ai model report

Is NbAiLab/nb-wav2vec2-1b-nynorsk safe to load?

CLEAN risk. 6 of 105 tracked dependencies carry advisories, 1 historically exposed. Driven by pulls num2words unpinned; it had malicious releases that the registry no longer serves.

View on HuggingFace

// posture
advisories 100provenance 75resolution 100xyz safety 50load safety 60supply chain 70scanner trust 100code execution 70
fig. 01 — posture on eight axes, 100 is clean. Weakest: xyz safety at 50.
// the finding
5.0xyz score out of 10CLEAN

LOWShips safetensors plus 1 PyTorch pickle weight file(s) (pytorch_model.bin) that Hugging Face loaders (transformers, diffusers, PEFT) load instead of the safetensors when a consumer passes use_safetensors=False; and 1 pickle checkpoint(s) for other tooling (language_model/5gram.bin); from_pretrained never loads these, even with use_safetensors=False, but code that torch.load()s them directly (a reference implementation, a training script) executes pickle.

2 more findings on this model.

Author
NbAiLab
Task
automatic-speech-recognition
Loader
transformers
License
apache-2.0
Downloads
1,147,783
Remote code
not required
// dependencies

19 direct, 14 with findings, 94 reachable, 1 historically exposed

PackageEvidenceAdvisoriesWorst
num2wordshistorically exposeddirect, runtime1CRITICALpast releases
nltkdirect, runtime52none pinned
Create a free accountto see all 105 dependencies, which versions are pinned, and what to change.
// what changed
  • HIGHrisk level raised (unknown to high)2026-09-09

Risk computed 2026-11-06 at 20:44 UTC. Dependencies come from the model's requirements, its declared loader and the code it ships; advisories from NVD, GHSA and OSV.

Model risk is computed from the model's declared and observed Python dependencies and the code it ships. If a finding is wrong, tell us.