Qwen3.8-27B Turkish CPT

Continued pretraining of Qwen3.8-27B on Turkish text, trained with QLoRA on a single RTX 3090 (24 GB). The goal was better Turkish understanding without losing English ability.

Results against the base model:

• Turkish test perplexity: 8.05 to 7.33 (-9%)
• TurkishMMLU (90-question stratified set): 60.0% to 75.6%
• Belebele reading comprehension, Turkish: 83.3% to 90.0%
• Belebele reading comprehension, English: 96.7%, fully retained

Released on Hugging Face as a LoRA adapter, a merged BF16 model and a Q4_K_M GGUF that runs in Ollama, LM Studio and llama.cpp.

Tech: PyTorch, PEFT / QLoRA, Hugging Face, llama.cpp, CUDA