Koko12345p/qwen-arabic-lora
Adapter vs Qwen/Qwen2.5-3B-Instruct. Koko12345p/qwen-arabic-lora is a LoRA adapter, not full weights: rank 16, on k_proj, o_proj, q_proj, v_proj in 36 layers (144 adapted matrices). It modifies Qwen/Qwen2.5-3B-Instruct at load time.