modeldiffr verify¶
Check a report for consistency and, optionally, against the real model files.
usage: modeldiffr verify [-h] [--models] [--suite-data SUITE_DATA] [--json] [--key KEY]
[--identity IDENTITY] [--issuer ISSUER] [--staging] [--rpc RPC] report
Arguments¶
report¶
string, required
Path to a report.json written by modeldiffr audit.
Options¶
--models¶
flag
Also hash both models' weights and compare them with the roots in the report. Hub models are fetched at the recorded commit. Needs the hf extra.
--suite-data¶
string
Rebuild a dataset suite to check its root: hub downloads the pinned datasets, a folder uses its TASK.jsonl files. Without it, dataset suites are skipped and built in suites are always checked.
--key, --identity, --issuer, --staging, --rpc¶
Check signatures and anchors stored next to the report: --key for report.dsse.json, --identity and --issuer for report.sigstore.json and for model.sig in the models, --rpc for attestations in anchors.json. See Signing and publishing commands.
--json¶
flag
Print the result as JSON instead of a table:
{
"ok": true,
"audit_id": "mdr1:71e1...",
"checks": [
{"name": "schema", "status": "pass", "detail": "valid against report schema 0.2"}
]
}
Each check has status pass, fail or skip. A check is skipped when the report has nothing to check, for example a custom suite whose items are not bundled, or a backend that does not report weight hashes.
Exit codes¶
| Code | Meaning |
|---|---|
0 |
Every check passed or was skipped |
2 |
The file could not be read or is not JSON |
4 |
At least one check failed |
See Verify a report for examples and Fingerprints and proof for what each check means.