benchmark cc-by-nc-nd-4 chatglm dataset evaluation gpt judge-model key-answer-extraction large-language-models llm llm-as-a-judge llm-as-evaluator lm-evaluation open-compass phi qwen regex reliability reliable-evaluation xfinder