erenextlabs/turkish-decision-benchmark

Turkish Decision Benchmark

The Turkish Decision Benchmark contains 2,194 realistic Turkish messages with 19,861 labeled decision questions. It is intended for measuring the performance of decision models (Decision 3.0, Pusula or others) on Turkish. The benchmark is part of the Pusula project.

This benchmark is intended for evaluation only and should not be used for training.

Contents

• Coverage: all 20 domains of the Pusula taxonomy and its 299 intents are represented, including banking, accounting and tax, law, health, e-commerce, logistics and public services. Twelve writing styles are included: formal e-mails, everyday chat, missing Turkish characters and typos, angry capitals, long and rambling text, one short line, Turkish and English code-switching, an elderly user's roundabout politeness, young and emoji-heavy text, several issues with the real request last, error codes and logs, and voice dictation without punctuation. About 10% of the messages are security probes (prompt injection, jailbreak, unauthorized access, harmful requests, spam), and fake personal data is included in about 30%.
• Urgency reflects the situation, not the tone: the word "acil" appears in 696 messages, 55% of which are routine or can wait a few days.
• Questions per message: alan / alt_alan / niyet (domain, subdomain and intent routing), the six universal questions (aciliyet, eylem_gerekli, insana_devret, guvenlik, duygu, kisisel_veri) and, for some messages, an ad-hoc business question with its own options.

Construction and labels

The messages were written by Claude (Anthropic) from randomized recipes (intent, style, length and situation), and near-duplicates were removed. The benchmark contains no real people, companies or personal data. Messages that the annotation model declined to label (a small number of harmful-content probes) are not included.

Every question was answered blind and independently by two annotators; disagreements were resolved by a third, more careful pass. All labels follow a single written labeling policy, which defines the boundaries of the judgment questions. Annotator agreement before adjudication, which indicates the ambiguity of each question, is shown below:

Question	Agreement (%)
alan / alt_alan / niyet	97.5 / 96.6 / 94.6
aciliyet	85.2
eylem_gerekli	92.7
insana_devret	87.2
guvenlik	99.2
duygu	90.2
kisisel_veri	96.8

Results

Accuracy (%) on all 19,861 questions:

Model	Accuracy
Decision 3.0 (vllm-sr/d3)	88.5
Pusula	92.6
Pusula Nano	82.7

Files and format

Two files are provided with the same content:

• data/test.jsonl is loaded by the datasets library and shown in the dataset viewer. The columns are id, state (the message), questions (the questions as a JSON string), gold (the gold answers as a JSON string) and source.
• tr_decision_bench_v1.jsonl is the original file in the Pusula / d3 request format, with one object per line:

{
  "id": "…",
  "state": "Turkish message",
  "questions": {"alan": {"type": "choice", "instructions": "…", "criteria": {"…": "…"}}, "…": {}},
  "labels": {"gold": {"alan": "finans_bankacilik", "aciliyet": "2", "eylem_gerekli": "true", "…": "…"}}
}

Each question is a complete d3 systemone question, so any d3-compatible model can be scored by sending {"state", "questions"} and comparing its answers with the gold answers (choice: option key, noul: "true" or "false", score: level index as a string).

import json
from datasets import load_dataset

ds = load_dataset("erenextlabs/turkish-decision-benchmark", split="test")
row = ds[0]
questions, gold = json.loads(row["questions"]), json.loads(row["gold"])

With the Pusula package, a d3-compatible model is evaluated as follows:

python -m pusula.d3train eval --d3 ~/d3 --data tr_decision_bench_v1.jsonl --out base_d3.json

The benchmark is developed by Erenext's AI Lab.

Citation

@software{pusula2026,
  title   = {Pusula: The First Full-Featured Turkish Decision Model},
  author  = {Şahin, Hasan and Sayman, Ahmet},
  year    = {2026},
  version = {1.0},
  url     = {https://github.com/erenextlabs/pusula},
  license = {Apache-2.0}
}

License

Released under the Apache 2.0 license, as part of the Pusula project.