In July, I helped my team build a private LLM benchmark. This benchmark is both confidential and tailored to our specific needs. The reason for creating this benchmark is that many public benchmarks have become contaminated as LLMs continue to develop. Despite claims from various LLMs that they surpass GPT-4 in areas like reasoning, I haven't experienced such outstanding performance in practice. Constructing the test set was challenging, and I drew inspiration from how many classical test sets are constructed. For evaluation, I read numerous papers and learned more about LLM-as-a-Judge.