Skip to content

Training workflow ​

Train a z3-mini model on your own labeled examples and review the result. You need a Zils account (sign up with Google or an email link) and labeled data you are authorized to use. You do not need to set up a database or run a GPU.

You can start a run in four ways:

  • The website. On the table page, drop a CSV, Excel, JSON or JSONL file of past decisions, put the column it should decide on the right of the line, and start the run.
  • The CLI. zils train prepare turns an export into the three files and zils train submit --tier open|private submits them (Train from the CLI).
  • A coding agent, through the MCP server's training tools (MCP).
  • zils fit train, from decisions your program logged (find where Zils fits).

Choose a tier ​

Every run is open tier or private:

Open tierPrivate
PriceFree, within its limits$0.20 per million training tokens, $2 minimum per run
Who trains itNetwork minersZils workers only
Training dataThe training split is sent to minersNever sent to miners
The trained modelServed by Zils, used through the API; cannot be downloadedServed by Zils, and its files can be downloaded
Limits3 runs per account in 24 hours; 5,000 examples and 5,000,000 training tokens per run; a daily limit across ZilsAt most five active jobs per account

Paid runs go ahead of open runs in the queue. See pricing and limits.

1. Prepare your data ​

Start with one decision you want the model to make. Follow Prepare a dataset to define the question, allowed answers, labels, and three separate JSONL files:

FilePurposeAccess during training
train.jsonlTrain candidate adaptersZils workers (private) or network miners (open tier)
calibration.jsonlCalibrate confidenceTrusted validator; not sent to miners
test.jsonlEvaluate candidatesTrusted validator; not sent to miners

Keep related source records together in one split. Choose your minimum accuracy and required improvement in probability error before submitting the job. z3-mini training supports at most 255 outcomes and 8,192 input tokens per example. Oversized inputs are rejected, not silently shortened.

2. Upload and submit ​

Sign in, prepare your job, choose its tier, and review the permission to export training data. An open-tier run also needs your permission to send the training split to network miners. Submit only data you are authorized to share. Uploads use private storage and temporary upload URLs associated with your job.

Upload all three splits before submitting for validation. If an upload is interrupted, resume the missing files. An uploaded file cannot be replaced in that job; cancel and create a new job to change the inputs. A failed file transfer restarts that file rather than resuming from the interrupted byte.

Your signed-in session authorizes training actions. The API key used for predictions is a separate credential.

3. Follow training progress ​

Validation checks the dataset and pins the job's model and acceptance criteria. The job then waits for a worker and available capacity. A private run goes only to Zils workers. An open-tier run goes to one network miner at a time, chosen from miners that have qualified their hardware.

The assigned worker trains a candidate adapter. The trusted validator checks the returned artifact, calibrates confidence, and evaluates it on the held-out test split. Miners do not receive calibration data, test data, or acceptance criteria.

The dashboard tracks upload, validation, assignment, training, and evaluation. Waiting for capacity is different from a failed job. There can be at most five active jobs per account.

4. Review the result ​

ResultWhat it means
acceptedA candidate met all frozen acceptance conditions
no_qualifying_modelEvaluation finished, but no candidate met every condition
Failed jobTraining or evaluation could not complete; review the reported error

Completed does not guarantee an improved model. A first version is compared with the calibrated starting model. An explicit upgrade is compared with the named previous accepted version on the new test data. A rejected candidate does not replace your current API model.

Review the aggregate results and comparison before using a model on real traffic. An accepted private run can include private, short-lived artifact download links; an open-tier model is used through the API only. Measured acceptance on this test set is not a guarantee on future data. See Understand the result and Model versions and upgrades.

5. Use an API-ready model ​

Training completion and API readiness are separate. After a candidate is accepted, the activation workflow verifies the serving runtime and registers the private model for your account. Only the workflow state ready confirms an API model ID. activating, activation_failed, or needs_review does not mean the new model is serving.

Use the confirmed model ID in the model field of your prediction request with your existing API key. An immutable ID pins a release; the task alias follows accepted upgrades. Follow the API quickstart to make a request. Other accounts cannot list or call your private model.

Data access and cancellation ​

On the open tier, the miner that trains your run can read its training split. The miner terms require miners to delete it within 7 days of the run finishing; Zils cannot verify that they do. Private storage and expiring URLs control access, but cannot recall downloaded copies. Export consent does not provide confidential compute. Calibration and test data stay with the trusted validator. Use a private run for data that must not leave Zils.

Canceling stops new claims and prevents a canceled job from completing through the queue. It does not immediately stop remote compute or delete data already downloaded by a miner. Uploaded files stay until scheduled cleanup removes them; copies a miner has already downloaded are outside that cleanup. The decision API's bulk-data retention controls are separate.

For terminal training, use the CLI guide. For photos, see Train on images. Each job freezes its own model profile.