stacked summaries

This organization exists to test and evaluate the (potential) benefits of "task-oriented pretraining" as popularized by the FLAN-t5 series of models within the summarization NLP task.

mission statement

The idea is to apply a similar concept but adjusted to be more specific w.r.t. the summarization task. Hopefully, this will train models that actually "know" how to condense and distill meaningful information from text rather than learning some naive style transfer of "this is what the dataset summaries sound like so I will do that with essential words."

The most apparent augmentation/task is "stacking" summaries that are shorter than MAX_LENGTH_TOKENS when combined, so the model has to learn to separate and group summaries for these independent concepts.

work statement

unless otherwise noted, work here is completed by "night shift" NLP enthusiasts/researchers in their free time.