DeepSeek-R1-Distill-Qwen-1.5B inference answer is wrong

I encountered the following problem when using llama cpp python on Mac. The model's answer is completely unreasonable. The red box contains the question and answer.


The configuration is as follows:
**model：** DeepSeek-R1-Distill-Qwen-1.5B.gguf
**Quantization：** Q4_K_M
**llama_cpp_python:** 0.3.7
**API：** create_chat_completion
**model input：**[{'role': 'system', 'content': 'you are a helpful assistant'}, {'role': 'user', 'content': 'There are 20 chickens and rabbits in a cage. It is known that these chickens and rabbits have 56 legs in total. Chickens have two legs and rabbits have four legs. How many chickens and rabbits are there in the cage?'}]