Datasets:

---

nvidia

/

OpenH-RF

Like 60

Follow

NVIDIA 70k

Five HDF5 acquisitions are marked unsafe by ClamAV signature Win.Exploit.CVE_2016_4332_1-6028013-0. Investigation located the matching sequences inside Blosc-compressed ultrasound sample chunks, supporting false positives rather than malformed HDF5 object-header messages.

This PR losslessly recompresses only tracks/track_0/data/raw_data in those five files, switching Blosc from bitshuffle to byte shuffle. It retains Zstandard, compression level 7, original chunk dimensions, data types, dataset values, attributes, and paths. Other dataset filters are unchanged. The September 26 Vanderbilt subject-ID correction is preserved.

Validation:

• Native zea 0.1.7: all five originals and replacements pass File.validate_spec(); all 129 raw-data frames read through file.data.raw_data match original values byte-for-byte. Strided slice reads also match. The existing Vanderbilt warning about missing time_to_next_transmit is present in originals and replacements.
• All dataset values compared against the current source files; numeric values checked byte-for-byte after decompression.
• All attributes, HDF5 data types, shapes, maxshapes, chunk sizes, fill values, and link kinds/destinations compared.
• Source SHA-256 hashes verified against the Hub, then rechecked before this commit.
• All five replacement files scan clean locally with ClamAV 1.5.4 and daily database 28139 (main 63, bytecode 339); no reported scan errors or exceeded-limit alerts.
• Scans use --allmatch --max-filesize=1900M --max-scansize=2000M --max-scantime=0 --alert-exceeds-max=yes.

All five replacement files passed local ClamAV scans. Decompressed ultrasound values are unchanged, and beamforming tests showed no visible image degradation. Hugging Face’s scan results remain pending and should be checked before merging.

Beamforming validation (2026-10-01):

• Ran the unmodified Vanderbilt and UBC Module C reconstruct.py entry points from open-h/OpenH-RF commit 7862883, overriding only input/output locations and selecting CPU. A wrapper captured pipeline outputs.
• Used the repository's frozen uv.lock environment: zea 0.1.7, JAX 0.10.2, macOS ARM64 CPU. Published image grids and reconstruction settings were retained.
• Frame 0 of all five originals and all five replacements reconstructed successfully, twice per input (20 reconstructions). This is a first-frame image regression check, not a full-cineloop beamforming test. All 129 raw frames were separately verified byte-for-byte as described above.
• Side-by-side visual inspection found no visible changes in sector geometry, structures, contrast, or speckle. Three pairs matched exactly; the other two differed by at most 0.00001145 dB, within observed same-file repeat variability. Nonfinite masks matched exactly; Vanderbilt's NaNs are the reference scan converter's fill outside the sector. UBC output is entirely finite.