We include the data for replicating the academic papers associated with this repo.
All datasets retain their original license, if applicable, and are included here only for replication of academic research on neural network selective classification, calibration, out-of-distribution detection, and interpretability-by-exemplar. Consult the citations in the papers for additional details. For ease of replication, for the classification experiments, we have included the hidden states from "microsoft/Phi-3.5-mini-instruct" and "mistralai/Mixtral-8x7B-Instruct-v0.1" taken as input to the SDM estimators for the classification datasets, as well as the model outputs for comparing alternative calibration methods. As such, given the file sizes, we have split the directories across files by model and task. The word ordering data only includes the text, since it is intended for experiments with fine-tuning sequence-prediction models.
Separately, for an Apache 2.0 dataset that is suitable for end-user applications, we recommend the OpenVerification1 dataset available via Hugging Face datasets: https://huggingface.co/datasets/ReexpressAI/OpenVerification1