Train a model
Runs LoRA fine-tuning on a method's graded pairs, returning an adapter and a held-out agreement report.
admin.jewels.train({ methodId })
Kicks off a LoRA fine-tuning run on the method's graded pairs. It returns immediately with a run id and the dataset report; the GPU trainer runs asynchronously for minutes to tens of minutes, so never pass --wait from an agent loop (it is for a human at a terminal). The tuning recipe comes from the method's manifest, and the run produces a downloadable adapter plus a held-out agreement report scored by the jewel's own grader. Poll with run.
Parameters
<methodId>
string
RequiredThe method to train a fine-tuning run for.
--wait
flag
Block until the run finishes. For a human at a terminal only; never from an agent loop.
CLI
remy-admin jewels train triage-issue