Reference Workflows

A reference workflow is a complete, agent-driven application pipeline that TAO ships as a skill in the TAO Skill Bank. Where a model skill runs a single action such as train, evaluate, or inference, a reference workflow chains many of those actions together, adds the data services that sit between them, and carries its own state so that a long-running job can pause, report, and resume.

You do not invoke a reference workflow from a command line. You describe the outcome you want in plain English, and the agent loads the matching workflow skill, resolves your workspace, presents a pre-flight summary for your approval, and then drives the pipeline. To install the TAO Skill Bank plugin and reach a first successful run, refer to Getting Started with NVIDIA TAO Toolkit.

This section documents the Data Enhanced Fine-tuning (DEFT) workflow family, the iterative model-improvement loops that TAO 7.2.0 ships.

• Data Enhanced Fine-tuning (DEFT)
  • Loop Anatomy
  • Iteration Budgets
  • Data Enhancement Flavors
  • Data Enhancement Modules
  • Metric Contracts
  • State and Resume
  • Outputs by Workflow
  • Execution Platforms
  • Installing the Workflows
  • Choosing a Workflow
  • Source of Truth