This is a fine-tuned version of the Qwen3-14B model optimized with Unsloth and 4-bit quantization for efficient inference. The model has been specifically trained on Turkish mathematical reasoning datasets to enhance its problem-solving capabilities in Turkish.

Key features

2x faster training using Unsloth optimization
4-bit quantization for reduced memory footprint
Enhanced Turkish mathematical reasoning capabilities
Compatible with Hugging Face's TRL library

Model Details

Base Model: unsloth/qwen3-14b-unsloth-bnb-4bit
License: Apache 2.0
Fine-tuned by: momererkoc
Language: Primarily Turkish (with English capabilities)
Quantization: 4-bit via BitsAndBytes

Training Data

The model was fine-tuned on specialized Turkish datasets:

Turkish Math 186k
OpenMathReasoning-mini

Usage

python

from huggingface_hub import login
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

login(token="hf_...")  

tokenizer = AutoTokenizer.from_pretrained("unsloth/qwen3-14b-unsloth-bnb-4bit",)
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/qwen3-14b-unsloth-bnb-4bit",
    device_map={"": 0}, token=""
)

model = PeftModel.from_pretrained(base_model,"momererkoc/qwen3-14b-reasoning-turkish-math186k")


question = "Bir çiftlikte 12 inek ve 8 tavuk vardır. Çiftliğe 5 inek daha eklenirse toplam hayvan sayısı kaç olur?"

messages = [
    {"role" : "user", "content" : question}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize = False,
    add_generation_prompt = True,
    enable_thinking = True, 
)
from transformers import TextStreamer
_ = model.generate(
    **tokenizer(text, return_tensors = "pt").to("cuda"),
    max_new_tokens = 3000,
    temperature = 0.6, 
    top_p = 0.95, 
    top_k = 20,
    streamer = TextStreamer(tokenizer, skip_prompt = True),
)

Performance

2x faster training compared to standard implementations
Reduced GPU memory requirements
Maintains strong Turkish language understanding
Enhanced mathematical reasoning capabilities

Optimization Details

The model uses:

Unsloth for accelerated training
4-bit quantization via BitsAndBytes
LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning

Acknowledgments

Special thanks to:

Unsloth team for optimization tools
Hugging Face for transformers ecosystem
Dataset creators for Turkish math datasets