ouyangyouzhang/QueryRewrite
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Qwen3-32B JSON格式化输出不稳定，微调输出格式

Browse Source
This commit is contained in:
ouyangyouzhang
2025-08-28 09:04:29 +08:00
parent cdaf6870d2
commit 660b5f6561
4 changed files with 110 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rag2_0/intent_recognition/DataModels.py
+41
View File
@@ -107,9 +107,34 @@ class Classification(BaseModel):
vertical_classification:str = Field(description="垂直领域一级分类")
sub_classification:str = Field(description="一级分类下的二级分类")
@classmethod
def get_format_instructions(cls):
return """
格式如下,必须严格以纯JSON格式输出
{
"vertical_classification": "垂直领域一级分类",
"sub_classification": "一级分类下的二级分类"
}
字段说明：
vertical_classification 类型：str 描述：垂直领域一级分类
sub_classification 类型：str 描述：一级分类下的二级分类
"""
class QueryRewrite(BaseModel):
rewrite:str = Field(description="问题改写")
@classmethod
def get_format_instructions(cls):
return """
格式如下：必须严格以纯JSON格式输出
{
"rewrite": "问题改写"
}
字段说明：
rewrite 类型：str 描述：问题改写之后的内容
"""
##########################槽位模型###########################
class SlotBase(BaseModel):
"""槽位基础模型"""
@@ -312,6 +337,22 @@ class StepBackPrompt(BaseModel):
can_use_back_prompt: bool = Field(description="原始查询是否可以进行后退提示(true/false),如果原始查询没有限定词或其他限定词语，则不能进行后退提示")
step_back_query: List[str] = Field(description="后退提示生成的抽象查询(多个)")
@classmethod
def get_format_instructions(cls):
return """
格式如下，必须严格以纯JSON格式输出
{
"original_query": "原始查询",
"can_use_back_prompt": "原始查询是否可以进行后退提示(true/false),如果原始查询没有限定词或其他限定词语，则不能进行后退提示",
"step_back_query": "后退提示生成的抽象查询(多个)"
}
字段说明：
original_query 类型：str 描述：用户输入的原始查询
can_use_back_prompt 类型：bool 描述：原始查询是否可以进行后退提示(true/false),如果原始查询没有限定词或其他限定词语，则不能进行后退提示
step_back_query 类型：List[str] 描述：后退提示生成的抽象查询(多个)
"""
class FollowUpQuestions(BaseModel):
"""后续问题数据模型"""
original_query: str = Field(description="原始查询")
rag2_0/intent_recognition/IntentRecognition.py
+24 -10
View File
@@ -188,17 +188,19 @@ class AsyncIntentRecognizer:
classification_parser = PydanticOutputParser(pydantic_object=Classification)
formatted_prompt = classification_prompt.format(user_input=query,
classification_info=classification_info,
output_format=classification_parser.get_format_instructions(),
output_format=Classification.get_format_instructions(),
# conversation_context=conversation_context,
chat_history=json.dumps(chat_history, ensure_ascii=False))
# 解析输出
try:
# 异步调用LLM
response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
response = await self._llm.ainvoke(formatted_prompt, response_format={"type": "json_object"}, extra_body={"enable_thinking": False})
# response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
# 尝试直接解析JSON响应
response.content = response.content.strip()
clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
clean_output = re.sub(r'\s+', '', clean_output)
parsed_output = classification_parser.parse(clean_output)
# 计算并打印耗时
@@ -262,11 +264,13 @@ class AsyncIntentRecognizer:
formatted_prompt = formatted_prompt.replace("{output_format}", terms_list_parser.get_format_instructions())
# 异步调用LLM
response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
response = await self._llm.ainvoke(formatted_prompt, response_format={"type": "json_object"}, extra_body={"enable_thinking": False})
# response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
# 尝试使用Pydantic解析器解析TermList
response.content = response.content.strip()
clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
clean_output = re.sub(r'\s+', '', clean_output)
parsed_output = terms_list_parser.parse(clean_output)
return parsed_output.terms
@@ -335,18 +339,22 @@ class AsyncIntentRecognizer:
</chat_history>
1、请从当前提问内容中提取电力造价行中定额编码、定额名称、清单编码、清单名称
2、请勿随机编造，如果没有提取到，返回空内容
3、返回结果为json格式
2、请勿随机编造，如果没有提取到内容返回空的JSON
3、返回结果为json格式,必须严格以纯JSON格式输出
```json
{{
"dinge_info_list":{{"dinge_code_list":["xxxx","xxxx"], "dinge_name_list":["xxxx","xxxx"]}},
"qingdan_info":{{"qingdan_code_list":["xxxx","xxxx"], "qingdan_name_list":["xxxx","xxxx"]}}
}}
```json
"""
try:
response = await self._llm.ainvoke(prompt, response_format={"type": "json_object"}, extra_body={"enable_thinking": False})
# response = await self._llm.ainvoke(prompt, extra_body={"enable_thinking": False})
response.content = response.content.strip()
clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
clean_output = re.sub(r'\s+', '', clean_output)
parsed_output = JsonOutputParser().parse(clean_output)
# 计算并打印耗时
@@ -378,16 +386,18 @@ class AsyncIntentRecognizer:
keywords_str = json.dumps(terms_dict, ensure_ascii=False)
query_rewrite_parser = PydanticOutputParser(pydantic_object=QueryRewrite)
formatted_prompt = query_rewrite_prompt_pro.format(query=query,
output_format=query_rewrite_parser.get_format_instructions(),
output_format=QueryRewrite.get_format_instructions(),
keywords=keywords_str,
chat_history=chat_history,
context=context)
# 解析输出
try:
# 异步调用LLM
response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
response = await self._llm.ainvoke(formatted_prompt,response_format={"type": "json_object"}, extra_body={"enable_thinking": False})
# response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
response.content = response.content.strip()
clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
clean_output = re.sub(r'\s+', '', clean_output)
parsed_output = query_rewrite_parser.parse(clean_output)
end_time = time.time()
process_time=end_time-start_time
@@ -630,9 +640,11 @@ class AsyncIntentRecognizer:
)
try:
# 异步调用LLM
response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
response = await self._llm.ainvoke(formatted_prompt,response_format={"type": "json_object"}, extra_body={"enable_thinking": False})
# response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
response.content = response.content.strip()
clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
clean_output = re.sub(r'\s+', '', clean_output)
# 尝试解析LLM响应
parsed_output = slot_parser.parse(clean_output)
return parsed_output
@@ -660,16 +672,18 @@ class AsyncIntentRecognizer:
query=query,
chat_history=json.dumps(chat_history, ensure_ascii=False) if chat_history else "[]",
# conversation_context=conversation_context,
output_format=step_back_parser.get_format_instructions()
output_format=StepBackPrompt.get_format_instructions()
)
try:
# 异步调用LLM
response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
response = await self._llm.ainvoke(formatted_prompt, response_format={"type": "json_object"}, extra_body={"enable_thinking": False})
# response = await self._llm.ainvoke(formatted_prompt, extra_body={"enable_thinking": False})
# 解析输出
response.content = response.content.strip()
clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
clean_output = re.sub(r'\s+', '', clean_output)
parsed_output = step_back_parser.parse(clean_output)
step_back_end_time = time.time()
step_back_time = step_back_end_time - step_back_start_time
rag2_0/intent_recognition/PromptTemplates.py
+2 -2
View File
@@ -298,7 +298,7 @@ step_back_prompt = """
后退问题:
{{
"original_query": "2023版本如何在Windows 11系统上导入单位工程量清单？",
"can_use_back_prompt": True,
"can_use_back_prompt": true,
"step_back_query": ["如何在Windows 11系统上导入单位工程量清单？", "如何导入单位工程量清单？"]
}}
@@ -306,7 +306,7 @@ step_back_prompt = """
后退问题:
{{
"original_query": "某个设备更换后，如何在系统中更新对应的定额？",
"can_use_back_prompt": True,
"can_use_back_prompt": true,
"step_back_query": ["如何更新设备对应的定额？", "如何更新定额？"]
}}