bug修复。

eval.py

@@ -85,7 +85,7 @@ def main(args):
     os.makedirs(save_result_dir)
 
     subject_list = ['computer_architecture', 'car_knowledge', 'car_use', 'car_market']
-    # subject_list = ['car_knowledge_in_train', 'car_use_in_train', 'car_market_in_train']
+    subject_list.extend(['car_knowledge_in_train', 'car_use_in_train', 'car_market_in_train'])
     # qa_subject_list = ['car_knowledge', 'car_use', 'car_market']
     qa_subject_list = ['car_market']
 
@@ -104,14 +104,14 @@ def main(args):
                                                    save_result_dir=save_result_dir)
         print("Acc:", correct_ratio)
 
-    # result_list = []
+    result_list = []
-    #
+
     # for subject_name in qa_subject_list:
     #     print("Now testing: " + subject_name)
     #     qa_file_path = os.path.join('data/qa', f'{subject_name}_qa.csv')
     #     qa_df = pd.read_csv(qa_file_path)
     #     result_list.append(evaluator.eval_qa(subject_name, qa_df, save_result_dir=save_result_dir))
-    #
+
     # if evaluator_class is not None:
     #     del evaluator
     #     evaluator = evaluator_class(

merge.py

@@ -0,0 +1,28 @@
+import pandas as pd
+import os
+
+def merge_df(save_result_dir, df1_name, df2_name, df1_row_name, df1_row_new_name, df2_row_name, df2_row_new_name, file_type='csv'):
+    df1_path = str(os.path.join(save_result_dir, df1_name))
+    df2_path = str(os.path.join(save_result_dir, df2_name))
+    if file_type == 'csv':
+        df1 = pd.read_csv(df1_path)
+        df2 = pd.read_csv(df2_path)
+    elif file_type == 'json':
+        df1 = pd.read_json(df1_path)
+        df2 = pd.read_json(df2_path)
+    else:
+        raise ValueError("Invalid file type. Please choose either 'csv' or 'json'.")
+    df2 = df2.rename(columns={df2_row_name: df2_row_new_name})
+    df1 = df1.rename(columns={df1_row_name: df1_row_new_name}).join(df2[df2_row_new_name])
+    result_file_name = f'{df1_name}_{df2_name}_merge.csv'
+    df1.to_csv(os.path.join(save_result_dir, result_file_name))
+
+
+if __name__ == "__main__":
+    dir_path = 'logs/pt_sft'
+    pt_file_name = 'output-pt-sft-1-0.95-0.5-1.2.json'
+    npt_file_name = 'output-npt-sft-1-0.95-0.5-1.2.json'
+    predict_row_name = 'Predict'
+    pt_predict_row_name = 'predict_finetune'
+    npt_predict_row_name = 'predict_origin'
+    merge_df(dir_path, pt_file_name, npt_file_name, predict_row_name, pt_predict_row_name, predict_row_name, npt_predict_row_name, 'json')

scoring/assessment_engine.py

@@ -44,18 +44,18 @@ class AssessmentEngine:
         else:
             print("Unknown file type:" + file_type)
             return
-        result_diff_df['rouge_score_finetune'] = 0
+        result_diff_df['rouge_score_finetune'] = '0'
-        result_diff_df['rouge_score_origin'] = 0
+        result_diff_df['rouge_score_origin'] = '0'
-        result_diff_df['acc_finetune'] = 0
+        result_diff_df['acc_finetune'] = '0'
-        result_diff_df['acc_origin'] = 0
+        result_diff_df['acc_origin'] = '0'
-        result_diff_df['fluency_finetune'] = 0
+        result_diff_df['fluency_finetune'] = '0'
-        result_diff_df['fluency_origin'] = 0
+        result_diff_df['fluency_origin'] = '0'
-        result_diff_df['diff_score'] = 0
+        result_diff_df['diff_score'] = '0'
-        result_diff_df['acc_response_finetune'] = 0
+        result_diff_df['acc_response_finetune'] = '0'
-        result_diff_df['acc_response_origin'] = 0
+        result_diff_df['acc_response_origin'] = '0'
-        result_diff_df['fluency_response_finetune'] = 0
+        result_diff_df['fluency_response_finetune'] = '0'
-        result_diff_df['fluency_response_origin'] = 0
+        result_diff_df['fluency_response_origin'] = '0'
-        result_diff_df['diff_score_response'] = 0
+        result_diff_df['diff_score_response'] = '0'
         start_time = time.time()
         finetune_rouge_score_sum = 0
         origin_rouge_score_sum = 0
@@ -163,11 +163,11 @@ class AssessmentEngine:
         else:
             print("Unsupported file type:" + file_type)
             return
-        result_df['rouge_score_finetune'] = 0
+        result_df['rouge_score_finetune'] = '0'
-        result_df['acc_finetune'] = 0
+        result_df['acc_finetune'] = '0'
-        result_df['fluency_finetune'] = 0
+        result_df['fluency_finetune'] = '0'
-        result_df['acc_response_finetune'] = 0
+        result_df['acc_response_finetune'] = '0'
-        result_df['fluency_response_finetune'] = 0
+        result_df['fluency_response_finetune'] = '0'
         rouge_score_sum = 0
         acc_score_sum = 0
         fluency_score_sum = 0

scoring/gpt_scorer.py

@@ -31,7 +31,8 @@ def request_gpt(prompt, retries=3):
             )
             return response.choices[0]['message']['content']
         except Exception as e:
-            print(f"An error occurred while scoring with ChatGPT: {e}, it's the {ordinal(i+1)} time.")
+            print(f"\nAn error occurred while scoring with ChatGPT: {e}, it's the {ordinal(i+1)} time.")
+    print("Failed to get response from ChatGPT. Use default score.")
     return None
 
 
@@ -54,7 +55,7 @@ class GPTScorer:
             chatgpt_score = extract_score(chatgpt_response)
             return chatgpt_response, chatgpt_score
         except Exception as e:
-            print("An error occurred while extract score:", e)
+            print("\nAn error occurred while extract score:", e)
             return None, '2'
 
     def generate_scoring_prompt(self, question, model_result, reference, origin_model_result=None):
@@ -82,12 +83,12 @@ class GPTScorer:
         elif self.eval_mode == "diff":
             base_prompt = [{
                 "role": "system",
-                "content": "你是一个汽车领域专家，接下来将向你提供一个问题、一个参考答案、一个大模型生成的结果和一个微调后大模型生成结果。"
+                "content": "你是一个汽车领域专家，接下来将向你提供一个问题、一个参考答案、一个大模型1生成的结果和一个大模型2生成的结果。"
-                           "请对比这些结果，判断微调后大模型的结果是否优于原模型。满分为3分。"
+                           "请对比这些结果，判断大模型2的结果和大模型1哪个更好。满分为3分。"
                            "信息的准确性应当被首要考虑，多余的未知真假的信息不应该带来加分。"
                            "对比时请关注结果和参考答案的契合度。"
-                           "评分标准为：认为回答优于原模型——3分。认为回答与原模型持平——2分。"
+                           "评分标准为：认为大模型2的结果更好——3分。认为两者结果持平——2分。"
-                           "认为回答不如原模型——1分。"
+                           "认为大模型1的结果更好——1分。"
                            "回复格式为：理由：xxx。因此，评分为x分。"
             }]
         if self.eval_mode == "diff":
@@ -96,8 +97,8 @@ class GPTScorer:
             prompt = base_prompt + [
                 {
                     "role": "user",
-                    "content": f"问题：{question}\n\n原模型生成的结果：{origin_model_result}\n\n"
+                    "content": f"问题：{question}\n\n大模型1生成的结果：{origin_model_result}\n\n"
-                               f"微调后模型生成的结果：{model_result}\n\n参考答案：{reference}"
+                               f"大模型2生成的结果：{model_result}\n\n参考答案：{reference}"
                 }
             ]
         else:

scoring/rogue_scorer.py

@@ -7,8 +7,6 @@ def get_rouge_score(s1, s2):
     rouge = Rouge()
     s1 = " ".join(jieba.cut(s1))
     s2 = " ".join(jieba.cut(s2))
-    # print(s1)
-    # print(s2)
     return rouge.get_scores(s1, s2)[0]
 
 

test.py

@@ -2,9 +2,9 @@ import argparse
 from scoring.assessment_engine import AssessmentEngine
 
 def main(args):
-    assessment_engine = AssessmentEngine("logs/other",args.openai_key)
+    assessment_engine = AssessmentEngine("logs/lxh",args.openai_key)
-    assessment_engine.eval_result_diff("0319output.csv")
+    assessment_engine.eval_result_diff("0416_DoRA_generated_predictions.csv")
-    assessment_engine.eval_result("output-pt-sft.json", "json")
+    # assessment_engine.eval_result("output-npt-sft-1-0.95-0.5-1.2.json", "json")
 
     # synthesis_rouge_score = 0.30358589506467687
     # print("微调模型ROUGE分数：", synthesis_rouge_score)