accessing-mlflow

To query MLflow runs by invocation_id, you can search across experiments using tags like invocation_id, model, and task_name. The invocation_id is interpreted as both a tag and a param on MLflow runs to help locate relevant runs.

Yes, you can retrieve fully resolved configs and nested evaluation metrics such as eval_factory_metrics.json from MLflow runs. This allows you to fetch artifacts for deep debugging to understand why evaluation results differ.

Yes, you need the mlflow-mcp server configured with the MLFLOW_TRACKING_URI environment variable. It enables natural-language queries to browse and operationalize analysis without manual UI navigation.

The best way to compare model performance is querying MLflow runs by tags like model and task_name across experiments. You can then extract nested metrics and configs for direct comparison of evaluation results.

MLflow run queries by invocation_id may fail if the mlflow-mcp server is not properly configured with MLFLOW_TRACKING_URI, or if the invocation_id is not correctly interpreted as both a tag and a param on the runs.

You can find specific MLflow runs using only an invocation ID by searching across experiments with the invocation_id tag. The system returns relevant run IDs and can extract associated artifact content like eval_factory_metrics.json.