Metadata Extraction Guide

Use modelaudit metadata to inspect model metadata without running a full security scan.

Safe defaults

By default, metadata extraction does not deserialize model objects. This is safer for untrusted artifacts and should be your default workflow.

modelaudit metadata model.safetensors

Common usage

# Directory summary + per-file metadata (table output)
modelaudit metadata ./models

# JSON output for automation
modelaudit metadata ./models --format json --output metadata.json

# YAML output
modelaudit metadata model.onnx --format yaml

# Security-focused metadata fields only
modelaudit metadata model.onnx --security-only

Directory limits

Directory metadata extraction stops after 10,000 file entries, 20,000 total filesystem entries, 1 GiB of aggregate file size, or 64 directory levels. When a limit is reached, table output shows an incomplete-extraction warning. JSON and YAML output include analysis_incomplete: true, scan_outcome: inconclusive, and a budget_events list describing the exceeded limit.

Special filesystem entries such as named pipes and sockets, including symlinks to those entries, are reported as errors and are not opened or sent to scanners. Regular files and in-directory symlinks to regular files remain eligible for metadata extraction.

Trusting loaders (advanced)

--trust-loaders allows scanner metadata loaders that may deserialize model content. Only use this for trusted artifacts, and only in isolated environments.

modelaudit metadata model.pkl --trust-loaders

Notes

• modelaudit metadata supports both files and directories.
• Unknown or unsupported files are reported with format: unknown.
• For full detection coverage (CVE checks, suspicious patterns, and policy outcomes), use modelaudit scan.