aliFurkan123/cultural-questions-dataset

Turkish General Knowledge & Trivia CoT Dataset (TR-GenK-CoT)

TR-GenK-CoT is a high-quality, synthetic, and carefully curated Turkish dataset designed for Instruction Tuning and Chain of Thought (CoT) reasoning. It contains exactly 500 completely unique, non-repetitive general knowledge and trivia conversations with rich step-by-step thinking processes.

The dataset is formatted using standard Chat Template formats (matching OpenAI/Hugging Face chat schemas) making it directly compatible with popular fine-tuning frameworks like Axolotl, LLaMA-Factory, Unsloth, or Hugging Face SFTTrainer.

Dataset Structure

The dataset contains a single train split consisting of 500 conversation samples. Each sample is a list of two turns (User and Assistant):

• —User Message: A naturally phrased general knowledge question in Turkish, written in a casual, human-like chat style (sometimes containing minor typos or lack of capital letters, mimicking real messaging behavior).
• —Assistant Message: A detailed, expert, yet conversational response containing:
• —A detailed reasoning process in the thinking field (in Turkish, minimum 2-3 sentences and 100+ characters).
• —The final clear answer in the content field.

Data Fields

• —role: (string) The speaker's role, either user or assistant.
• —content: (string) The actual text of the message (user's question or assistant's answer).
• —thinking: (string) The Chain of Thought (CoT) reasoning. It is null for user messages and populated with detailed steps for assistant messages.
• —images: (null) Set to null (this is a text-only dataset).
• —tool_calls: (null) Set to null.

Topic Distribution

The 500 entries span a wide variety of subjects, ensuring high informational density and no repetitive concepts:

1. 1.Turkish Geography & History (140+ items): Historical sites (Göbeklitepe, Ephesus, Sümela), national parks, rivers, mountains, and Republican/Ottoman history.
2. 2.World Geography & Capital Cities (80 items): Unique country-capital pairs focusing on frequently confused capitals (e.g., Ottawa for Canada, Bern for Switzerland).
3. 3.Chemistry & Physics (80 items): Chemical elements with scientifically accurate atomic numbers and symbols, and physical/astronomical constants.
4. 4.World & Turkish Literature (60 items): Classics (Aşk-ı Memnu, İnce Memed, Don Quixote, Hamlet) and author-book trivia.
5. 5.Famous Figures & Philosophers (20 items): Achievements of historical thinkers (Socrates, Galileo, Marie Curie, Tesla, Mozart, etc.).
6. 6.Logical & Mathematical Puzzles (120+ items): Algorithmic algebra and logic questions, complete with step-by-step numerical calculations in the thinking field.

Sample Data Point

[
  {
    "content": "ya arkadaşlar Fas'ın başkenti neresiydi, hep Kazablanka ile karıştırıyorum",
    "images": null,
    "role": "user",
    "thinking": null,
    "tool_calls": null
  },
  {
    "content": "Fas'ın başkenti Rabat'dir. Genellikle en bilinen veya en büyük şehir olan Kazablanka ile karıştırılır ama resmi yönetim merkezi Rabat şehridir.",
    "images": null,
    "role": "assistant",
    "thinking": "Kullanıcı Fas ülkesinin başkentini soruyor. Genelde Kazablanka ile karıştırılan bu durumun doğrusunu, yani Rabat şehrini açıklayacağım.",
    "tool_calls": null
  }
]

How to Use

You can easily load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("your_username/tr-genk-cot")
print(dataset["train"][0])

Prompt Formatting

To train models with this dataset, configure your chat template to include the <thinking> tag (or similar tokens) for the assistant's reasoning process. E.g., for DeepSeek-R1 style formatting:

<|im_start|>user
[User Question]<|im_end|>
<|im_start|>thought
[Thinking Process]<|im_end|>
<|im_start|>assistant
[Final Answer]<|im_end|>

License

This dataset is released under the MIT License.