Bases: BaseLLMReward, BaseListWiseReward
Context-Aware Reward Modeling
Source code in rm_gallery/gallery/rm/carmo.py
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 | class CARMO(BaseLLMReward, BaseListWiseReward):
"""Context-Aware Reward Modeling"""
def _before_evaluate(self, sample: DataSample, **kwargs) -> dict:
instruction = sample.input[-1].content
query = CriteriaGenerationPrompt.format(instruction=instruction)
response = self.llm.simple_chat(query)
principles = CriteriaGenerationPrompt.parse(response).principles
completions = [output.answer.content for output in sample.output]
return dict(
principles=principles,
instruction=instruction,
completions=completions,
)
def _after_evaluate(
self, response: RelativeEvaluationPrompt, sample: DataSample, **kwargs
) -> RewardResult:
"""
Converts LLM response to list-wise ranking metrics.
Parameters:
response (RelativeEvaluationPrompt): Parsed LLM comparison
Returns:
RewardResult: Relative ranking of responses
"""
scores = [0 for i in range(len(sample.output))]
scores[response.best - 1] = 1
return RewardResult(
name=self.name,
details=[
RewardDimensionWithRank(
name=self.name, reason=response.reason, rank=scores
)
],
)
|