How do I export a Giskard scan?

`results = giskard.scan(model, dataset)` returns a report object; `results.to_html('giskard.html')` writes a standalone page with every detected issue and its examples.

What does the scan find?

Performance bias across slices, robustness failures under perturbation, overconfidence, spurious correlations and — for LLMs — prompt injection and harmfulness probes, depending on the model type.

Whoever signs off on the model: a risk function, a product owner, a clinician, a compliance reviewer. None of them are going to open your notebook, which is why the export matters.

Yes — run the scan in the training pipeline and PATCH one report id per model. Each retrain becomes a revision, so a newly appearing bias is a diff rather than a discovery months later.

Can reviewers comment on one issue?

Yes. Threads anchor to the rendered content, so 'this slice is 40 rows, it is noise' or 'this one is a real regression' sits on the issue itself.

Share a Giskard scan report

Giskard scans a model the way a security tool scans a codebase: it goes looking for the failure modes you did not write a test for — a slice where performance collapses, a perturbation that flips the prediction, a spurious correlation doing the work you thought the features were doing.

The output is a report object in a notebook. The audience for it is a risk reviewer who has never opened a notebook.

Export the scan

import giskard

results = giskard.scan(giskard_model, giskard_dataset)
results.to_html("giskard.html")

One self-contained file: every detected issue, its severity, and the example rows that triggered it.

Publish it

Drag giskard.html into the app, or from the training pipeline:

curl -fsS -X PATCH "https://commareports.com/api/v1/reports/$REPORT_ID" \
  -H "Authorization: Bearer $COMMA_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d "$(jq -n --rawfile html giskard.html \
        --arg title "Model scan — $MODEL_VERSION" '{title: $title, html: $html}')"

Scripts run inside a sandboxed iframe (allow-scripts, no allow-same-origin), so the collapsible issue panels and example tables work at the URL.

Sign-off needs a durable artifact

"We checked for bias" is not a record. A published scan with comments on it is:

Anchored threads on the issue — "this slice is 40 rows, not a finding" next to the slice, or "this is the known seasonality effect, accepted in MRM-221" next to the drift. See commenting on HTML.
Revisions per retrain, so a bias that appears in v7 and not v6 is visible as a diff at the same URL. PATCH one report id per model.
Access per report — scans include example rows from your dataset. Keep them private or team-scoped. See the sharing model.
A routine if the model is retrained on a schedule and the scan should be too.

Limits

Entry HTML: 5 MB. Assets: 25 MB per file, 250 MB and 500 files total. Scans that embed many example rows grow quickly — narrow the dataset slice for the shareable view.
60 requests/minute per token.

Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision history.

Publish a model report →