IndicPHI

Synthetic Indian clinical documents with PHI/PII span labels for 23 languages (22 scheduled Indian languages + English). All identifiers are synthetic surrogates, not real patients.

Two Hub configs at the repo root:

Config	Files	Rows	What it is
default	train.jsonl, eval.jsonl	22,554 / 3,982	Full documents: text, character spans, metadata
gliner	gliner_train.json, gliner_eval.json	22,889 / 4,054	GLiNER windows: tokenized_text + token ner

Also on the repo: curated.jsonl / curated.parquet (undivided curated set) and figures/.

Curated documents	26,536
Train / validation	22,554 / 3,982 (15.01%, language-stratified)
Languages	23
Document types	14
Clinical domains	6 (no paediatric rows)
Entity types	50
Span mentions	390,105
GLiNER windows	22,889 train / 4,054 validation
GLiNER mean tokens	189 (max 384; none truncated)
ECR	1.000
Run	20260830T222852 · seed 25001

Personas come from nvidia/Nemotron-Personas-India. English drafts are generated with Sarvam-105B (NeMo Data Designer). S4b tag-preserving translation uses Sarvam-105B first; Gemma 4 31B (gemma-4-31B-it) ran locally on an RTX 3090 as the fallback after Sarvam failed script or tag checks (rare-script languages: Bodo, Manipuri, Santali, Kashmiri, Sindhi, Dogri, Sanskrit; same ladder can apply to other languages). Grok-4.3 is the linguistic judge. Rows then pass a deterministic auditor (checksums, DICS, script purity) and NeMo Curator dedup.

Load

from datasets import load_dataset

ds = load_dataset("Sidharth1743/indicphi")  # JSONL documents
row = ds["validation"][0]
print(row["text"][:200])
print(row["spans"][:3])

gliner = load_dataset("Sidharth1743/indicphi", "gliner")
print(len(gliner["train"][0]["tokenized_text"]), gliner["train"][0]["ner"][:3])

Record format (default)

Field	Type	Description
document_id	string	{uuid}__0
uuid	string	Persona / document id
text	string	Surface text (tags already resolved to surrogates)
spans	list	{start, end, label, text} character offsets, end exclusive
tokenized_text	list[string]	Tokens used for GLiNER-style training
ner	list	[start_token, end_token, label] inclusive token indices
document_language_code	string	ISO code (hi, ta, brx, …)
doc_type_id	string	One of 14 clinical forms
domain_id	string	Clinical domain (6 values in this drop)
split	string	train or eval
token_align_errors	list	Empty on this release

Example (English SMS, validation):

HSP-UT-AAA2E4
Naveen Mishra, MRN-7831D797B4, urgent psychiatry screening at Rae Bareli District Hospital for panic attack. APT-240521-001 with Dr. Anil Kumar. Call 6903169442 for details.

start	end	label	text
0	13	HOSPITAL_ID	HSP-UT-AAA2E4
14	27	PATIENT_NAME	Naveen Mishra
29	43	MRN	MRN-7831D797B4
76	104	HOSPITAL_NAME	Rae Bareli District Hospital
123	137	APPOINTMENT_ID	APT-240521-001
143	157	DOCTOR_NAME	Dr. Anil Kumar

GLiNER format (gliner)

Use this config to fine-tune GLiNER. Each row is:

{"tokenized_text": ["MRN-15FE5B1DAF", "HSP-AS-935C09", "টোকা", "..."], "ner": [[0, 0, "MRN"], [1, 1, "HOSPITAL_ID"], ...]}

ner is [start_token, end_token, LABEL] with inclusive token indices.

These files are not a dump of the JSONL tokenized_text field. The JSONL tokenizer was Latin-centric and split Indic graphemes into characters (টোকা → ট,ো,ক,া), so mean length was 589 tokens and 68.3% of train documents exceeded GLiNER’s 384-token window.

The uploaded gliner_*.json files re-tokenize with Unicode letter/mark/number clusters (hyphenated IDs stay one token) and split leftover long docs into overlapping windows on entity boundaries. Span count is unchanged (390,105). Window counts are higher than document counts because of that split (train 22,554 → 22,889; eval 3,982 → 4,054). Mean tokens per window is 189; 0 windows are over 384.

All 50 labels appear across the two GLiNER splits (STUDENT_ID is train-only; CREDIT_CARD_NUMBER and CVV are eval-absent).

Languages

Quota was 1,272 personas per language (Bodo 1,268). Split is stratified by language; all documents from one persona UUID go to train or eval, never both.

Code	Language	Curated	Train	Val	Yield
en	English	1215	1033	182	96%
kok	Konkani	1208	1027	181	95%
mr	Marathi	1204	1023	181	95%
mai	Maithili	1201	1021	180	94%
or	Odia	1200	1020	180	94%
te	Telugu	1197	1017	180	94%
gu	Gujarati	1195	1016	179	94%
ne	Nepali	1187	1009	178	93%
kn	Kannada	1186	1008	178	93%
hi	Hindi	1185	1007	178	93%
as	Assamese	1184	1006	178	93%
bn	Bengali	1177	1000	177	93%
ur	Urdu	1171	995	176	92%
doi	Dogri	1159	985	174	91%
ml	Malayalam	1157	983	174	91%
pa	Punjabi	1126	957	169	89%
sd	Sindhi	1120	952	168	88%
ks	Kashmiri	1109	943	166	87%
ta	Tamil	1107	941	166	87%
brx	Bodo	1092	928	164	86%
sa	Sanskrit	1089	926	163	86%
mni	Manipuri	1052	894	158	83%
sat	Santali	1015	863	152	80%

Santali, Manipuri, Sanskrit, Tamil, Kashmiri, and Bodo lose more rows at the linguistic judge (script / fluency), not at the split.

Document types and domains

Types (14): discharge_summary (2090), radiology_report (2075), er_triage_notes (2063), opd_slip (2061), phc_register (2052), telemedicine_transcript (2048), prescription (2031), referral_letter (1884), lab_report (1854), automated_sms (1752), asha_worker_note (1722), insurance_claim (1710), hospital_billing (1614), surgical_note (1580).

Domains in this release: tb_ncd (4632), oncology_chronic (4611), general_medicine (4596), maternal_health (4488), psychiatry_behavioral (4265), surgical (3944). Config also defines paediatric (max age 17). This curated set has no paediatric rows — sampled personas are adults (age 18–111).

Entity types (50)

Personal: PATIENT_NAME, RELATIVE_NAME, DOB, AGE, GENDER, RELIGION, CASTE, OCCUPATION

Government IDs: AADHAAR_NUMBER, PAN_NUMBER, PASSPORT_NUMBER, DRIVING_LICENCE, VOTER_ID, BPL_RATION_CARD, TAX_ID

Contact / location: PHONE_NUMBER, TELEPHONE_LANDLINE, EMAIL_ADDRESS, RESIDENTIAL_ADDRESS, DISTRICT, VILLAGE, PIN_CODE, IP_ADDRESS, URL

Financial: BANK_ACCOUNT_NUMBER, IFSC_CODE, BANK_ROUTING_NUMBER, CREDIT_CARD_NUMBER, CVV, PIN

Other: EMPLOYEE_ID, STUDENT_ID, IMEI_NUMBER, MAC_ADDRESS, VEHICLE_REGISTRATION

Healthcare IDs: ABHA_ID, ABHA_ADDRESS, MRN, PATIENT_ID, HOSPITAL_ID, ADMISSION_NUMBER, ENCOUNTER_ID, APPOINTMENT_ID, REFERRAL_ID

Providers / facility: DOCTOR_NAME, ASHA_WORKER_NAME, HOSPITAL_NAME, BED_NUMBER, WARD_NUMBER, INSURANCE_POLICY_NUMBER

Mean ISED is 0.41 because closed-class types (gender, age, religion, IFSC) repeat by design. High-cardinality IDs stay diverse (HOSPITAL_ID 0.997, MRN 0.983, PHONE_NUMBER 0.978).

Intended use

• default: inspect documents, character spans, language / doc-type metadata.
• gliner: fine-tune GLiNER (or similar token NER) on the re-tokenized windows.

Do not use as real medical records, for clinical decisions, or as if the identifiers belong to real people.

Limitations

• Fully synthetic. Clinical facts are generated, not extracted from hospitals.
• Some Indic documents still contain English names, hospital names, or Latin IDs (allowed for several entity types).
• Yield is lower for Santali, Manipuri, Sanskrit, Tamil, Kashmiri, and Bodo.
• No paediatric domain rows in this drop.
• DICS requires repeated identity tags (phone, name, MRN, …) to share one value; inconsistent translations were dropped.
• Do not train GLiNER on JSONL tokenized_text; that tokenizer splits Indic scripts. Use the gliner config.
• GLiNER also spends part of its 384-token budget on label prompts, so max_len at train/inference still matters.

License

MIT. Personas are sampled from NVIDIA Nemotron-Personas-India; respect that dataset’s license as well.

Citation

@misc{indicphi2026,
  title        = {IndicPHI: Synthetic Indic Clinical PHI/PII NER Dataset},
  author       = {HackGrid},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {Run 20260830T222852, 26536 curated documents, 23 languages}
}

Pipeline: IndicPHI repository (NeMo Data Designer + Curator, Sarvam-105B generation/translation, Gemma 4 31B local RTX 3090 S4b fallback, Grok-4.3 judge).