A decontaminated version of the nq dataset from the BEIR benchmark, with samples found in the mgte-en pre-training dataset removed.
Contamination was detected using a two-pass approach against the full mgte-en dataset (484 GB, 1,235 parquet files):
All texts (queries and corpus documents) were normalized (lowercased, unicode NFKD, whitespace collapsed) and hashed with xxHash-64. The… See the full description on the dataset page:
https://huggingface.co/datasets/lightonai/nq-decontaminated.