Muhammed Köse

@MuhammedKsee

AI Developer focused on Turkish NLP and Vision Language Models. Fine-tuning large language models with localized datasets on H100 infrastructure. Founder of Visuality.

🎓 Karabük University · 2026 🚀 Founder @ Visuality 📍 Turkey

Featured model

🇹🇷 Qwen3.5-9B-Instruct-Turca-TurkishLLM

Supervised fine-tune of Qwen/Qwen3.5-9B on 500K InstrucTurca samples for Turkish instruction following, reasoning, and NLG. No catastrophic forgetting — English capabilities fully preserved.

text-generation turkish lora sft qwen3.5-9b peft 0.18.1 apache-2.0

English reasoning — 0-shot

MMLU

77.9%

acc · 0-shot

HellaSwag

78.3%

acc_norm · 0-shot

ARC-Challenge

53.8%

acc_norm · 0-shot

Reasoning & code

GSM8K strict

84.9%

exact_match · 5-shot

GSM8K flex

84.4%

exact_match · 5-shot

HumanEval

26.2%

pass@1 · 0-shot

TruthfulQA

47.2%

acc · 0-shot

Turkish NLP — 0-shot

Belebele TR

81.4%

acc · 0-shot

Turkish MMLU

65.6%

acc · avg

XCOPA TR

67.8%

acc · 0-shot

About

Current focus

Fine-tuning Qwen3.5 on localized Turkish instruction datasets using synthetic data pipelines on H100 infrastructure.

Research

Vision Language Models (VLMs), multimodal architectures, and efficient local inference with GGUF / llama.cpp.

Notable project

custom-clip-vit-b-coco — Vision Transformer optimized for RTX 3050 Ti, scalable to A100/H100.

Visuality

Founder of an AI-focused initiative building agentic systems and workflow automation with n8n.