README.md

metadata

license: mit

AstroClassification and Redshifts Pretraining Dataset

This dataset was used in the masked autoencoding pretraining step for the AstroClassification and Redshifts tasks introduced in Connect Later: Improving Fine-tuning for Robustness with Targeted Augmentations. This is a dataset of simulated astronomical time-series (e.g., supernovae, active galactic nuclei), and the task is to classify the object type (AstroClassification) or predict the object's redshift (Redshifts).

• Repository: https://github.com/helenqu/connect-later
• Paper: will be updated
• Point of Contact: Helen Qu (helenqu@sas.upenn.edu)

Dataset Loading

The load_dataset HuggingFace dataset loading syntax will use the custom dataset loading script included in this repository to load this dataset in the expected format for pretraining with the run_pretrain.sh script in the Connect Later repository.