AhmetSemih/Deepseek-mcq-reasoning-dataset

Turkish Reasoning Dataset

A Turkish reasoning dataset generated from alibayram/turkish_mmlu using DeepSeek-V3.2 (deepseek-reasoner). Each sample contains a multiple-choice academic question paired with a step-by-step rationale and internal thinking trace.

Dataset Summary

• —Source: Turkish MMLU (academic exam questions from TUS, KPSS, YKS, etc.)
• —Size: 1,000 samples
• —Language: Turkish
• —Generator Model: DeepSeek-V3.2 (deepseek-reasoner)
• —Purpose: Fine-tuning language models for Turkish reasoning tasks

Dataset Structure

How It Was Created

1. 1.1,000 samples were randomly selected from alibayram/turkish_mmlu (seed=42).
2. 2.Each question was sent to the DeepSeek-V3.2 API with the correct answer provided.
3. 3.The model generated a step-by-step rationale explaining why the answer is correct and why other options are incorrect.
4. 4.Both the internal thinking trace (thinking) and the final polished rationale (response) were captured.

Intended Use

Fine-tuning Reasoning Models

Use both thinking and response columns to train models with <think> tags:

<|user|> {question}
<|assistant|> <think>{thinking}</think>{response}

Fine-tuning Standard Models

Use the response column only:

<|user|> {question}
<|assistant|> {response}

Citation

If you use this dataset, please cite:

@dataset{turkish_reasoning_2026,
  title={Turkish Reasoning Dataset},
  author={Ahmet Semih Gumus},
  year={2026},
  url={https://huggingface.co/datasets/AhmetSemih/Turkish_reasoning_dataset},
  note={Generated from alibayram/turkish_mmlu using DeepSeek-R1}
}

License

Apache 2.0