Rustamshry/Qwen3-0.6B-turkish-math-reasoning-80K

Model Card for Model ID

This model was fine-tuned on 80,000 Turkish math problems, targeting better understanding and generation of mathematically structured responses in Turkish. The dataset covers arithmetic, algebra, word problems, and other foundational math skills, allowing the model to serve as a multilingual math tutor or reasoning engine in Turkish.

Model Details

Model Description

• Language(s) (NLP): Turkish
• License: MIT
• Finetuned from model: unsloth/Qwen3-0.6B
• Domain: Mathematical Reasoning

Uses

Direct Use

• 🧮 Math problem solving in Turkish Can assist users in understanding and solving elementary to intermediate math problems written in Turkish.
• 📚 Educational tools and tutoring systems Suitable for integration into digital tutors, math practice apps, or classroom AI assistants for Turkish-speaking students.
• 💬 Multilingual reasoning research Can be used to evaluate Turkish-language mathematical reasoning tasks in LLM benchmarks.

Bias, Risks, and Limitations

🌐 Language bias Performance is limited to Turkish; multilingual or code-mixed input may confuse the model.

How to Get Started with the Model

Use the code below to get started with the model.

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

login(token="")  

tokenizer = AutoTokenizer.from_pretrained("unsloth/Qwen3-0.6B",)
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/Qwen3-0.6B",
    device_map={"": 0}, token=""
)

model = PeftModel.from_pretrained(base_model,"Rustamshry/Qwen3-0.6B-turkish-math-reasoning-80K")


question = """
Problem 2. $a, b$ iki farklı gerçel sayı ve $c$ öyle bir pozitif gerçel sayı olsun ki
$$a^4-2019a=b^4-2019b=c\mathrm{.}a^\{4\}-2019\; a=b^\{4\}-2019\; b=c.$$ $-\sqrt{c}<a b<0$ olduğunu kanıtlayın.
"""

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),
)

Training Data

The dataset ituperceptron/turkish-math-186k is a Turkish-language collection designed for training and evaluating language models on mathematical reasoning tasks. It comprises approximately 186,000 entries, each containing structured fields such as instruction, input, and output. The dataset is available in Parquet format and is intended for text generation tasks, particularly focusing on mathematical problem-solving in Turkish.

Framework versions

• PEFT 0.14.0