---
base_model: unsloth/gemma-3-12b-it
tags:
- text-generation
- transformers
- unsloth
- llama
- trl
- grpo
license: gemma
language:
- tr
library_name: peft
---

# TurkishReasoner-Gemma3-12B

## Model Description
TurkishReasoner-Gemma3-12B is a specialized reasoning model fine-tuned from Google's Gemma3-12B specifically for Turkish language reasoning tasks. This model excels at structured problem-solving with step-by-step reasoning capabilities, making it ideal for complex mathematical, logical, and analytical problems in Turkish.

## Key Features
- Built on Google's multimodal Gemma3-12B foundation
- Fine-tuned specifically for Turkish reasoning using GRPO (Group Relative Policy Optimization)
- Supports both text and image inputs for comprehensive reasoning tasks
- Delivers structured, step-by-step reasoning with clear solution formatting
- Maintains the base model's 128K token context window
- Trained on high-quality Turkish reasoning datasets including GSM8K-tr

## Technical Specifications
- Base Model: Google/Gemma3-12B
- Parameters: 12 billion
- Input: Text and images (multimodal capabilities)
- Hardware Requirements: ~20GB VRAM (NVIDIA RTX 6000 Ada or equivalent)
- Training Infrastructure: NVIDIA Ada6000 GPU

## Usage
This model is optimized for reasoning-intensive applications in Turkish, including:
- Educational tools requiring detailed mathematical explanations
- Research applications exploring complex problem-solving
- Applications requiring structured reasoning with visual components
- Turkish-language AI assistants with advanced reasoning capabilities

## Example Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from peft import PeftModel
import torch

base_model = AutoModelForCausalLM.from_pretrained("unsloth/gemma-3-12b-it")
model = PeftModel.from_pretrained(base_model, "Chan-Y/TurkishReasoner-Gemma3-12B").to("cuda")
tokenizer = AutoTokenizer.from_pretrained("unsloth/gemma-3-12b-it")

pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
)

messages = [
    {"role": "system", "content": """Sen kullanıcıların isteklerine Türkçe cevap veren bir asistansın ve sana bir problem verildi.
Problem hakkında düşün ve çalışmanı göster.
Çalışmanı <start_working_out> ve <end_working_out> arasına yerleştir.
Sonra, çözümünü <SOLUTION> ve </SOLUTION> arasına yerleştir.
Lütfen SADECE Türkçe kullan."""},
    {"role": "user", "content": "121'in karekökü kaçtır?"},
]

response = pipe(messages, return_full_text=False)[0]["generated_text"]
print(response)
```

For more information or assistance with this model, please contact the developers:
- Cihan Yalçın: https://www.linkedin.com/in/chanyalcin/
- Şevval Nur Savcı: https://www.linkedin.com/in/%C5%9Fevval-nur-savc%C4%B1/