最近在基于GLM-4的智能体多轮对话任务训练过程中发现finetune.py脚本中的两个bug：

1. finetune_demo/finetune.py中：
   process_batch函数中，以下代码部分存在漏洞：
   for message in conv:
   message = process_message(message)
   loss_mask_val = False if message['role'] in ('system', 'user', 'observation') else True
   new_input_ids = tokenizer.apply_chat_template([message], tokenize=True, return_dict=False)[2:]
   以及和process_batch_eval函数中以下代码部分存在漏洞：
   for message in conv:
   if len(input_ids) >= max_input_length:
   break
   else:
   message = process_message(message)
   new_input_ids = tokenizer.apply_chat_template([message], tokenize=True, return_dict=False)[2:]
   追溯apply_chat_template函数中结尾处代码：
   if tokenize:
   output = self.batch_encode_plus(
   [result] if isinstance(result[0], int) else result,
   padding=padding,
   truncation=truncation,
   max_length=max_length,
   return_tensors=return_tensors,
   is_split_into_words=True,
   add_special_tokens=False
   )
   if return_dict:
   return output
   else:
   return output["input_ids"]
   可知：
   tokenizer.apply_chat_template([message], tokenize=True, return_dict=False) 返回为具有一个向量的矩阵，此时直接的[2:]切片操作导致训练数据完全丢失，导致无法有效训练。而我们期望提取的数据是当前message编码后的向量数据（去掉[gMASK]），因此 解决办法是修改代码：
   new_input_ids = tokenizer.apply_chat_template([message], tokenize=True, return_dict=False)[2:]
   为：
   new_input_ids = tokenizer.apply_chat_template([message], tokenize=True, return_dict=False)[0][2:]

修改后，多轮对话模型可正常训练。

2. 当前finetune_demo/finetune.py脚本的实际代码不支持agent chat 训练数据，存在以下漏洞：
   从finetune_demo/finetune.py脚本中函数：
   def process_message(message):
   if 'tools' in message and message['role'] == 'system':
   for tool in message['tools']:
   parameters = tool['function']['parameters']['properties']
   tool['function']['parameters']['properties'] =
   {k: v for k, v in parameters.items() if
   v is not None}
   elif 'tools' in message:
   del message['tools']

以及 process_batch函数中：
else:
for message in conv:
message = process_message(message)
loss_mask_val = False if message['role'] in ('system', 'user', 'observation') else True
new_input_ids = tokenizer.apply_chat_template([message], tokenize=True, return_dict=False)[0][2:]
input_ids += new_input_ids
loss_masks += [loss_mask_val] * len(new_input_ids)

可推出结论：
    现有程序没有对agent chat 训练中的assistant生成的 tool 信息的数据进行建模和转换，部分实际训练数据如下：
      {
  "role": "user",
  "content": "Hi, I am looking for some book recommendations. I am interested in history and science fiction."
},
{
  "role": "assistant",
  "content": "{\"name\": \"get_recommended_books\", \"arguments\": {\"interests\": [\"history\", \"science fiction\"]}}"
},
{
  "role": "observation",
  "content": "{\"books\": [\"Sapiens: A Brief History of Humankind by Yuval Noah Harari\", \"A Brief History of Time by Stephen Hawking\", \"Dune by Frank Herbert\", \"The Martian by Andy Weir\"]}"
},
{
  "role": "assistant",
  "content": "Based on your interests in history and science fiction, I would recommend the following books: \"Sapiens: A Brief History of Humankind\" by Yuval Noah Harari, \"A Brief History of Time\" by Stephen Hawking, \"Dune\" by Frank Herbert, and \"The Martian\" by Andy Weir."
}

其中，assistant生成的 tool 信息为：
{
"role": "assistant",
"content": "{"name": "get_recommended_books", "arguments": {"interests": ["history", "science fiction"]}}"
},

根据GLM-4 agent chat 推理测试发现：GLM-4 生成的API任务规划调度数据格式为：functionName\n{"param":"p"}

继续追踪数据建模算法到函数apply_chat_template中：
input = self.build_single_message(
item["role"],
item.get("metadata", ""),
item["content"],
tokenize=tokenize
)
和函数：
def build_single_message(self, role, metadata, message, tokenize=True):
""" tokens of "" <|{role}|>{metadata}\n{message} """
assert role in ["system", "user", "assistant", "observation"], role
if tokenize:
role_tokens = [self.convert_tokens_to_ids(f"<|{role}|>")] + self.tokenizer.encode(f"{metadata}\n",
disallowed_special=())
message_tokens = self.tokenizer.encode(message, disallowed_special=())
tokens = role_tokens + message_tokens
return tokens
else:
return str(f"<|{role}|>{metadata}\n{message}")

可推出：
GLM-4的agent chat 训练数据建模模板为：
[gMASK]<|system|>\n tools messages<|user|>\nQ<|assistant|>\nPrefix response<|assistant|>functionName\n{"param":"p"}<|observation|>\n tool return<|assistant|>Answer

以上充分证明了：
agent chat 训练时，必须把assistant的tool信息
{
"role": "assistant",
"content": "{"name": "get_recommended_books", "arguments": {"interests": ["history", "science fiction"]}}"
}
转换为格式：
{
"role": role,
"metadata": function name,
"content": arguments {} 参数
}

解决办法：

 为了使代码整洁，在 finetune.py的process_message函数中追加代码：
# convert tarin data of agent chat.
if message['role'] == 'assistant':
    content = message['content']
    if isinstance(content, str) and content.startswith("{") and content.endswith("}"):
        try:
            content_ = eval(content)
            if isinstance(content_, dict) and "name" in content_ and "arguments" in content_:
                message['content'] = json.dumps(content_["arguments"], ensure_ascii=False)
                message['metadata'] = content_["name"]
        except:
            pass

完整函数如下：

def process_message(message):
if 'tools' in message and message['role'] == 'system':
for tool in message['tools']:
parameters = tool['function']['parameters']['properties']
tool['function']['parameters']['properties'] =
{k: v for k, v in parameters.items() if
v is not None}
elif 'tools' in message:
del message['tools']

# convert tarin data of agent chat.
if message['role'] == 'assistant':
    content = message['content']
    if isinstance(content, str) and content.startswith("{") and content.endswith("}"):
        try:
            content_ = eval(content)
            if isinstance(content_, dict) and "name" in content_ and "arguments" in content_:
                message['content'] = json.dumps(content_["arguments"], ensure_ascii=False)
                message['metadata'] = content_["name"]
        except:
            pass
return message

效果：成功完成了agent chat 多轮对话工具调用的训练。