Training data

Training data comes from traces — completed runs with both prompt and response.

Open Dashboard → Training data (/dashboard/training).

Export JSONL

Download a dataset for external fine-tuning tools:

Query parameters for GET /api/training/export:

• format=openai or format=alpaca
• actionName= — filter by action (e.g. agent_lab_run, coach_curated)
• curatedOnly=1 — only traces marked in Coach
• limit= — max rows (default 500)

Example:

curl -o training.jsonl \
  "http://localhost:3000/api/training/export?format=openai&curatedOnly=1" \
  -H "Cookie: <session>"

Build an Ollama model

From the Training page, Build model sends traced examples to Ollama’s create API (POST /api/training/create-model).

It generates a Modelfile:

• FROM your base model (e.g. llama3.2)
• SYSTEM prompt
• MESSAGE user / MESSAGE assistant pairs from traces

Use the new model name in Agent Lab or Chat after creation.

Dry run (API)

Preview the Modelfile without creating a model:

POST /api/training/create-model
{
  "name": "my-agent",
  "baseModel": "llama3.2",
  "dryRun": true
}

Tips

• Use Coach picks only for smaller, higher-quality datasets.
• Keep action names consistent (agent_lab_run, agent_chat, coach_curated) for filtering.
• Local runs cost $0; coach traces include estimated costUsd on the execution record.