Sign in to view source links and access this dataset
Description
PRMBench Preview is the official dataset for the PRMBench benchmark, authored by hitsmy and last updated on 2025-01-07. It contains 6,216 data instances designed to evaluate process-level reward models. Each instance includes a question, a solution process, and a modified process with errors for fine-grained error type identification.
Use Cases
Benchmarking reward model performance based on the annotated error types
Training error detection models based on the solution and modified process pairs
Evaluating the ability of AI systems to identify fine-grained errors in multi-step solutions
Strengths
Contains 6,216 data instances for evaluation
Specifically designed for evaluating process-level reward models
Limitations
Column-level documentation is absent; field semantics must be inferred after download
Row count is known, but other scale metrics like file size are unknown
Provenance
Source
hitsmy
Freshness
Last updated 2025-01-07 09:12:31
License is unknown; terms of use must be verified.