ewhauser/outis-synthetic-phi

Outis synthetic clinical PHI corpora

Synthetic English clinical text annotated with character-offset spans for 14 types of protected health information (PHI). These are the exact training and checkpoint-selection corpora of the default Outis detector, ewhauser/outis-deberta-v3-base.

Every note is synthetic. Nothing here was drawn from real patient records, and no score on this data estimates accuracy on real clinical notes.

Splits

Split	Rows	Banks	PHI spans	PHI-free notes	Use in the default model
train	115,615	46	1,262,294	10,487	Training
validation	10,074	5	134,182	839	Checkpoint selection

The held-out evaluation bank behind the model card's headline numbers and all sealed or blind qualification banks are not included.

Labels

Label	train	validation
NAME	518,125	51,401
DATE	222,798	23,806
FACILITY	99,432	12,023
AGE	77,480	7,646
PHONE	72,797	7,872
ID	71,288	5,381
MRN	43,507	4,253
LOCATION	35,815	5,512
ADDRESS	31,310	3,714
EMAIL	22,182	3,012
ACCOUNT	20,437	2,628
URL	18,562	2,537
IP	15,864	2,216
SSN	12,697	2,181

Span boundaries and types follow the project's annotation policy (clinical-phi-v2) and its places addendum. Every age of the patient, a relative or a household member is annotated, not only ages over 89.

Schema

Field	Type	Contents
id	string	Row identifier, unique across the dataset
bank	string	Source corpus (the original file name without .jsonl)
text	string	The note
spans	list of struct	start, end (character offsets, end exclusive), label, text (always text[start:end]) and identity_group (set by a few banks, otherwise null)
annotation_policy	string	Always clinical-phi-v2
source_split	string	The split recorded inside the source bank
metadata	string	JSON object with the bank's own fields (cohort, template family, phenomena, layout and so on); these differ from bank to bank

import json
from datasets import load_dataset

ds = load_dataset("ewhauser/outis-synthetic-phi")
row = ds["train"][0]
for span in row["spans"]:
    print(span["label"], repr(span["text"]))
print(json.loads(row["metadata"]).get("phenomena"))

Exact originals

raw/train/ and raw/validation/ hold the original JSONL files, byte for byte the files the model was trained and selected on, each with its generator manifest (local absolute paths removed). banks.json lists every bank's split, row count and SHA-256; the hashes match the corpus_audit section of the model's training_summary.json.

How the data was made

The notes come from the project's own generators and from independently authored synthetic banks, each written for a stated gap such as names that are ordinary words, care-site boundaries, spoken dates, transcripts, FHIR resources or ages in context. People, dates, ages and identifiers are invented. The places addendum allows real US place and care-site names as surface values, taken only from US federal public-domain sources: the Census Bureau Gazetteer Files and the CMS Hospital General Information file.

Before use, every training bank passed an overlap audit against the sealed qualification banks, has no text or identity overlap with the development or held-out banks, and was filtered against the blind banks' hash-only exclusion lists. Bank file names record the cleaning passes applied (for example mix3clean and allages2).

Limitations

• Synthetic text is more regular than real notes; templates and phrasing recur within a bank.
• Invented names and identifiers can coincide with real people or numbers by chance.
• Real place and care-site names appear as values, so location spans are not all fictional.
• Exact boundaries follow one annotation policy; other corpora draw spans differently.

License

Project-authored synthetic text and annotations are dedicated to the public domain under CC0 1.0. The real place and care-site names come from US federal public-domain files.