bug fix

Author: qiuhuachuan

convert_to_training_set.py

@@ -1,12 +1,16 @@
 import ujson
 import os
 
+
 def get_all_sessions():
     source_dir = f'./data'
-    
+
     all_sessions = []
     context = []
     counter = 0
+    # demo data
+    # for idx in range(56):
+    
     for idx in range(56032):
         try:
             with open(f'{source_dir}/{idx}.json', 'r', encoding='utf-8') as f1:
@@ -32,8 +36,9 @@ def get_all_sessions():
     print('done')
     return all_sessions
 
+
 def get_dialogue_history(dialogue_history_list: list):
-    
+
     dialogue_history_tmp = []
     for item in dialogue_history_list:
         if item['role'] == 'counselor':
@@ -43,10 +48,10 @@ def get_dialogue_history(dialogue_history_list: list):
         dialogue_history_tmp.append(text)
 
     dialogue_history = '\n'.join(dialogue_history_tmp)
-    
+
     return dialogue_history + '\n' + '咨询师：'
-    
-    
+
+
 
 def get_instruction_data(dialogue_history):
     instruction = f'''现在你扮演一位专业的心理咨询师，你具备丰富的心理学和心理健康知识。你擅长运用多种心理咨询技巧，例如认知行为疗法原则、动机访谈技巧和解决问题导向的短期疗法。以温暖亲切的语气，展现出共情和对来访者感受的深刻理解。以自然的方式与来访者进行对话，避免过长或过短的回应，确保回应流畅且类似人类的对话。提供深层次的指导和洞察，使用具体的心理概念和例子帮助来访者更深入地探索思想和感受。避免教导式的回应，更注重共情和尊重来访者的感受。根据来访者的反馈调整回应，确保回应贴合来访者的情境和需求。请为以下的对话生成一个回复。
@@ -61,7 +66,7 @@ def get_instruction_data(dialogue_history):
     all_sessions = get_all_sessions()
     train_dir = './train_dir'
     os.makedirs(train_dir, exist_ok=True)
-    
+
     with open(f'{train_dir}/train.json', 'w', encoding='utf-8') as f:
         for item in all_sessions:
             dialogue_history_list = item['dialogue_history_list']
@@ -74,6 +79,5 @@ def get_instruction_data(dialogue_history):
             )
             text = ujson.dumps({'instruction': instruction, 'output': output}, ensure_ascii=False)
             f.write(text+'\n')
-            
+
     print('********')
-    

lora_tune.sh

@@ -1,8 +1,8 @@
 training_type=smile
 
-CUDA_VISIBLE_DEVICES=3,4,5,6,7
+CUDA_VISIBLE_DEVICES=0,1,2,3
 # THUDM/chatglm2-6b
-nohup deepspeed --include=localhost:3,4,5,6,7 --master_port 8888 train.py \
+nohup deepspeed --include=localhost:0,1,2,3 --master_port 8888 train.py \
             --train_path train_dir/train.json \
             --model_name_or_path THUDM/chatglm2-6b \
             --per_device_train_batch_size 1 \

merge_lora.py

@@ -1,5 +1,5 @@
 import torch
-from smile.model import MODE
+from model import MODE
 import argparse
 from peft import PeftModel