Every sample appears twice in the extracted .jsonl files

#1
by tgraber4 - opened

After downloading with thrember.download_dataset(..., split="all", file_type="all"), the .jsonl files contain
about twice the documented number of records.

  • Expected: 3,238,315 (2,626,000 train + 606,000 test + 6,315 challenge)
  • Observed: 6,470,315 records in 448 files. Every weekly file has exactly 2x the documented count (e.g. 60,000
    Win32 vs. 30,000).
  • Affected: all six file types (Win32, Win64, Dot_Net, APK, ELF, PDF). The challenge set is not affected (exactly
    6,315).
  • What I checked: for the first train week of each type, every sha256 appears exactly twice, and the unique
    counts match the README. The two copies are either identical or differ only in the caps, ttps and mbc fields
    (one copy filled, the other empty). APK and PDF copies are always identical.
  • Cause: I believe the duplicates come from the HuggingFace data and not the download code, which only downloads
    and extracts each zip.

Sign up or log in to comment