Can I share a Weave evaluation publicly?

Only by making the whole W&B project public, which is usually not an option for work data. Exporting the evaluation table and publishing that gives you a per-report sharing decision instead of a per-project one.

How do I export a Weave evaluation?

`weave.init()` then query the calls for your Evaluation op through the client, build a DataFrame of input, output and scores, and `.to_html()` it.

Does the comparison view export?

The underlying rows do, which is what a reviewer needs. Build the columns you would have compared — model A score, model B score, delta — and sort by the delta.

Why not just add the reviewer to the W&B team?

Because a seat costs money and an invite costs a week. Most reviewers need to read one table once; a link is proportionate.

Can I publish every run?

PATCH the same report id per eval suite and each run appends a revision, so the URL is the current run and the archive at the same time.

Share a W&B Weave evaluation

Weave's evaluation view is genuinely good, and it lives inside a Weights & Biases project. Every person who needs to read it needs a seat in that project — which is fine for the ML team and absurd for the one domain expert who can tell you whether answer 14 is actually wrong.

Export the evaluation rows

import weave, pandas as pd

client = weave.init("support-agent")

rows = []
for call in client.get_calls(filter={"op_names": ["Evaluation.predict_and_score"]}):
    rows.append({
        "input": call.inputs.get("example", {}).get("question"),
        "output": (call.output or {}).get("model_output"),
        **{k: v for k, v in ((call.output or {}).get("scores") or {}).items()},
    })

pd.DataFrame(rows).to_html("weave-eval.html", index=False, float_format="%.3f", escape=True)

Publish the file — drag it into the app, or from the job that ran the eval:

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 weave-eval.html \
        --arg title "Support agent eval — $(date +%F)" '{title: $title, html: $html}')"

What you get that a project invite does not give you

Anchored threads — the objection lands on the row it is about. See commenting on HTML.
Revisions — one report id per suite, one revision per run, no dashboard to build.
Access per report — private, team-only, or domain-gated, decided per report rather than per project. See the sharing model.

Close the loop with the agent

Reviewer corrections are the highest-value training data you will get this quarter. An agent can read them back through the MCP server and fold them into the eval set — let an agent respond to comments.

Limits

Entry HTML: 5 MB. Model outputs dominate the size; truncate them or export per scorer.
Assets: 25 MB per file, 250 MB and 500 files per report.
60 requests/minute per token.

Try it

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

Publish an eval report →