Cozum-4B

Cozum-4B is a custom-trained, 4 billion parameter language model specifically optimized for Turkish cultural context, Turkish instruction following, and mathematical problem-solving. Built upon the Qwen architecture, it is fine-tuned to go beyond simple translation, providing responses that feel natural, culturally aware, and highly accurate for native Turkish speakers.

Model Details

• Architecture: Qwen (4B parameters)
• License: Apache 2.0
• Language(s): Turkish, English
• Training Framework: LLaMA-Factory

Training Datasets

The model underwent Supervised Fine-Tuning (SFT) using a highly curated mixture of localized and structural datasets. The training pipeline was specifically designed to teach the model local nuances, idioms, and general Turkish culture, alongside robust logical reasoning:

• Aya Turkish Filtered: A specially filtered subset of the Aya dataset, optimized to remove noise while capturing authentic, human-curated Turkish cultural knowledge and natural phrasing.
• Aya Dataset: A comprehensive, multilingual instruction fine-tuning dataset utilized to reinforce broad knowledge and structural capabilities.
• Alpaca Turkish Combined: A combined Turkish instruction dataset designed to improve task completion and ensure the model aligns with user expectations in a localized context.
• GSM8K (Turkish Translation - gsm8k_tr): High-quality grade-school math word problems translated into Turkish to boost the model's step-by-step mathematical reasoning capabilities without losing language integrity.

Training Configuration

• Method: Supervised Fine-Tuning (SFT)
• Tools Used: LLaMA-Factory
• Model Format: Safetensors (Sharded for optimized loading)

Usage

The model is fully equipped with its necessary tokenizer and processor configuration files. You can load it directly using the transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Yigit-Karaman/Cozum-4B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

# Example prompting for cultural knowledge
messages = [
    {"role": "user", "content": "Bana Türk kahvesi kültürünün ve 'bir fincan kahvenin kırk yıl hatırı vardır' sözünün anlamını açıklar mısın?"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))