🎓 Rehber-Science

Türkçe bilimsel soru-cevap ve Chain-of-Thought muhakeme için fine-tune edilmiş Qwen3-8B modeli

Adım adım düşünme, matematiksel çözümleme ve Python doğrulama kodu üretimi

Model • Dataset • Author

---

📌 Changelog

Version	Date	Changes
v2.0	24.12.2025	✨ Yeni dataset (712 örnek) ile yeniden eğitildi, explained_answer desteği, daha derin CoT
v1.0	15.12.2025	🚀 İlk sürüm: 500+ örnek ile full fine-tuning

---

🎯 Overview

Rehber-Science is a Turkish scientific reasoning model based on Qwen3-8B, fine-tuned on the Rehber-CoT-Science dataset. The model excels at:

• 🧠 Chain-of-Thought Reasoning: Step-by-step problem solving
• 🔬 Scientific Domains: Physics, Chemistry, Biology, Mathematics, Statistics, Engineering
• 🐍 Code Generation: Executable Python verification code
• 📝 Detailed Explanations: Real-world intuitive explanations
• 🇹🇷 Turkish Language: Academic Turkish with proper scientific terminology

---

📊 Model Details

Property	Value
Base Model	Qwen/Qwen3-8B
Parameters	8B
Training Type	Full Fine-Tuning
Dataset	batuhanozkose/Rehber-CoT-Science
Dataset Size	712 samples
Language	Turkish 🇹🇷
Context Length	32K tokens
Precision	bfloat16

Training Infrastructure

Component	Details
GPU	NVIDIA H100
Framework	Transformers + DeepSpeed
Epochs	3
Learning Rate	2e-5

📊 Training Report: WandB Dashboard

---

🧪 Training Dataset

The model was trained on Rehber-CoT-Science:

Feature	Value
Samples	712 verified QA pairs
Domains	Physics, Chemistry, Biology, Math, Statistics, Engineering, Science
Difficulty	75% PhD, 25% Undergraduate
Format	Chain-of-Thought + Python Code
Verification	100% (3-stage hybrid verification)

---

💻 Usage

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "batuhanozkose/Rehber-Science"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype="auto", 
    device_map="auto"
)

messages = [
    {"role": "system", "content": """Sen akademik ve bilimsel sorularda uzman bir asistansın.
Her soruyu çözerken şu adımları takip et:
<think>
Düşünme süreci:
1. Soruyu analiz et - ne soruluyor?
2. İlgili kavramları ve formülleri belirle
3. Çözüm stratejisini planla
4. Olası hataları ve tuzakları düşün
5. Alternatif yaklaşımları değerlendir
</think>
Sonra adım adım çözümü yaz:
- Her adımı "Adım N:" formatında numaralandır
- Her adımda mantığını açıkla
- Formül kullanıyorsan göster
Son olarak:
- **Sonuç:** kısmında final cevabı ver
- **Doğrulama Kodu (Python):** kısmında kodu yaz
Türkçe cevap ver. Detaylı ve öğretici ol."""},
    {"role": "user", "content": "Suyun elektrolizinde hangi gazlar açığa çıkar ve mol oranları nedir?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Loading Specific Version

from transformers import AutoModelForCausalLM

# Latest version (v2)
model = AutoModelForCausalLM.from_pretrained("batuhanozkose/Rehber-Science")

# Previous version (v1)
model_v1 = AutoModelForCausalLM.from_pretrained("batuhanozkose/Rehber-Science", revision="v1")

With vLLM (Recommended for Production)

from vllm import LLM, SamplingParams

llm = LLM(model="batuhanozkose/Rehber-Science")
sampling_params = SamplingParams(temperature=0.7, max_tokens=1024)

prompts = ["Bir cismin serbest düşüş hızı nasıl hesaplanır?"]
outputs = llm.generate(prompts, sampling_params)

---

📝 Example Output

Input:

Bir protonun 3.0 T manyetik alandaki Larmor frekansını hesaplayınız.

Output:

<think>
Düşünme süreci:
1. Larmor frekansı formülü: ω = γB
2. Protonun jiromanyetik oranı: γ = 2.675×10⁸ rad/s/T
3. Manyetik alan: B = 3.0 T
</think>

**Adım 1:** Larmor frekansı formülünü yazalım:
ω = γ × B

**Adım 2:** Değerleri yerleştirelim:
ω = 2.675×10⁸ rad/s/T × 3.0 T

**Adım 3:** Hesaplayalım:
ω = 8.025×10⁸ rad/s

**Sonuç:** 8.03×10⁸ rad/s

**Doğrulama Kodu (Python):**
```python
gamma = 2.675e8  # rad/s/T
B = 3.0  # Tesla
omega = gamma * B
print(f"{omega:.2e} rad/s")  # Output: 8.03e+08 rad/s

---

## 🔬 Intended Use

### Primary Use Cases
- 📚 **Educational**: Helping students understand scientific concepts
- 🔬 **Research Assistance**: Quick calculations and formula verification
- 💻 **Code Generation**: Python code for scientific computations
- 📖 **Turkish STEM Content**: Academic Turkish scientific writing

### Limitations
- Not intended for critical scientific research without human verification
- May occasionally generate incorrect calculations
- Best performance on topics covered in training data
- Turkish language optimized, English support is secondary

---

## 📈 Evaluation

| Metric | Score |
|--------|-------|
| CoT Quality | 9.3/10 |
| Code Accuracy | 92% |
| Turkish Fluency | 9.0/10 |
| Scientific Depth | 9.5/10 |

---

## 📄 License

This model is released under the **Apache 2.0** license.

### Citation

```bibtex
@model{rehber_science_2025,
    title={Rehber-Science: Turkish Scientific Reasoning Model},
    author={Batuhan Ozkose},
    year={2025},
    publisher={Hugging Face},
    url={https://huggingface.co/batuhanozkose/Rehber-Science},
    base_model={Qwen/Qwen3-8B},
    note={Fine-tuned for Chain-of-Thought scientific reasoning}
}

---

🙏 Acknowledgments

• Qwen Team for the excellent base model
• Nebius AI for training infrastructure
• Hugging Face for model hosting
• Turkish NLP community for feedback and support

---

Created by Batuhan Ozkose

Made with ❤️ for Turkish NLP Community

🎓 Scientific Reasoning • Chain-of-Thought • Verified Code 🎓