Working surface: the source checkout’s finetune JSON CLI.
Bring: Python 3.10+ and repository access; River requires Python 3.12+.
Finish with: a retained run ID, dataset identity, and reviewed plan checksum.
Install from the reviewed checkout
From thestateset-agents repository root:
training extra. The checkout also provides
make install-training-locked for its Python 3.12/Linux training/API dependency lock.
Create a small format fixture
Createreviewed-chat.jsonl with the following two lines. They are synthetic examples of
format, not independently reviewed training data:
config.json, replacing the dataset path with the actual absolute path:
revision with the
model repository’s actual nonzero, full 40-character commit. Do not use an invented hash.
Validate configuration and inspect data
schema_version, ok, command, and result. Inspect
result.valid and its issues after validation; ok: true means a report was produced,
even when the configuration is invalid. inspect-data reads the file without exposing
examples by default; --include-samples is an explicit private-data read.
workflow reports settings, dataset, runtime, launch review, and evaluation readiness.
Missing local training packages or CUDA can block execution while data preparation remains
available. Readiness does not authenticate providers or certify model fit.
Prepare and retain the original run
result.run.run_id, result.run.plan_sha256, result.review, and result.launch_argv.
Preparation copies the validated input into a private snapshot and binds its data hash and
configuration. Later edits to the original file do not alter the reviewed input.
RUN_ID with the returned ID. Preparing again creates a new run, so preserve the
original response when recovering from a lost client connection.
Checkpoint: status identifies the same prepared run and plan, with no
worker submission or training completion implied.
Launch when the environment is ready
After installing the training dependencies, configuring CUDA/Buildr resources, reviewing the model pin, and authorizing the workload, launch the exact prepared plan:status
and inspect failure, metrics, completion metadata, and execution uncertainty. An identical
launch retry returns the original submission instead of starting another worker.
For remote launches, use the supported backend and its explicit billing policy, and add
--allow-paid only for an authorized workload. Modal and River have no enforced dollar
ceiling in this interface. See backend setup.
Continue in the terminal studio, automate with the
JSON CLI, or prepare
industry data with isolated splits.