Skip to main content

MLflow Runs Data Model for GenAI

In MLflow 3, Runs represent evaluation iterations of your GenAI application and are attached directly to Models as part of the model's evolution. Unlike earlier MLflow versions where runs lived under experiments, runs now capture specific evaluation sessions that test and validate model performance.

Overview​

A Run represents a single evaluation iteration of your GenAI model - think of it as a snapshot of how your model or application performed during a specific testing session.

Runs as Model Evolution Snapshots​

Each Run captures a specific moment in your model's development lifecycle:

Relationship to Other Entities​

Runs connect your model development to systematic evaluation:

Key relationships:

  • Runs can link to Models: Iterative model or app development is tracked through runs
  • Runs generate Traces: Evaluation execution creates trace records
  • Runs produce Assessments: Quality judgments on model performance
  • Runs use Datasets: Systematic testing against curated examples
  • Runs apply Scorers: Automated evaluation functions

Benefits of Model-Attached Runs​

The MLflow 3 approach of attaching runs to models provides:

  • 🎯 Model-Centric Organization: Evaluation history travels with the model
  • 📈 Evolution Tracking: Clear progression of model performance over time
  • 🔍 Focused Analysis: Evaluation results directly tied to specific logged model
  • 🚀 Simplified Workflows: Natural connection between model development and testing

This model-centric approach makes it easier to understand how your GenAI application has evolved and which evaluation iterations led to improvements or regressions.