Docs

Training data

What a training pair is, how Studio writes them, and how to review, import and export a project's pairs.

A model learns your tools from training pairs. Each pair is one utterance and the tool call it should produce:

InputOutput
show unpaid invoices from this quarter[{"name": "list_invoices", "arguments": {"status": "unpaid", "issued": "this_quarter"}}]

The output is a list because one utterance can need more than one call, though most need one.

Where pairs come from

SourceHow
Written by StudioThe wizard's Generate step writes pairs for each tool with Gemini. See Create a model.
A templateThe WooCommerce template comes with its pairs.
ImportedA JSONL file you wrote or exported earlier. See Import and export.
Your trafficOn the Retrain tab, real requests become pairs once you check them. See Retrain.

Examples of what the model can't do

Alongside the pairs for your tools, Studio writes examples of requests your tools can't answer, so the model learns to say so instead of choosing the closest tool. There are two kinds:

KindExample, for an invoicing appWhat the endpoint returns
Out of scope"what's the weather in Paris"No calls, out_of_scope: true and a short reply you can show the user
Related but unsupported"explain how VAT works"No calls from the model, so the AI fallback can try, if it's on

Call your model describes each response.

Reviewing a project's data

The Data tab has two sections:

  • Tools lists each tool with its description and arguments. Open a tool to see each argument's type, whether it is required (marked *) and its allowed values.
  • Training pairs lists every pair as Input and Output, with the style it was written in.

Things to look for:

  • Wrong calls. An input whose output names the wrong tool or a wrong value teaches the model that mistake.
  • Missing phrasings. If your users say "overdue" and every pair says "unpaid", add pairs that say "overdue".
  • Thin tools. A tool with only a few pairs is the one the model is most likely to miss.

The Data tab is read-only. To change pairs, export them, edit the file and import it again, or correct real requests on the Retrain tab.

Import and export

Export JSONL on the Data tab, or on the wizard's Generate step, downloads the project's pairs. JSONL is a text file with one JSON object per line:

{"input_text": "show unpaid invoices from this quarter", "output_json": [{"name": "list_invoices", "arguments": {"status": "unpaid", "issued": "this_quarter"}}], "approved": 1}
{"input_text": "what's the weather in Paris", "output_json": [{"name": "__out_of_scope__", "arguments": {}}], "approved": 1}
FieldMeaning
input_textThe utterance.
output_jsonThe calls it should produce. Use __out_of_scope__ for an unrelated request and __fallback__ for a related one your tools can't answer.
approved1 to train on the pair, 0 to keep it out of training.

To import, choose Import JSONL on the wizard's Generate step. If the project already has pairs, the button reads Replace with JSONL.

An import replaces every pair in the project. To add to the existing pairs, export them first, add your lines to that file, and import the result.

A file can be up to 10 MB and hold up to 20,000 pairs. Lines that can't be read are skipped, and Studio lists the first few problems so you can fix them.

Revised

Was this page helpful?