Why this matters now Most public instruction datasets lack native-language chain-of-thought (CoT) examples, so models either hallucinate reasoning in translation or never learn explicit stepwise traces. This dataset supplies thousands of Turkish CoT-enabled dialogues so you can train or evaluate models that produce and consume Turkish internal reasoning natively.
What Sets It Apart
- Native-language CoT coverage: ~4,868 examples where the assistant emits a full Turkish reasoning trace between
<think>...</think>and a final answer, avoiding synthetic translations or post-hoc annotations. - Messages-ready format: each row is a JSON object with system/user/assistant chat roles (suitable for SFT pipelines that accept chat-format training data).
- Taxonomy-driven, culturally localized prompts: questions were generated with a 5‑dim taxonomy (14 domains, 13 reasoning types, 4 difficulties, 8 formats) and use Turkey-specific contexts (TL currency, local place and person names).
- Scope and density: average assistant output ≈3,900 characters and ~99% Turkish reasoning, with a detailed distribution across difficulty, domain, and format to enable balanced training/validation splits.
- Synthetic-but-structured generation: prompts and CoT answers were produced by a reasoning model (MiniMax-M3) under a two-stage pipeline (question generation → CoT generation), yielding consistent formatting for large-scale supervised fine-tuning.
Who it's for and tradeoffs
Great fit if you want to fine-tune or evaluate Turkish-capable LLMs for explainable, stepwise outputs (SFT or instruction-tuning), build Turkish conversational agents that must expose reasoning, or benchmark CoT generalization across domains. Use the messages field directly for chat-style SFT.
Look elsewhere or be cautious if you require human-crafted, gold-standard CoT for high-stakes evaluation: most examples are model-generated, labels like difficulty/format are target values from the generation process, and ~1% of traces contain English reasoning. Expect model-originated artifacts and potential biases; incorporate human review for critical applications. The dataset is Apache-2.0 licensed, allowing commercial use but requiring compliance with license terms.
