This link gathers 72 noisy versions of the MS-Marco-Passage Ranking dataset consisting of three noise types (insertion, deletion, substitution), two different distributions of errors in the text (Batch 1 where errors are distributed in few words in the text and Batch2 where errors are more evenly spread out between words) and 12 different intensities of noise (CER varying from 3% to 36% with intervals of 3%).
The exact dataset that has been used is the MS-Marco-passagetest2020-top1000. The… See the full description on the dataset page:
https://huggingface.co/datasets/edwardgiamphy/Noisy-MSMARCO-Passage-Ranking.
