lightonai/fever-decontaminated

fever (Decontaminated)

A decontaminated version of the fever dataset from the BEIR benchmark, with samples found in the mgte-en pre-training dataset removed.

Decontamination methodology

Contamination was detected using a two-pass approach against the full mgte-en dataset (484 GB, 1,235 parquet files):

Pass 1: Exact hash matching

All texts (queries and corpus documents) were normalized (lowercased, unicode NFKD, whitespace collapsed) and hashed with xxHash-64. The same normalization + hashing was applied to every query and document field in mgte-en. Any sample whose hash appeared in mgte-en was flagged as contaminated.

Pass 2: 13-gram containment (GPT-3 style)

Following the methodology introduced in the GPT-3 paper (Brown et al., 2020), word-level 13-grams were extracted from all remaining samples. For each sample, containment was computed as:

containment = |ngrams_in_sample ∩ ngrams_in_mgte| / |ngrams_in_sample|

Samples with containment >= 0.5 were flagged as near-duplicates.

Qrels filtering

Relevance judgments (qrels) referencing any removed query or corpus document were also removed.

Decontamination results

Component	Original	Clean	Removed
Corpus	5,416,568	5,117,452	299,116
Queries	123,142	122,842	300

Qrels per split

Split	Original	Clean	Removed
test	7,937	7,806	131
train	140,085	138,310	1,775
validation	8,079	7,908	171

Usage

from datasets import load_dataset

corpus = load_dataset("lightonai/fever-decontaminated", "corpus", split="corpus")
queries = load_dataset("lightonai/fever-decontaminated", "queries", split="queries")

Citation

Please cite the original BEIR benchmark:

@inproceedings{thakur2021beir,
  title={BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models},
  author={Thakur, Nandan and Reimers, Nils and Rücklé, Andreas and Srivastava, Abhishek and Gurevych, Irena},
  booktitle={NeurIPS Datasets and Benchmarks},
  year={2021}
}

License

MIT (same as original BEIR)