Read the results

The run page shows the progress and the results of a run. This guide explains the tabs on the run page, the scores, and how to export or serve the best version.

Tabs on the run page

The run page shows only the tabs that apply to the run. Some tabs depend on the path, the status and your experience level.

Scores

Skynet selects versions on the validation cases. It reports the final score on the test cases. The engines never see the test cases. Read Datasets and splits.

The Overview tab shows:

• Improvement: the change from the starting version to the best version.
• Score progression: the score over the run.
• Score breakdown: each metric, with the Baseline, Optimized and Change columns.

While the run is active, the overview also shows Elapsed, Remaining and Rate.

For a Prompt, Code or Text run, the overview also shows Named scores when your metric reports named scores.

For a DSPy program run, the Data tab has a Test set evaluation. It shows the baseline and optimized score for each test case.

When the best version is the starting version

Sometimes no candidate beats the starting version. Skynet then returns the starting version. The Artifact tab tells you that a regression guard was applied.

If the run stopped at its budget, the run page tells you which version Skynet saved. Read Runs.

Export the best version

Prompt, Code and Text runs

The Artifact tab lists the Versions of your artifact. For each version you can:

• Copy the version.
• Download the version as a file.
• For a repository run, download the version as a zip of the whole repository.

You can compare a version with another version. The tab also shows the Scorer feedback for a version.

DSPy program runs

The Artifact tab has an Export results panel. Open the Download menu. It shows the formats that the run has:

Serve the best version

You can call an optimized DSPy program from your own code. The Usage tab has two parts:

• A form. Enter the input fields and select Run to get a prediction.
• Integration code in cURL, Python, JavaScript, Go and DSPy.

The program always runs on the model settings that it was optimized with. You must give all input fields. Skynet ignores extra fields.

To call the program, create an API token in Settings → API. Then send a request:

export APP_API_TOKEN=skyd_...
curl -X POST "$API_URL/serve/<optimization_id>" \
  -H "Authorization: Bearer $APP_API_TOKEN" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{"inputs": {"question": "What is DSPy?"}}'

import os
import uuid
import requests

response = requests.post(
    f"{os.environ['API_URL']}/serve/<optimization_id>",
    headers={
        "Authorization": f"Bearer {os.environ['APP_API_TOKEN']}",
        "Idempotency-Key": str(uuid.uuid4()),
    },
    json={"inputs": {"question": "What is DSPy?"}},
)
print(response.json()["outputs"])

Replace question with the input fields of your program. The Integration code on the Usage tab fills in the real URL and field names for you.

The response contains outputs, the input_fields, the output_fields and the model_used.

To get partial outputs as they arrive, send the same request to /serve/<optimization_id>/stream. The response is a stream of server-sent events.

Read API for authentication and the other endpoints.

Share the results

You can share a run with other people, or list it in the public Explore gallery. Read Share and publish.