Skip to content

Prepare data for z3 training ​

Start with one decision and labeled examples you are authorized to use. The website's table page and zils train prepare turn a spreadsheet or export into the files below; the hosted queue trains candidates on the selected immutable profile and evaluates them against a frozen comparison model.

Select an available pool and channel before creating the run. Omitting selection keeps the mini stable default. Every new run stores the resolved profile; recovery keeps the original profile even when the channel changes. See model pools.

Define the task ​

Record the context available at decision time, allowed outcomes, and correct label. Exclude information that only becomes available afterward. The training API accepts three JSONL splits using this case format:

JSON
{"id":"ticket-001","group_id":"conversation-001","family":"support-routing","state":{"text":"Please send my invoice."},"question":{"type":"choice","instructions":"Choose the responsible team.","criteria":{"billing":"Invoices and payments","technical":"Software troubleshooting"}},"label":"billing"}

Each line is one complete record. label identifies an outcome; group_id keeps related source records together; family identifies the task group for evaluation. The training format differs from the prediction API's model, state, and questions request.

z3-mini supports at most 255 outcomes and 8,192 input tokens per example. Inputs that exceed the limit are rejected, not truncated. See z3-mini.

Separate the data ​

FilePurposeMiner access
train.jsonlFit candidate adaptersZils workers (private) or network miners (open tier)
calibration.jsonlFit confidence calibrationValidator only
test.jsonlCompare the frozen modelsValidator only

Split related conversations, documents, or other source groups together before upload. Case IDs must be globally unique; groups and exact prompts cannot cross splits. Calibration and test must contain matching task families, all present in training. Review semantic duplicates and label quality yourself.

Reserve an independent final set if you repeatedly inspect test results. For upgrades, avoid overlap with data used to train or calibrate the previous model too.

Set acceptance criteria ​

Freeze minimum accuracy and minimum absolute Brier improvement before evaluation. First versions compare against the calibrated base. Explicit upgrades compare against the named previous accepted version. See version selection.

Training may complete with no_qualifying_model. The product-matching study recorded an additional adapter that reduced held-out accuracy and worsened probability error. Training completion does not justify replacement on its own.

Upload with an explicit trust boundary ​

Whoever trains the run can read the training split. Private runs are trained only by Zils workers; open-tier runs are trained by network miners, who must delete the data within 7 days of the run finishing (see open tier). The export-consent flag records your decision; it does not provide confidential compute or recall downloaded copies.

Follow Training workflow for sign-in, tiers and uploads. When capacity is available, the platform assigns a worker: a Zils worker for a private run, a qualified network miner for an open-tier run. Only accepted, verified models are activated. Private training costs $0.20 per million training tokens with a $2 minimum per run; open-tier training is free within its limits. See pricing.

For terminal uploads, use the training CLI. For image jobs, see Train on images.