Search Results for "benchmark testing"

ARC-AGI-1 Task Generator creates new ARC-style reasoning tasks whose distribution is designed to resemble the public ARC-AGI-1 evaluation set. It provides fresh problems that models are less likely to have encountered during training or previous evaluation. The project is intended to complement public benchmark scores by testing whether reasoning abilities transfer to newly generated examples. Generated tasks use the standard ARC train-and-test JSON structure. This makes the output compatible with existing ARC evaluation harnesses and analysis pipelines. The repository also includes scripts for labeling, describing, visualizing, and generating tasks, including stratified generation. ...