PRMBench
What it is
PRMBench is a comprehensive benchmark and toolkit designed for evaluating process-level reward models (PRMs). It includes 'mr_eval' for automated evaluation and 'mr_annotate' for data generation and annotation.
The problem it solves
Standard reward models often focus on outcome accuracy, which can lead to 'reward hacking' or models that arrive at the correct answer through flawed reasoning.
What you could use it for
- Evaluating the reasoning steps of Large Language Models (LLMs) rather than just final answers.
- Benchmarking process-level reward models (PRMs) on complex math and reasoning tasks.
- Generating and annotating datasets for training reward models.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.