README.md

metadata

language:
  - tr
license: cc-by-4.0
size_categories:
  - 100M<n<1B
task_categories:
  - text-generation
  - question-answering
tags:
  - reasoning
  - logic
  - turkish
  - pretraining
  - synthetic
pretty_name: Turkish Reasoning Data

Turkish Reasoning Data

A large-scale synthetic Turkish logic-puzzle dataset designed for pretraining and fine-tuning language models on structured reasoning tasks.

Dataset Summary

Property	Value
Language	Turkish (tr)
Format	Parquet (HuggingFace Datasets compatible)
Approximate tokens	~100 million
Number of examples	~720,000
Split	train only
License	CC BY 4.0

Load with datasets

from datasets import load_dataset

ds = load_dataset("DevHunterAI/turkish-reasoning-data", split="train")
print(ds[0])
# {'text': '<doc>\nAşağıdaki ipuçlarına...\n</doc>'}

Reasoning Categories

Each example belongs to one of eight logic-puzzle categories:

Category (TR)	Category (EN)	Weight
Sıralama bulmacaları	Ordering puzzles	25 %
Doğru/Yanlış ifade	True/False statement evaluation	15 %
Çelişki tespiti	Contradiction detection	15 %
Kim ne yaptı	Person–action–object matching	15 %
Ev/kat bulmacaları	Zebra-style house/floor puzzles	10 %
Seçim argümanı	Argument strength evaluation	10 %
Sayısal sıralama	Numerical magnitude reasoning	10 %
Zaman çizelgesi	Timeline / before–after reasoning	10 %

Format

Every example has a single text field containing a <doc>...</doc> block:

<doc>
Aşağıdaki ipuçlarına göre Ahmet, Zeynep, Mert kişilerinin
sıralamayı belirleyin:

- Zeynep, Mert'den önce gelir.
- Ahmet 1. sıradadır.

Doğru sıralama nedir?

Adım adım çözüm:
  1. Ahmet
  2. Zeynep
  3. Mert

Sonuç: Doğru sıralama Ahmet > Zeynep > Mert şeklindedir.
</doc>

Intended Use

• Pretraining Turkish language models on structured reasoning
• Fine-tuning (split text on the solution separator to get instruction/response pairs)
• Evaluation of step-by-step reasoning in Turkish

Generation

Generated programmatically with random.seed(42) using randomized templates covering Turkish names, professions, cities, colors, objects, and events.

Limitations

• Fully synthetic — does not cover real-world or open-ended reasoning
• Puzzle difficulty is limited; puzzles involve 3–5 entities
• Solutions are template-generated

Citation

@dataset{devhunterai2025turkish_reasoning,
  title     = {Turkish Reasoning Data},
  author    = {DevHunterAI},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/DevHunterAI/turkish-reasoning-data}
}