LLMLingua
LLMLingua copied to clipboard
[Question]: ZeroScrolls reproduction
Describe the issue
-
The publicly available validation split only contains a small number of samples (around 20 per task), and the ground truths are not available for the test split, which is a lot bigger. Are the results in the LLMLingua paper obtained from the validation split, or did you submit your results to https://www.zero.scrolls-benchmark.com/submission?
-
The ZeroScrolls dataset provides the inputs already formatted inside the prompt-templates - did you compress the inputs from the dataset as is, or did you extract the contexts from the inputs and compress only the contexts, keeping the prompt templates intact?