uzcaliskan/magibu-llm-fine_tuned_sondaj
The uzcaliskan/magibu-llm-fine_tuned_sondaj is a 4.3 billion parameter Qwen3.5 family model, fine-tuned by uzcaliskan using Unsloth for efficient training. Quantized to GGUF (Q4_K_M), this model demonstrates superior performance in Turkish MMLU evaluations, achieving a 57.02% success rate, outperforming an 8B Llama 3.1 model by 6.02 percentage points. It is optimized for Turkish question-answering tasks and local inference on GGUF-compatible platforms.
Loading preview...
Model Overview
uzcaliskan/magibu-llm-fine_tuned_sondaj is a 4.3 billion parameter model from the Qwen3.5 family, developed by uzcaliskan. It was fine-tuned using Unsloth and Hugging Face's TRL library, then quantized to GGUF (Q4_K_M) for efficient local deployment. The model is released under an Apache-2.0 license.
Key Capabilities & Performance
This model excels in Turkish language understanding, specifically in question-answering tasks. Benchmarked against llama3.1:8b on a 6,200-question Turkish MMLU dataset, magibu-llm-fine_tuned_sondaj achieved a 57.02% success rate, surpassing the 8B Llama 3.1 model by 6.02 percentage points (57.02% vs 51.00%). This indicates effective fine-tuning for Turkish-specific tasks, despite its smaller parameter size (4.3B vs 8B).
Usage and Considerations
The model is provided in GGUF format, making it compatible with tools like llama.cpp and Ollama for local inference. While demonstrating higher accuracy in Turkish MMLU, the model exhibited a slower average response time (1.51 sec/question) compared to 0.71 sec/question) during benchmarking. This performance difference might be influenced by hardware or loading conditions, suggesting further testing on consistent hardware for definitive conclusions. The evaluation methodology included both exact answer matching and semantic similarity using llama3.1:8b (paraphrase-multilingual-mpnet-base-v2.