Datasets:
4-digit Numbers/Charaters in the Doucment Beginning
Hi HPLT Team,
Thank you for your efforts in creating and sharing the HPLT dataset—it has been a valuable resource for our work.
While reviewing the data, I noticed that some documents begin with seemingly random four-digit numbers or character sequences, as shown in the examples below. I’m wondering whether these strings carry any specific meaning or if they might be artifacts introduced during preprocessing.
Clarifying this would be very helpful for our use case. Looking forward to your response—thanks again!
Best regards,
Liping
Hi Liping,
Thank you for your feedback and for pointing our attention to this issue! We saw this issue before and discovered that these tokens come from the chunked HTTP responses in the InternetArchive crawls (they are chunk sizes, see https://en.wikipedia.org/wiki/Chunked_transfer_encoding for more details). Our pipeline does not correctly handle chunked HTTP responses now, but we plan to fix this in the next HPLT data releases.
Such responses are most frequent in ArchiveBot (all examples on your screenshots are from ArchiveBot) and less frequent in other Internet Archive crawls. CommonCrawl does its own dechunking so HPLT documents coming from CommonCrawl should not contain this issue. If you find these tokens harmful for your use case, consider filtering by the 'collection' field:
- including only HPLT documents coming from CommonCrawl (collections starting with 'cc') should solve this issue, but you will loose about 40% of documents;
- excluding HPLT documents coming from ArchiveBot (collection==archivebot) should significantly reduce the proportion of these tokens while removing a tiny portion of documents only.