This directory contains the test set portion of our CriticAgent benchmark, designed for evaluating reward models for agents capable of tool use and reasoning. Please note that only the test set is being publicly released. The training set is not included in this release.
The test set consists of 5,000 annotated steps. Each step follows the structure outlined below:
{
"traj_id": <string, the id of the step's original trajectory>… See the full description on the dataset page:
https://huggingface.co/datasets/CriticAgent/Submission.
