Update README.md
Browse files
README.md
CHANGED
|
@@ -20,12 +20,44 @@ To use MindGLM with the Hugging Face Transformers library:
|
|
| 20 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 21 |
tokenizer = AutoTokenizer.from_pretrained("ZhangCNN/MindGLM")
|
| 22 |
model = AutoModelForCausalLM.from_pretrained("ZhangCNN/MindGLM")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
'
|
| 31 |
|
|
@@ -58,4 +90,96 @@ For any queries, feedback, or collaboration opportunities, please reach out to:
|
|
| 58 |
- Email: [zcm200605@163.com]
|
| 59 |
- wechat: [Zhang_CNN]
|
| 60 |
- Affiliation: [university of glasgow]
|
| 61 |
-
- We hope MindGLM proves to be a valuable asset in the realm of digital psychological counseling for the Chinese-speaking community. Your feedback and contributions are always welcome!
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 21 |
tokenizer = AutoTokenizer.from_pretrained("ZhangCNN/MindGLM")
|
| 22 |
model = AutoModelForCausalLM.from_pretrained("ZhangCNN/MindGLM")
|
| 23 |
+
history = []
|
| 24 |
+
end_command = "结束对话"
|
| 25 |
+
max_length = 600
|
| 26 |
+
max_turns = 12
|
| 27 |
|
| 28 |
+
instruction = ""
|
| 29 |
+
instruction = "假设你是友善的心理辅导师,避免透露任何AI或技术的信息。请主动引导咨询者交流并帮助他们解决心理问题。"
|
| 30 |
+
|
| 31 |
+
# 获取用户输入
|
| 32 |
+
prompt = input("求助者:")
|
| 33 |
+
|
| 34 |
+
# 添加前缀
|
| 35 |
+
prompt = instruction + "求助者:" + prompt + "。支持者:"
|
| 36 |
+
response, history = model.chat(tokenizer, prompt, history=[])
|
| 37 |
+
print("支持者: " + response)
|
| 38 |
+
|
| 39 |
+
while True:
|
| 40 |
+
# 获取用户输入
|
| 41 |
+
user_input = input("求助者:")
|
| 42 |
+
|
| 43 |
+
# 检查是否收到结束指令
|
| 44 |
+
if user_input == end_command:
|
| 45 |
+
break
|
| 46 |
+
|
| 47 |
+
# 添加前缀
|
| 48 |
+
prompt = "求助者:" + user_input + "。支持者:"
|
| 49 |
+
response, history = model.chat(tokenizer, prompt, history=history)
|
| 50 |
+
print("支持者: " + response)
|
| 51 |
+
|
| 52 |
+
# 计算历史记录总字符长度,包括新对话
|
| 53 |
+
total_length = len(instruction) + sum(len(f'{item[0]} {item[1]}') for item in history)
|
| 54 |
+
|
| 55 |
+
# 如果历史记录字符数超过限制或者对话轮数超过限制,删除一些旧的对话
|
| 56 |
+
while total_length > max_length or len(history) > max_turns:
|
| 57 |
+
# 删除最旧的一条对话
|
| 58 |
+
removed_item = history.pop(1) # 第一条是instruction,删除第二条
|
| 59 |
+
# print('删除:',removed_item)
|
| 60 |
+
total_length -= len(f'{removed_item[0]} {removed_item[1]}')
|
| 61 |
|
| 62 |
'
|
| 63 |
|
|
|
|
| 90 |
- Email: [zcm200605@163.com]
|
| 91 |
- wechat: [Zhang_CNN]
|
| 92 |
- Affiliation: [university of glasgow]
|
| 93 |
+
- We hope MindGLM proves to be a valuable asset in the realm of digital psychological counseling for the Chinese-speaking community. Your feedback and contributions are always welcome!
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
# MindGLM: 针对中文心理咨询任务的对齐大语言模型
|
| 97 |
+
1. 简介
|
| 98 |
+
MindGLM 是针对中文心理咨询任务进行微调和对齐的大型语言模型。MindGLM 由基础模型 ChatGLM2-6B 发展而来,旨在与人类的心理咨询偏好产生共鸣,为数字心理咨询提供可靠、安全的工具。
|
| 99 |
+
|
| 100 |
+
2. 主要特点
|
| 101 |
+
- 针对心理咨询进行微调: MindGLM 经过细致的训练,能够理解和回应心理咨询,确保以同理心做出准确的回应。
|
| 102 |
+
|
| 103 |
+
- 符合人类偏好: 该模型经过了严格的调整过程,确保其响应符合心理咨询领域的人类价值观和偏好。
|
| 104 |
+
|
| 105 |
+
- 高性能: MindGLM 在定量和定性评估中均表现出卓越的性能,使其成为数字心理干预的首选。
|
| 106 |
+
|
| 107 |
+
3. 使用方法
|
| 108 |
+
将 MindGLM 与 "transformer "一起使用:
|
| 109 |
+
|
| 110 |
+
'
|
| 111 |
+
|
| 112 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 113 |
+
tokenizer = AutoTokenizer.from_pretrained("ZhangCNN/MindGLM")
|
| 114 |
+
model = AutoModelForCausalLM.from_pretrained("ZhangCNN/MindGLM")
|
| 115 |
+
history = []
|
| 116 |
+
end_command = "结束对话"
|
| 117 |
+
max_length = 600
|
| 118 |
+
max_turns = 12
|
| 119 |
+
|
| 120 |
+
instruction = ""
|
| 121 |
+
instruction = "假设你是友善的心理辅导师,避免透露任何AI或技术的信息。请主动引导咨询者交流并帮助他们解决心理问题。"
|
| 122 |
+
|
| 123 |
+
# 获取用户输入
|
| 124 |
+
prompt = input("求助者:")
|
| 125 |
+
|
| 126 |
+
# 添加前缀
|
| 127 |
+
prompt = instruction + "求助者:" + prompt + "。支持者:"
|
| 128 |
+
response, history = model.chat(tokenizer, prompt, history=[])
|
| 129 |
+
print("支持者: " + response)
|
| 130 |
+
|
| 131 |
+
while True:
|
| 132 |
+
# 获取用户输入
|
| 133 |
+
user_input = input("求助者:")
|
| 134 |
+
|
| 135 |
+
# 检查是否收到结束指令
|
| 136 |
+
if user_input == end_command:
|
| 137 |
+
break
|
| 138 |
+
|
| 139 |
+
# 添加前缀
|
| 140 |
+
prompt = "求助者:" + user_input + "。支持者:"
|
| 141 |
+
response, history = model.chat(tokenizer, prompt, history=history)
|
| 142 |
+
print("支持者: " + response)
|
| 143 |
+
|
| 144 |
+
# 计算历史记录总字符长度,包括新对话
|
| 145 |
+
total_length = len(instruction) + sum(len(f'{item[0]} {item[1]}') for item in history)
|
| 146 |
+
|
| 147 |
+
# 如果历史记录字符数超过限制或者对话轮数超过限制,删除一些旧的对话
|
| 148 |
+
while total_length > max_length or len(history) > max_turns:
|
| 149 |
+
# 删除最旧的一条对话
|
| 150 |
+
removed_item = history.pop(1) # 第一条是instruction,删除第二条
|
| 151 |
+
# print('删除:',removed_item)
|
| 152 |
+
total_length -= len(f'{removed_item[0]} {removed_item[1]}')
|
| 153 |
+
|
| 154 |
+
'
|
| 155 |
+
|
| 156 |
+
4. 训练数据
|
| 157 |
+
MindGLM 结合使用开源数据集和自建数据集进行训练,以确保全面了解心理咨询场景。这些数据集包括 SmileConv、comparison_data_v1、psychology-RLAIF、rm_labelled_180 和 rm_gpt_375。
|
| 158 |
+
|
| 159 |
+
5. 训练过程
|
| 160 |
+
该模型采用了三阶段训练方法:(均使用了LoRA)
|
| 161 |
+
|
| 162 |
+
监督微调: 使用 ChatGLM2-6B 基础模型,用心理咨询专用数据集对 MindGLM 进行微调。
|
| 163 |
+
|
| 164 |
+
奖励模型训练: 对奖励模型进行训练,以对微调模型的响应进行评估和评分。
|
| 165 |
+
|
| 166 |
+
强化学习: 使用 PPO(近端策略优化)算法对模型进行进一步调整,以确保其反应符合人类的偏好。
|
| 167 |
+
|
| 168 |
+
6. 局限性
|
| 169 |
+
虽然 MindGLM 是一款功能强大的工具,但用户也应了解其局限性:
|
| 170 |
+
|
| 171 |
+
它是为心理咨询而设计的,但不应取代专业医疗建议或干预。
|
| 172 |
+
|
| 173 |
+
模型的反应是基于训练数据的,虽然它与人类的偏好相一致,但可能并不总是提供最合适的反应。
|
| 174 |
+
|
| 175 |
+
7. 使用许可
|
| 176 |
+
请参考用于训练的数据集的许可条款。MindGLM的使用应遵守这些许可。
|
| 177 |
+
|
| 178 |
+
8. 联系信息
|
| 179 |
+
如有任何疑问、反馈或合作机会,请联系:
|
| 180 |
+
|
| 181 |
+
- 姓名: [张淙冕]
|
| 182 |
+
- 电子邮件 [zcm200605@163.com]
|
| 183 |
+
- 微信 [Zhang_CNN]
|
| 184 |
+
- 所属单位: [格拉斯哥大学]
|
| 185 |
+
- 我们希望 MindGLM 能够成为中文社区数字心理咨询领域的宝贵财富。我们随时欢迎您的反馈和贡献!
|