Hugging Face's logo

Datasets:
HuggingFaceFW
/
finepdfs

Tasks:
Text Generation
Modalities:
Tabular
Text
Formats:
parquet
Languages:
Arifama-Miniafia
Ankave
Abau
Size:
100M - 1B
ArXiv:
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files
xet
Community
20

Decontamination against benchmarks?

#11
by jo-kn - opened
Discussion
jo-kn

First, thank you for the valuable work!
Was the training data decontaminated (e.g. via minhash) against the test benchmarks that you used here?

hynky
 FineData org

Hi, no, we haven't run the benchmark decontamination. Due to the scale of the data, we don't believe it could play an influential role. Additionally, the domain is PDF, and most of the benchmarks we used are sourced from web data sources.

jo-kn

Unfortunately, I do think it could make a difference. Just search for GSM8k test questions in pdf files via Google, e.g.: https://www.google.com/search?q=filetype%3Apdf+Janet%E2%80%99s+ducks+lay+16+eggs+per+day.

jo-kn

(Btw, for links for MMLU-Redux, GSM8K, XStoryCloze, and CommonsenseQA which you used here seem to be incorrect or not valid anymore.)

Sign up or log in to comment