erenfazlioglu/turkish-instruct-reasoning-dpo-3.4m

🇹🇷 Turkish Instruct · Reasoning · DPO — ~3.4M

The largest open native-Turkish instruction-tuning suite: SFT + chain-of-thought reasoning + DPO preferences, with an independently verified reasoning tier and a unique Turkey-grounded slice.

En büyük açık native Türkçe talimat-eğitim seti: SFT + adım-adım muhakeme (CoT) + DPO tercih çiftleri; bağımsız doğrulanmış muhakeme katmanı ve Türkiye-temelli özgün dilim içerir.

📦 Examples	~3.44M ( SFT 3.26M · DPO 181k )
🗣️ Language	Turkish (native, not machine-translated)
🧠 Verified reasoning	~285k examples independently verification-checked
🇹🇷 Turkey-grounded	417k examples built on real Turkish text
📐 Format	chat messages[] (SFT) · {prompt,chosen,rejected} (DPO)

---

🌍 In other languages

• English: ~3.4M native Turkish examples for instruction-tuning (SFT), chain-of-thought reasoning, and DPO preference alignment — with a verified reasoning tier and a Turkey-grounded slice.
• Türkçe: Talimat-eğitimi, muhakeme ve DPO hizalama için ~3.4M native Türkçe örnek; doğrulanmış muhakeme katmanı + Türkiye-temelli dilim.
• Español: ~3,4M de ejemplos nativos en turco para instruction-tuning, razonamiento (CoT) y alineación por preferencias (DPO).
• Français : ~3,4M d'exemples turcs natifs pour l'instruction-tuning, le raisonnement (CoT) et l'alignement par préférences (DPO).
• Deutsch: ~3,4 Mio. native türkische Beispiele für Instruction-Tuning, Reasoning (CoT) und DPO-Präferenzausrichtung.
• Русский: ~3,4 млн нативных турецких примеров для instruction-tuning, рассуждений (CoT) и выравнивания по предпочтениям (DPO).
• 中文： 约 340 万条原生土耳其语样本，用于指令微调（SFT）、思维链推理（CoT）和 DPO 偏好对齐。
• العربية: نحو 3.4 مليون مثال أصلي باللغة التركية للضبط التعليمي والاستدلال التسلسلي ومواءمة التفضيلات (DPO).
• Português: ~3,4M de exemplos nativos em turco para instruction-tuning, raciocínio (CoT) e alinhamento por preferências (DPO).
• 日本語: 指示チューニング・思考連鎖推論・DPO選好整合のためのネイティブトルコ語データ約340万件。
• हिन्दी: इंस्ट्रक्शन-ट्यूनिंग, चेन-ऑफ-थॉट रीज़निंग और DPO प्रिफरेंस अलाइनमेंट के लिए ~3.4M मूल तुर्की उदाहरण।

---

📚 Composition

sft — chat format messages: [{role, content}]

source	count	description
up3_instructurca	~2.43M	Native-Turkish regeneration of InstrucTurca instructions (clean question + fresh answer)
q1_turkey_grounded	~417k	Tasks built on real Turkish social-media / news text (summary, sentiment, stance, NER, QA…) — the Turkey-grounded moat
q2_reasoning	~263k	Turkish math / logic / planning chain-of-thought — verified ✓
math_omi2	~109k	OpenMathInstruct-2 problems, native-Turkish worked solutions (source answer-verified)
eba_reasoning	~12k	National-exam questions with step-by-step explanations — verified ✓
eba_open_qa	~12k	Same questions reframed as open-ended Q&A
eba_mc_qa	~11k	National multiple-choice questions + correct answer
math_gsm8k	~8k	GSM8K, native-Turkish solutions — gold-answer verified ✓

dpo — {prompt, chosen, rejected} (~181k)

Turkish preference pairs for DPO / alignment.

---

🚀 Usage

from datasets import load_dataset

sft = load_dataset("erenfazlioglu/turkish-instruct-reasoning-dpo-3.4m", "sft", split="train")
dpo = load_dataset("erenfazlioglu/turkish-instruct-reasoning-dpo-3.4m", "dpo", split="train")

# verified reasoning only
verified = sft.filter(lambda x: x["verified"] and x["source"] in ("q2_reasoning", "eba_reasoning"))
# Turkey-grounded slice only
turkey = sft.filter(lambda x: x["source"] == "q1_turkey_grounded")

🛠️ How it was built

• Synthetic generation: instructions and answers were generated/rewritten with large language models, scaled via batched generation.
• Turkey-grounded slice (q1): built on real, public Turkish social-media / news text → teaches the model Turkish context, names and current affairs.
• Verified reasoning: q2_reasoning and eba_reasoning outputs were checked by an independent second verification pass (flawed ones corrected or removed); math slices are matched against gold/source answers.
• Cleaning: instruction-level dedup + empty/too-short/malformed filtering + provider-mention scrub.

⚠️ Limitations (honest)

• Synthetic / distilled — outputs may contain errors or biases; only the *_reasoning and math_* slices are additionally verified. Other slices are not verified.
• up3 origin: instruction topics derive from English sets (Turkish text is native, but context is often non-Turkey). For Turkey context use q1 / eba.
• q1: built on public posts; the author identity is not stored, but @mention handles may appear in text. Contains political / current-affairs content.
• DPO rejected answers are deliberately lower quality (not human preference labels).

📝 License & attribution

Mixed provenance (license: other). Source instructions: InstrucTurca, OpenMathInstruct-2 (CC-BY-4.0), GSM8K (MIT), national education content (EBA), public social-media posts. Assistant outputs are synthetically generated. Released for research; review the terms of the underlying sources before use.

Curated by Eren Fazlıoğlu — X/Twitter: @erenfazlioglutr

📌 Citation

@misc{fazlioglu_2026_turkish_instruct,
  title  = {Turkish Instruct, Reasoning and DPO Dataset (~3.4M)},
  author = {Eren Fazlıoğlu},
  year   = {2026},
  url    = {https://huggingface.co/datasets/erenfazlioglu/turkish-instruct-reasoning-dpo-3.4m},
  note   = {X/Twitter: @erenfazlioglutr}
}

Beğendiyseniz ❤️ bırakın ve paylaşın — Türkçe açık-kaynak AI topluluğuna katkı için. If you find this useful, please like & share.