17,273 truncated shards in olmocr_science_pdfs-* (zstd "premature end")
Hi, thanks for releasing dolma3_pool! While processing a full download of data/olmocr_science_pdfs-* we found 17,273 shards (~551 GiB) whose zstd stream is truncated. zstd -t / zstd -dc fails with Read error (39) : premature end, and the last JSON line is cut off mid-document. This looks different from #4, which covered 6 common_crawl-* shards that have since been fixed.
It's not a download problem. All 17,273 local files have exactly the size the Hub reports. A random sample of 30 is byte-identical to the LFS SHA-256 on main (revision 6462556697df1a8f5c953727e9c686629ad98b68).
The pattern looks like a processing/upload batch issue. In most affected folders the first 1,000 shards in sorted order are fine and every shard after them is truncated:
| folder | truncated / total | first truncated shard |
|---|---|---|
olmocr_science_pdfs-education_and_jobs |
4479 / 5479 | shard_00012968 (first 1000 OK) |
olmocr_science_pdfs-science_math_and_technology-part2 |
4039 / 5039 | shard_00034605 (first 1000 OK) |
olmocr_science_pdfs-health |
3529 / 4529 | shard_00013140 (first 1000 OK) |
olmocr_science_pdfs-crime_and_law |
1619 / 1644 | shard_00000485 (first 25 OK) |
olmocr_science_pdfs-finance_and_business |
1314 / 2314 | shard_00012817 (first 1000 OK) |
olmocr_science_pdfs-literature |
1159 / 1159 | all |
olmocr_science_pdfs-politics |
608 / 1608 | shard_00013651 (first 1000 OK) |
olmocr_science_pdfs-software_development |
433 / 1433 | shard_00011827 (first 1000 OK) |
olmocr_science_pdfs-industrial |
91 / 1091 | shard_00013968 (first 1000 OK) |
olmocr_science_pdfs-art_and_design |
1 / 285 | shard_00005885.jsonl.zst only |
olmocr_science_pdfs-science_math_and_technology-part1 |
1 / 10000 | shard_00032494.jsonl.zst only |
All other olmocr_science_pdfs-* folders decompress cleanly.
How we found it: every shard was decompressed end to end with the zstd CLI. The affected files still contain a few thousand complete documents each. Because the stream simply ends, there's no way to tell how many documents are missing after the cut.
To check a folder:
for f in data/olmocr_science_pdfs-literature/*.jsonl.zst; do zstd -tq "$f" || echo "$f"; done
Could you take a look and re-upload the affected shards? We're happy to share the full list of 17,273 paths if that helps.