Skip to content

Troubleshoot the DSPy workflow

Preserve failed task directories and diagnose configuration or data before explicitly starting a new task. This reference covers the Limited image.

Provider configuration is not saved

Run the helper as ubuntu in an interactive terminal that is not being recorded. The model name begins with openai/; the API base is a remote HTTPS endpoint without embedded credentials, URL query or fragment. Use the actual API base and model supported by your endpoint. Local loopback HTTP is not a customer mode.

If the file already exists, follow Provider settings to preserve it before changing it. The file must remain Ubuntu-owned and mode 0600. Symlinked or multiply linked configuration files are rejected.

Requests fail or time out

Check outbound HTTPS, the endpoint API base, model availability, key permissions and quota. A single request times out after 30 seconds; there are no configured SDK retries or adapter fallback requests. Keep the failed output folder. After fixing the cause, rerun explicitly into a new directory and account for new costs. Do not include the key or request bodies in a support message.

The dataset is rejected

Each split needs 3–24 examples and every one of the three supported labels. IDs and normalized ticket text must be unique across training, validation and heldout data. Check row fields and the 2,000-character text limit against the dataset rules. Changing the labels requires your own SDK program rather than changing only the input file.

Accuracy does not improve

The CLI selects by validation accuracy; it does not promise a heldout gain. Inspect both heldout scores, candidate selection and actual saved demonstration count. Use validation data for further iteration. Do not tune on heldout data; collect a new final evaluation set when appropriate.

Readback fails or a run directory exists

Choose a new output path for a new task. Preserve all original files from an existing task; missing or changed files fail its hash and metric checks. Reread only a completed run with its full dataset, program and prediction files together. A local readback makes no endpoint request. If setup folders are missing, inspect sudo journalctl -u sushi-dspy-firstboot.service --no-pager; do not make the managed runtime writable to repair workspace permissions.

Last updated November 16, 2026 · Suggest an improvement