uznlp-uz/uz_edbench

UZ-EDBench

UZ-EDBench is an Uzbek-language benchmark for medical specialty classification and urgency prediction from clinical encounter descriptions. It contains 3,360 encounters: 2,348 in a development cohort (train.tsv) and 1,012 in an independent external validation cohort (test.tsv).

Each row represents an encounter with an original complaint, a cleaned description, structured clinical attributes, a target medical specialty, and an urgency label. Text is primarily in Uzbek Latin script. The release also includes a reference table of 24 medical specialty labels.

Files and cohorts

File	Hugging Face configuration	Split	Rows	Role
train.tsv	default	train	2,348	Development cohort for model training and internal validation
test.tsv	default	test	1,012	Independent external validation cohort for final evaluation
UZ-EDBench.Doctors.tsv	doctors	train	24	Medical specialty reference table

All three files are UTF-8, tab-separated tables with a header row. Counts exclude the headers. The 24 reference rows are not encounters and are not included in the total of 3,360.

The external validation cohort is exposed as the Hugging Face test split. There is no separate internal validation split. Create an internal validation partition or cross-validation folds from train.tsv when tuning models. Keep test.tsv reserved for final external evaluation.

The train split name in the doctors configuration is a loading convention for the reference table; it does not represent a patient cohort.

Tasks

• Medical specialty classification: predict target_doctor_label from an encounter description.
• Urgency classification: predict urgency_level from an encounter description.
• Structured clinical information extraction: study the relationship between the narrative and fields such as symptoms, duration, and suspected system.

The files contain encounter-level records. They do not provide token-level BIO tags or annotated entity spans for a standard named entity recognition benchmark.

Encounter schema

train.tsv and test.tsv share the following 13 columns:

Column	Description
ID	Source record identifier. Values can repeat; this is not a unique patient identifier.
text_raw	Original Uzbek complaint or encounter description.
text_clean	Cleaned or reformulated encounter description.
age_group	Recorded age band, such as0-12, 13-17, 18-35, 36-60, or 60+; may also be unknown or missing.
gender	Recorded gender code:M, F, M/F, or unknown; may be missing.
main_symptoms	Free-text description or list of the main symptoms.
duration	Free-text duration of the complaint.
severity	Recorded severity value, generally from1 to 5; may be missing. A clinical scale definition is not supplied in the TSV files.
suspected_system	Suspected body system or clinical domain, expressed as free text.
target_doctor_label	Target medical specialty code, linked toLabel in the reference table. See the data quality notes for noncanonical values.
target_doctor_name	Human-readable Uzbek name of the target specialist.
urgency_level	Urgency annotation, normallypast, o‘rtacha, yuqori, or shoshilinch.
notes	Additional clinical or annotation notes; may be missing.

Urgency labels

Value	English gloss
past	Low
o‘rtacha	Moderate
yuqori	High
shoshilinch	Emergency

These are dataset label glosses, not a validated clinical triage scale.

Medical specialty reference

UZ-EDBench.Doctors.tsv contains six columns, with the following exact headers:

Column	Description
№	Reference row number.
Label	Canonical medical specialty code.
Shifokor turi (to‘liq nomi)	Full Uzbek specialist name.
Asosiy yo‘nalish / tizim	Main clinical domain or body system.
Tipik simptom/kasalliklar (qisqacha)	Brief examples of typical symptoms or conditions.
Izoh	Explanatory notes.

Label	Specialist name
TER	Terapevt / Umumiy amaliyot shifokori
CAR	Kardiolog
NEU	Nevropatolog / Nevrolog
GAS	Gastroenterolog
PUL	Pulmonolog
END	Endokrinolog
DER	Dermatolog
OPH	Oftalmolog (Ko‘z shifokori)
PSY	Psixiatr / Psixoterapevt / Psixolog
URO	Urolog
GYN	Ginekolog
PED	Pediatr
TRM	Travmatolog / Ortoped
ENT	LOR – Otorinolaringolog
ONC	Onkolog
INF	Infeksionist
ALL	Allergolog / Immunolog
REH	Reabilitolog / Fizioterapevt
RHE	Revmatolog
NPH	Nefrolog
HEM	Gematolog
PSUR	Plastik va rekonstruktiv jarroh
SUR	Umumiy jarroh
EMR	Tez yordam / Shoshilinch tibbiy yordam shifokori

Loading the dataset

Install the Hugging Face datasets library:

pip install datasets

from datasets import load_dataset

encounters = load_dataset("uznlp-uz/uz_edbench", name="default")
train = encounters["train"]  # 2,348 development encounters
test = encounters["test"]    # 1,012 external validation encounters

doctors = load_dataset("uznlp-uz/uz_edbench", name="doctors", split="train")

The configurations in this card explicitly map the files to their splits and set the tab separator, following the Hugging Face data file configuration documentation.

To read downloaded files locally with pandas:

import pandas as pd

train = pd.read_csv("train.tsv", sep="\t", dtype=str, keep_default_na=False)
test = pd.read_csv("test.tsv", sep="\t", dtype=str, keep_default_na=False)
doctors = pd.read_csv(
    "UZ-EDBench.Doctors.tsv", sep="\t", dtype=str, keep_default_na=False
)

assert len(train) == 2348
assert len(test) == 1012
assert len(doctors) == 24

Evaluation

Use the development cohort for fitting models, selecting features, and tuning hyperparameters. Fix preprocessing and model choices before evaluating on the external validation cohort.

For specialty classification, report macro-F1, accuracy, per-class precision and recall, and a confusion matrix. For urgency classification, report macro-F1 and per-class recall, including recall for shoshilinch. State the input fields, label normalization rules, and any excluded rows so that results are reproducible.

For a text-only benchmark, explicitly choose text_raw or text_clean. Exclude the target label, its corresponding name, and annotation-derived fields such as notes from inputs when they could reveal the answer. If structured fields are used, report that setup separately and justify their availability at prediction time.

Data quality and limitations

• Record counts are row counts, not verified counts of unique patients. ID values repeat within each cohort.
• The supplied files contain repeated descriptions within cohorts. No exact text_raw or text_clean string is shared between the two cohorts; this check alone does not establish patient-level or semantic independence.
• Age bands use both hyphens and en dashes, and several fields contain missing values or free-text variants. Normalize these consistently and document the rules.
• The development cohort includes noncanonical specialty values (Pediatr, EMR → CAR, and EMR → GAS?), a missing specialty label, and an out-of-vocabulary urgency value. Review these entries before using a fixed class vocabulary; document any mappings or exclusions.
• Specialist names vary in the encounter tables. Use the reference table's Label field as the canonical specialty vocabulary.
• The cohort roles above follow the dataset provider's description. Collection sites and dates, recruitment criteria, annotation procedures, and patient-level linkage are not established by the TSV files themselves.
• This benchmark supports research on Uzbek clinical language. Dataset annotations and model predictions should not be treated as clinical advice or evidence of readiness for clinical deployment.

License

The dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. Attribute UZ-EDBench and include the repository URL or dataset DOI when using or redistributing it.