[Discussion] File count and per-file size audit of ABC-130k at revision 29136bc9 vs. usedStorage
Posting this as a discussion, not a bug report — just sharing findings from a migration audit in case it helps others.
While migrating the dataset to an internal mirror, I noticed the repo's reported storage size doesn't match the sum of file sizes in the current main revision. Here are the numbers, all pulled from the HF API on 2026-07-06 against revision 29136bc9b9e38d320b00ffcddbbe4cd0e3278c58 (main, lastModified 2026-07-02):
Sources
- Repo metadata (GET /api/datasets/XDOF/ABC-130k): usedStorage = 22,060,544,378,693 bytes (~22.06 TB)
- Tree API, walked every task dir under data/train and data/val with recursive=true, summed size for each type == "file" entry: 8,505,284,694,384 bytes (~8.51 TB)
Summary
| Metric | Bytes | TB |
|---|---|---|
| usedStorage (repo metadata) | 22,060,544,378,693 | 22.06 |
| Tree API sum (current main) | 8,505,284,694,384 | 8.51 |
| Difference | 13,555,259,684,309 | 13.56 |
Total file count (current revision, all .mcap): 56,593 files — consistent across the tree API walk and the local manifest.
All 56,593 files are .mcap (no non-mcap files in the tree).
Interpretation
usedStorage appears to reflect LFS backend storage (which can include historical/unreferenced blobs), while the tree API's size field reflects the current main snapshot. The ~13.5 TB gap is consistent with the dataset having gone through multiple revisions during upload. This is not a problem with the dataset itself — the current snapshot is internally consistent — but it's worth knowing if you're doing a size-based audit or planning storage for a mirror.
Method is reproducible: walk /api/datasets/XDOF/ABC-130k/tree/main/data/{train,val} → for each task dir, request recursive=true → sum size of file entries.
Hope this is useful to anyone else auditing or mirroring the dataset.