Troubleshooting¶
Find the error from your terminal below. Numbers and values in angle brackets vary by run.
Every model call failed¶
every model call failed (N of N); first error: <detail>. Check the model name, the API key, the API base, and your network connection.
Every example in the evaluation returned a provider error, so Strop stopped without a score. Start with the first error shown. Check the model name, API key, API base URL, and network connection. After correcting the cause, retry the original command.
The dataset file changed since the split was created¶
the dataset file changed since the split was created; run 'strop data split' again
The current dataset's content hash differs from the one recorded in the active split. If you
edited the file on purpose, run strop data check and then strop data split to regenerate
the split. If the edit was accidental, restore the dataset used to create the active split.
Invalid rows in the dataset¶
N invalid row(s) in the dataset:
One or more rows failed validation. Run strop data check to see row numbers and errors, or
strop data check --json for the complete list. Fix the rows and run strop data split again.
To deliberately exclude the invalid rows, run strop data split --skip-invalid. Later
evaluations use the active split's setting.
The validation partition has too few rows¶
the validation partition has N rows, fewer than the M needed to compare prompts; add data
Optimization needs at least optimization.min_validation_samples validation rows to compare
prompts. Add valid labeled rows or adjust the split ratios to
allocate more rows to validation. Run strop data split again and check the validation count.
Keep the final test set held out from optimization.
Too few rows for this split¶
N rows are too few for this split: the <partition names> partition would be empty; add data or change the split ratios
A partition with a positive ratio would receive no rows. Add data or change the
split ratios, then run strop data split again.
The test partition was already used¶
the test partition was already used (evaluation <id>, vN). Testing again after changing the prompt biases the result; pass --allow-rerun to continue anyway
The evaluation in the message already used this split's test partition. Refer to the saved
result, or use fresh held-out data for a new independent evaluation. To deliberately repeat
the test on this split, run strop test --confirm --allow-rerun. A repeat uses the same data
and can give a biased estimate if you changed the prompt after seeing the earlier result.
A version is waiting for a decision¶
vN is waiting for a decision; run 'strop approve vN' or 'strop reject vN' first
A candidate is pending manual approval. Inspect the candidate and use the version shown in
the message with strop approve vN or strop reject vN. Then run strop optimize --resume.
See How it works for the approval workflow.
There is no unfinished run to resume¶
there is no unfinished run to resume; start one with 'strop optimize'
The current project has no resumable run. Check that you are in the intended project. If you
want a new run, start it with strop optimize.
I do not want colors or the banner¶
Color and the STROP banner appear only in a terminal. Set NO_COLOR=1 to turn color off, and
STROP_NO_BANNER=1 to hide the banner. Output that is piped to a file or another program, and
--json output, never contains either.