93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279 | class PairwiseComparisonEvaluator(BaseEvaluator):
"""对两两比较进行评估。
通过LLM评判哪个回复更好，从而评估给定问题的回复与“参考”回复的质量。
输出给定的`response`是否比`reference`回复更好。
Args:
service_context (Optional[ServiceContext]):
用于评估的服务上下文。
eval_template (Optional[Union[str, BasePromptTemplate]]):
用于评估的模板。
enforce_consensus (bool): 是否强制一致性（如果我们改变答案的顺序，则一致性）。默认为True。"""
def __init__(
self,
llm: Optional[LLM] = None,
eval_template: Optional[Union[BasePromptTemplate, str]] = None,
parser_function: Callable[
[str], Tuple[Optional[bool], Optional[float], Optional[str]]
] = _default_parser_function,
enforce_consensus: bool = True,
# deprecated
service_context: Optional[ServiceContext] = None,
) -> None:
self._llm = llm or llm_from_settings_or_context(Settings, service_context)
self._eval_template: BasePromptTemplate
if isinstance(eval_template, str):
self._eval_template = PromptTemplate(eval_template)
else:
self._eval_template = eval_template or DEFAULT_EVAL_TEMPLATE
self._enforce_consensus = enforce_consensus
self._parser_function = parser_function
def _get_prompts(self) -> PromptDictType:
"""获取提示。"""
return {
"eval_template": self._eval_template,
}
def _update_prompts(self, prompts: PromptDictType) -> None:
"""更新提示。"""
if "eval_template" in prompts:
self._eval_template = prompts["eval_template"]
async def _get_eval_result(
self,
query: str,
response: str,
second_response: str,
reference: Optional[str],
) -> EvaluationResult:
"""获取评估结果。"""
eval_response = await self._llm.apredict(
prompt=self._eval_template,
query=query,
answer_1=response,
answer_2=second_response,
reference=reference or "",
)
# Extract from response
passing, score, feedback = self._parser_function(eval_response)
if passing is None and score is None and feedback is None:
return EvaluationResult(
query=query,
invalid_result=True,
invalid_reason="Output cannot be parsed",
feedback=eval_response,
)
else:
return EvaluationResult(
query=query,
response=eval_response,
passing=passing,
score=score,
feedback=eval_response,
pairwise_source=EvaluationSource.ORIGINAL,
)
async def _resolve_results(
self,
eval_result: EvaluationResult,
flipped_eval_result: EvaluationResult,
) -> EvaluationResult:
"""解决来自评估和翻转评估的评估结果。
Args:
eval_result（EvaluationResult）：answer_1先显示时的结果
flipped_eval_result（EvaluationResult）：answer_2先显示时的结果
Returns:
EvaluationResult：最终评估结果
"""
# add pairwise_source to eval_result and flipped_eval_result
eval_result.pairwise_source = EvaluationSource.ORIGINAL
flipped_eval_result.pairwise_source = EvaluationSource.FLIPPED
# count the votes for each of the 2 answers
votes_1 = 0.0
votes_2 = 0.0
if eval_result.score is not None and flipped_eval_result.score is not None:
votes_1 = eval_result.score + (1 - flipped_eval_result.score)
votes_2 = (1 - eval_result.score) + flipped_eval_result.score
if votes_1 + votes_2 != 2: # each round, the judge can give a total of 1 vote
raise ValueError("Impossible score results. Total amount of votes is 2.")
# get the judges (original and flipped) who voted for answer_1
voters_1 = [eval_result] * (eval_result.score == 1.0) + [
flipped_eval_result
] * (flipped_eval_result.score == 0.0)
# get the judges (original and flipped) who voted for answer_2
voters_2 = [eval_result] * (eval_result.score == 0.0) + [
flipped_eval_result
] * (flipped_eval_result.score == 1.0)
if votes_1 > votes_2:
return voters_1[0] # return any voter for answer_1
elif votes_2 > votes_1:
return voters_2[0] # return any vote for answer_2
else:
if (
eval_result.score == 0.5
): # votes_1 == votes_2 can only happen if both are 1.0 (so actual tie)
# doesn't matter which one we return here
return eval_result
else: # Inconclusive case!
return EvaluationResult(
query=eval_result.query,
response="",
passing=None,
score=0.5,
feedback="",
pairwise_source=EvaluationSource.NEITHER,
)
async def aevaluate(
self,
query: Optional[str] = None,
response: Optional[str] = None,
contexts: Optional[Sequence[str]] = None,
second_response: Optional[str] = None,
reference: Optional[str] = None,
sleep_time_in_seconds: int = 0,
**kwargs: Any,
) -> EvaluationResult:
del kwargs # Unused
del contexts # Unused
if query is None or response is None or second_response is None:
raise ValueError(
"query, response, second_response, and reference must be provided"
)
await asyncio.sleep(sleep_time_in_seconds)
eval_result = await self._get_eval_result(
query, response, second_response, reference
)
if self._enforce_consensus and not eval_result.invalid_result:
# Flip the order of the answers and see if the answer is consistent
# (which means that the score should flip from 0 to 1 and vice-versa)
# if not, then we return a tie
flipped_eval_result = await self._get_eval_result(
query, second_response, response, reference
)
if not flipped_eval_result.invalid_result:
resolved_eval_result = await self._resolve_results(
eval_result, flipped_eval_result
)
else:
resolved_eval_result = EvaluationResult(
query=eval_result.query,
response=eval_result.response,
feedback=flipped_eval_result.response,
invalid_result=True,
invalid_reason="Output cannot be parsed.",
)
else:
resolved_eval_result = eval_result
return resolved_eval_result
|