Skip to main content

​
mellea.backends.process_reward_models.huggingface.prms

PRM Implementations for Local HuggingFace Backends.

​
Classes

​
HFGenerativePRM

A Generative PRM that works with a huggingface backend. Methods:

​
score

score(self, query: str, response: str) -> tuple[list[float], list[list[float]]]
Returns a final and per-step score for a given input query and response. Args:
  • query: User query
  • response: Assistant Response to score

​
prepare_inputs

prepare_inputs(self, user_content: str, steps: list[str]) -> BatchEncoding
Prepare the inputs for inference with the model. Args:
  • user_content: the user query
  • steps: assistant response, broken down into steps

​
HFRegressionPRM

A Regression PRM that works with a huggingface backend. Methods:

​
score

score(self, query: str, response: str) -> tuple[list[float], list[list[float]]]
Returns a final and per-step score for a given input query and response. Args:
  • query: User query
  • response: Assistant Response to score

​
prepare_inputs

prepare_inputs(self, user_content: str, steps: list[str]) -> BatchEncoding
Prepare the inputs for inference with the model. Args:
  • user_content: the user query
  • steps: assistant response, broken down into steps