Datasets:

---

yunplus

/

SuperWing

Like 7

Summary

data_vol/data_vol.31.npy.zst and data_vol/data_vol.32.npy.zst are byte-for-byte identical. One of them appears to be an unintentional duplicate, and the geometry range that should follow shard 31 (geom 3122..3203, 82 geometries / 562 samples) is entirely absent from the dataset.

Evidence

Both files are content-addressable and yield the same LFS SHA-256 and the same byte size:

The SHA-256 hashes match exactly, so the two files are identical at the byte level. Decompressing either one produces the same volume array (84 geometries, 575 samples, geom 3038..3121).

Impact

Summing the per-shard sample counts of the 46 data_vol.*.npy.zst files gives 28,869 rows, while index.npy has 28,856 rows (a difference of +13).

Matching the shards against index.npy shows:

• data_vol.31 = data_vol.32 = geom 3038..3121 (duplicate block),
• then the next shard data_vol.33 starts at geom 3204,
• so the 82 geometries geom 3122..3203 (562 samples) are missing from the dataset.

Net effect: +575 (duplicate shard) - 562 (absent block) = +13, which matches the 28,869 vs 28,856 discrepancy. This breaks any consumer that assumes the shards tile a contiguous geometry range (e.g. cumulative-offset slicing against index.npy).

Reproduction

curl -s "https://huggingface.co/api/datasets/yunplus/SuperWing/tree/main/data_vol?expand=true" \
  | jq -r '.[] | select(.path | endswith("npy.zst")) | [.path, .size, .lfs.oid] | @tsv'

The two rows for data_vol.31.npy.zst and data_vol.32.npy.zst share the same size and lfs.oid.

Suggestion

Could you re-upload the correct data_vol.32.npy.zst (the block that should carry geom 3122..3203), or confirm the intended shard layout? Happy to help verify once fixed.