azbench

A zero-shot multiple-choice benchmark for Azerbaijani, 2,148 questions from three existing sources, all in Azerbaijani and all formatted the same way. Code, runner and results: github.com/aghasalim/azbench.

config	source	items	what it tests	licence
belebele-az	Belebele, azj_Latn (Bandarkar et al., 2024)	900	reading comprehension of a passage	CC BY-SA 4.0
include-az	INCLUDE, Azerbaijani (Romanou et al., 2024)	548	real academic and professional exam questions, 38 subjects	Apache 2.0
tumlu-az	TUMLU-mini, Azerbaijani (Isbarov et al., 2025)	700	native school-exam questions, 7 subjects	CC BY 4.0

I did not write these questions. What this dataset adds is one format and one prompt across all three, so a model's Azerbaijani score means the same thing on each. Each config keeps its source's licence; belebele-az is share-alike.

Fields

field	meaning
id	<source>/<index in the source split>
subject	subject or domain from the source
prompt	the exact Azerbaijani prompt azbench sends, passage included for Belebele
choices	the four options, in order A to D
answer	index of the correct option, 0 to 3
answer_letter	the same as a letter

from datasets import load_dataset
ds = load_dataset("aghasalim/azbench", "tumlu-az", split="test")

Scoring

Zero-shot, one question per call, the reply scored by the first letter A to D it contains. A reply with no letter counts as wrong.

Chance is not 25 % on every config. The answer letters are not balanced in tumlu-az: 299 of 700 answers are D, so always answering D scores 42.7 %. Report scores against this baseline, not against 25 %.

config	most common answer	always answering it scores
belebele-az	B	27.9 %
include-az	C	26.3 %
tumlu-az	D	42.7 %

Results

Only models that answered every question are listed. Both are 3B models at Q4_K_M through Ollama, zero-shot, temperature 0.

model	belebele-az	include-az	tumlu-az
Qwen2.5-3B-Instruct	45.2 %	36.9 %	36.0 %
SIBA Meristem 1.0.1 (Qwen2.5-3B fine-tuned on Azerbaijani)	38.9 %	40.0 %	37.7 %
always the most common answer	27.9 %	26.3 %	42.7 %

The fine-tune gains on exam questions and loses 6.3 points on reading comprehension. On tumlu-az neither model beats always answering D. Partial runs of larger models and the raw counts are in the GitHub repository.

Sources

Bandarkar, L. et al. The Belebele Benchmark: a Parallel Reading Comprehension Dataset in 122 Language Variants. ACL 2024. https://huggingface.co/datasets/facebook/belebele

Romanou, A. et al. INCLUDE: Evaluating Multilingual Language Understanding with Regional Knowledge. 2024. https://huggingface.co/datasets/CohereLabs/include-base-44

Isbarov, J. et al. TUMLU: A Unified and Native Language Understanding Benchmark for Turkic Languages. 2025. https://huggingface.co/datasets/jafarisbarov/TUMLU-mini

Citation

If you use azbench, cite the three sources above, and this dataset as:

@misc{mustafazada2026azbench,
  title  = {azbench: an Azerbaijani multiple-choice benchmark},
  author = {Mustafazada, Aghasalim},
  year   = {2026},
  url    = {https://huggingface.co/datasets/aghasalim/azbench}
}