Claire21/genomefold-training-data

GenomeFold hg38 training data

This repository contains the processed hg38 input archives used by GenomeFold. It includes CTCF, CUTLL1, GM12878, gm12878, H1-hESC, HFFc6, HepG2, IMR90, imr90, Jurkat, K562, Tcell, DNA-sequence, and reference inputs.

The 15 original .tar.zst archives total 82,377,711,432 bytes (76.720 GiB). They are stored under data/. Evaluation outputs are not duplicated in this Dataset repository.

Archive layout

The archives are uploaded as complete files. Zenodo-specific .chunkNNN and .partNNN byte fragments are not used here.

The names K562_part01 and K562_part02 identify two complete logical K562 archives created by the original packaging process; they are not byte fragments that need to be concatenated.

Each .tar.zst has an adjacent .sha256 sidecar. The sidecars retain the original server paths, so verification after download should use the stored digest with the downloaded basename:

cd data

for checksum_file in *.tar.zst.sha256; do
  archive=${checksum_file%.sha256}
  expected=$(awk '{print $1}' "$checksum_file")
  printf '%s  %s\n' "$expected" "$archive" | sha256sum -c -
done

Download

hf download Claire21/genomefold-training-data \
  --repo-type dataset \
  --local-dir genomefold-training-data

After checksum verification, extract an archive with:

tar --use-compress-program=unzstd \
  -xf gf_training_data__input_data_hg38_CUTLL1.tar.zst

or:

zstd -dc gf_training_data__input_data_hg38_CUTLL1.tar.zst | tar -xf -