Datasets:

rtemis
/

synthetic-clinical-notes

id stringlengths 3 3	age int32 34 83	sex stringclasses 2 values	smoking_status stringclasses 3 values	hypertension stringclasses 2 values	coronary_artery_disease stringclasses 2 values	heart_failure stringclasses 2 values	atrial_fibrillation stringclasses 2 values	diabetes stringclasses 2 values	hypothyroidism stringclasses 2 values	asthma_or_copd stringclasses 2 values	chronic_kidney_disease stringclasses 2 values	depression stringclasses 2 values	stroke_or_tia stringclasses 2 values	epilepsy stringclasses 2 values	parkinson_disease stringclasses 2 values	dementia stringclasses 2 values	multiple_sclerosis stringclasses 2 values	migraine stringclasses 2 values	peripheral_neuropathy stringclasses 2 values	cancer_history stringclasses 2 values	cancer_site stringclasses 8 values	metastatic_cancer stringclasses 2 values	cancer_status stringclasses 3 values
P01	58	female	never	yes	no	no	no	yes	no	no	no	no	no	no	no	no	no	yes	no	yes	breast	no	remission
P02	72	male	former	no	yes	no	yes	no	no	no	no	no	yes	no	no	no	no	no	no	no	none	no	none
P03	45	female	never	no	no	no	no	no	yes	yes	no	yes	no	no	no	no	yes	no	no	no	none	no	none
P04	66	male	current	yes	no	no	no	no	no	yes	no	no	no	no	no	no	no	no	no	yes	lung	yes	active
P05	79	female	never	yes	no	yes	no	no	yes	no	yes	no	no	no	yes	yes	no	no	no	no	none	no	none
P06	34	male	former	no	no	no	no	no	no	no	no	no	no	yes	no	no	no	yes	no	no	none	no	none
P07	61	male	current	yes	yes	no	no	yes	no	no	no	no	no	no	no	no	no	no	yes	yes	prostate	no	remission
P08	52	female	never	no	no	no	no	no	no	no	no	yes	no	yes	no	no	no	no	no	yes	brain	no	active
P09	70	male	former	yes	no	no	yes	yes	no	no	no	no	yes	no	no	no	no	no	yes	yes	colorectal	no	remission
P10	48	female	never	no	no	no	no	no	no	no	no	yes	no	no	no	no	no	yes	no	yes	hematologic	no	remission
P11	83	male	former	yes	no	yes	no	no	no	no	yes	no	yes	no	no	yes	no	no	no	yes	skin	no	remission
P12	39	female	never	no	no	no	no	no	no	no	no	no	no	no	no	no	no	no	no	no	none	no	none

rtemis synthetic clinical notes

Synthetic clinical notes for demonstrations of structured data extraction with large language models: 12 past-medical-history paragraphs, one per fictional patient, and reference labels for 23 fields per patient. The notes and labels are written by hand and describe no real person.

The dataset has two configs, joined by id:

notes: the note of each patient and its difficulty.
labels: the reference value of each field for each patient.

Structure

data/notes.parquet: 12 rows, 3 columns.

Column	Type	Description
`id`	string	Patient identifier, `P01` to `P12`
`difficulty`	categorical	`easy`, `moderate`, or `hard`; 4 notes each
`note`	string	Past medical history, one paragraph of free text

Easy notes state each condition in full words.
Moderate notes use clinical abbreviations (HTN, T2DM, CKD, s/p), negations, and family history.
Hard notes establish some conditions through a medication or an imaging finding (metformin for diabetes, old infarcts on MRI for stroke), and mention conditions that were ruled out or are borderline.

data/labels.parquet: 12 rows, 24 columns: id and 23 fields.

Fields	Type	Levels
`age`	integer	Years
`sex`	categorical	`female`, `male`
`smoking_status`	categorical	`never`, `former`, `current`
16 conditions	categorical	`no`, `yes`
`cancer_history`, `metastatic_cancer`	categorical	`no`, `yes`
`cancer_site`	categorical	`none`, `breast`, `prostate`, `lung`, `colorectal`, `brain`, `hematologic`, `skin`, `other`
`cancer_status`	categorical	`none`, `active`, `remission`

The 16 conditions, by body system:

Cardiovascular: hypertension, coronary_artery_disease, heart_failure, atrial_fibrillation
Endocrine: diabetes, hypothyroidism
Pulmonary: asthma_or_copd
Renal: chronic_kidney_disease
Psychiatric: depression
Neurological: stroke_or_tia, epilepsy, parkinson_disease, dementia, multiple_sclerosis, migraine, peripheral_neuropathy

Each condition is present in at least one note and absent in another. There are no missing values.

Labeling rules

A condition is yes when the note establishes a current or past diagnosis in the patient, including through a medication or an imaging finding, and no when it is not mentioned, is denied or ruled out, or occurs only in a family member.
stroke_or_tia includes transient ischemic attack and infarcts on brain imaging. epilepsy includes recurrent seizures of any cause, such as seizures due to a brain tumor. dementia includes dementia of any cause.
smoking_status is current for any current tobacco use, including cigars.
cancer_status is active under treatment or with known disease, remission when treated with no evidence of disease, and none when cancer_history is no. cancer_site is the primary site: brain metastases of a lung cancer have site lung and metastatic_cancer yes.

croissant.json states the rule of every field in its description.

Load

The Parquet files store each categorical column dictionary-encoded, with its levels in order, so Parquet readers return factors (R) or categorical columns (Python).

R - using rtemis:

notes <- rtemis::read("hf://datasets/rtemis/synthetic-clinical-notes/data/notes.parquet")
labels <- rtemis::read("hf://datasets/rtemis/synthetic-clinical-notes/data/labels.parquet")

R - using croissant, with categorical columns as factors in their declared level order:

library(croissant)
path <- "hf://datasets/rtemis/synthetic-clinical-notes/croissant.json"
notes <- read_record_set(path, "notes")
labels <- read_record_set(path, "labels")

Python - using polars:

import polars as pl
notes = pl.read_parquet("hf://datasets/rtemis/synthetic-clinical-notes/data/notes.parquet")

Use

The notes are input to a language model with an output schema whose fields match the columns of labels; the extracted values are then compared with the labels, overall and by difficulty. In R, rtemis.llm defines the schema with schema() and field() and runs the batch with llmapply().

Metadata

croissant.json is the dataset's Croissant 1.1 metadata, written with the croissant R package. It describes each file with its SHA-256 checksum, each field with its labeling rule, and the Responsible AI properties of the data. Each categorical column references an enumeration record set (notes_difficulty_levels, labels_sex_levels, ..., labels_cancer_status_levels) that lists its levels in order.

License

CC0 1.0: public domain dedication.

Citation

rtemis synthetic clinical notes (2026). https://huggingface.co/datasets/rtemis/synthetic-clinical-notes

@misc{rtemis_synthetic_clinical_notes_2026,
  title = {rtemis synthetic clinical notes},
  author = {Gennatas, E. D.},
  publisher = {rtemis},
  year = {2026},
  version = {1.0.0},
  url = {https://huggingface.co/datasets/rtemis/synthetic-clinical-notes}
}

Downloads last month: -

Number of rows:

Total file size:

64.5 kB