| 144 | |
| 145 | |
| 146 | def print_result(): |
| 147 | # 打印指标结果 |
| 148 | print("=============统计数据=========================") |
| 149 | print(f"function_call_sum: {function_call_sum}") |
| 150 | print(f"function_call_correct: {function_call_correct}") |
| 151 | print(f"function_call_fail: {function_call_fail}") |
| 152 | print(f"function_call_fail_functioncall: {function_call_fail_functioncall}") |
| 153 | print(f"function_call_fail_name: {function_call_fail_name}") |
| 154 | print(f"function_call_fail_param: {function_call_fail_param}") |
| 155 | print(f"function_call_fail_name_illusion: {function_call_fail_name_illusion}") |
| 156 | print(f"assistant_ans_sum: {len(assistant_ans_relevancy_list)}") |
| 157 | print(f"assistant_ans_relevancy: {np.mean(assistant_ans_relevancy_list)}") |
| 158 | print("=============实验结果=========================") |
| 159 | function_call_correct_rate = function_call_correct/function_call_sum |
| 160 | function_call_fail_rate = function_call_fail/function_call_sum |
| 161 | function_call_fail_functioncall_rate = function_call_fail_functioncall/function_call_fail if function_call_fail else 0 |
| 162 | function_call_fail_name_rate = function_call_fail_name/function_call_fail if function_call_fail else 0 |
| 163 | function_call_fail_param_rate = function_call_fail_param/function_call_fail if function_call_fail else 0 |
| 164 | function_call_fail_name_illusion_rate = function_call_fail_name_illusion/function_call_fail if function_call_fail else 0 |
| 165 | print(f"工具识别正确率fccr: {function_call_correct_rate}") |
| 166 | print(f"工具识别失败率fcfr: {function_call_fail_rate}") |
| 167 | print(f"工具调用识别失败占比fcffr: {function_call_fail_functioncall_rate}") |
| 168 | print(f"工具名识别失败占比fcfnr: {function_call_fail_name_rate}") |
| 169 | print(f"工具参数识别失败占比fcfpr: {function_call_fail_param_rate}") |
| 170 | print(f"工具幻觉识别失败占比fcfnir: {function_call_fail_name_illusion_rate}") |
| 171 | print(f"助手回复答案相关度aar: {np.mean(assistant_ans_relevancy_list)}") |
| 172 | print("==============================================") |
| 173 | # 保存数据 |
| 174 | with open("test_result_data.jsonl","w") as fw: |
| 175 | for line in test_result_lines: |
| 176 | print(line, file=fw) |
| 177 | |
| 178 | |
| 179 | def test_process(test_lines, gpu_index): |