Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation
Our work studies the ratio between signal, a benchmark's ability to separate models; and noise, a benchmark's sensitivity to random variability during training steps.
This dataset contains evaluation results. For utilites to use this dataset and to reproduce the findings in our paper, please see our github.