99 lines
4.5 KiB
Python
99 lines
4.5 KiB
Python
# -- encoding: utf-8 --
|
||
|
||
"""
|
||
Description: 基于LLM通过用户设置维度和相应描述进行QA对评估
|
||
Create: 2023/11/7 9:26
|
||
"""
|
||
import json
|
||
import re
|
||
import time
|
||
from pathlib import Path
|
||
from typing import List, Dict, Any
|
||
|
||
from loguru import logger
|
||
|
||
|
||
from deepdata.core.base_op import LLM
|
||
|
||
|
||
class QAConditionEvaluator(LLM):
|
||
def __init__(self, *args, **kwargs):
|
||
super(QAConditionEvaluator, self).__init__(*args, **kwargs)
|
||
self.pattern = r'结果[:：] ?[YN]'
|
||
self.template_path = Path(__file__).parent / "resources/template.txt"
|
||
self.examples_path = Path(__file__).parent / "resources/examples.json"
|
||
self.task_id = kwargs.get("taskId", "default_id")
|
||
self.dimensions = kwargs.get("dimension", [
|
||
{
|
||
"dimension": "回答是否有针对性",
|
||
"description": "回答应对问题中的所有疑问点提供正面、直接的回答，"
|
||
"不应引起疑惑。同时，答案不应有任何内容的遗漏，需构成一个完整的陈述。"
|
||
},
|
||
{
|
||
"dimension": "问题是否独立",
|
||
"description": "仅分析问题，问题的主体和客体都比较明确，即使有省略，也符合语言习惯。"
|
||
"在不需要补充其他信息的情况下不会引起疑惑。"
|
||
},
|
||
{
|
||
"dimension": "语法是否错误",
|
||
"description": "问题为疑问句，答案为陈述句; 不存在词语搭配不当的情况;连接词和标点符号不存在错用情况；"
|
||
"逻辑混乱的情况不存在；语法结构都正确且完整;"
|
||
}
|
||
])
|
||
|
||
self.llm = self.get_llm(*args, **kwargs)
|
||
self.prompts = self.build_llm_prompt(*args, **kwargs)
|
||
|
||
@staticmethod
|
||
def _process_examples(dimension_example: List) -> str:
|
||
if not dimension_example:
|
||
return "\n"
|
||
res = "\n以下是一些案例供你参考："
|
||
for single_example in dimension_example:
|
||
res += (f"\n问题：{single_example['question']}"
|
||
f"\n回答：{single_example['answer']}"
|
||
f"\n分析思路：{single_example['evaluate']}"
|
||
f"\n结果：{single_example['result']}\n")
|
||
return res
|
||
|
||
def execute(self, sample: Dict[str, Any]) -> Dict[str, Any]:
|
||
start = time.time()
|
||
qas = json.loads(sample[self.text_key])
|
||
single_content_res = []
|
||
for qa in qas:
|
||
single_qa_res = []
|
||
for dimension, prompt in self.prompts.items():
|
||
local_result = self._llm_call_parse(qa, prompt, retry=2)
|
||
single_qa_res.append({"dimension": dimension, "result": local_result})
|
||
qa_response = {"qaId": qa["qaId"], "result": single_qa_res}
|
||
single_content_res.append(qa_response)
|
||
|
||
sample[self.text_key] = "Sucess"
|
||
self.save_sample(single_content_res, sample)
|
||
cost_time = time.time() - start
|
||
logger.info(f"task id: {self.task_id}, method: QAConditionEvaluator costs {cost_time:.6f} s")
|
||
return sample
|
||
|
||
def build_llm_prompt(self, *args, **kwargs) -> Dict:
|
||
templates = self.template_path.read_text(encoding="utf-8")
|
||
examples_dict = json.loads(self.examples_path.read_text(encoding="utf-8"))
|
||
prompts_dict = {}
|
||
for dimension in self.dimensions:
|
||
name, des = dimension["dimension"], dimension["description"]
|
||
dimension_example = self._process_examples(examples_dict.get(name))
|
||
dimension_prompt = templates.format(criterion=des, examples=dimension_example, question="{question}",
|
||
answer="{answer}")
|
||
prompts_dict[name] = dimension_prompt
|
||
return prompts_dict
|
||
|
||
def _llm_call_parse(self, data: Dict, prompt: str, retry: int = 2):
|
||
try:
|
||
for _ in range(retry):
|
||
response = self.llm(prompt.format(question=data["question"], answer=data["answer"]))
|
||
result = re.findall(self.pattern, response)
|
||
if result:
|
||
return "Y" in result[0]
|
||
except RuntimeError as e:
|
||
logger.error(f"method: QAConditionEvaluator execution error, cause by {e}")
|
||
return False
|