig_idx, split, variant, name_map, cmi, provenance}; name_map maps each teval v2 name in the document to its substitute.

Luganda facility names

The Luganda files contain surrogate health-facility names drawn from the Uganda Master Facility List, a public government registry of health facilities. Its licence terms for redistribution are under legal review; the names are public institutional names, not personal data. If the review finds a restriction, the affected names will be replaced and this dataset re-versioned.

Funding

This material is based on research funded by the Gates Foundation. The findings and conclusions contained within are those of the authors and do not necessarily reflect positions or policies of the Gates Foundation.