MLflow Integration

ModelDoctor seamlessly integrates with MLflow to track diagnostic health scores, metadata, and artifacts alongside your standard training metrics.

Installation

Ensure mlflow is installed in your environment. It is not a direct dependency of the ModelDoctor core library:

pip install mlflow

Usage

Use the log_report function to attach a ModelDoctor report to your active MLflow run.

import mlflow
import modeldoctor as md
from modeldoctor.integrations.mlflow import log_report

# Train your model and run diagnostics
report = md.diagnose(model, X_train, y_train, X_test, y_test)

with mlflow.start_run():
    # Log standard training metrics
    mlflow.log_metric("accuracy", 0.95)

    # Log the ModelDoctor report alongside your standard metrics
    log_report(report, artifact_path="modeldoctor")

What Gets Logged?

Metrics

The following scalar metrics are tracked in your MLflow run:

Metric	Description
`md_health_score`	The overall 0–100 health score.
`md_critical_findings`	Count of `CRITICAL` severity findings.
`md_warning_findings`	Count of `WARNING` severity findings.

Artifacts

The integration automatically renders and uploads several files to the specified MLflow artifact_path (defaulting to modeldoctor/):

File	Description
`dashboard.html`	The interactive HTML dashboard.
`report.json`	The raw JSON payload for downstream programmatic consumption.
`summary.md`	A human-readable Markdown summary of the findings.

Viewing Results in the MLflow UI

After logging, start the MLflow tracking UI to view results:

mlflow ui

Navigate to your experiment and click the run to see the md_health_score, md_critical_findings, and md_warning_findings metrics alongside your standard training metrics.

Limitations

The integration assumes an MLflow run is already active (mlflow.start_run()). If one is not active, it will create a new run, which may fragment your tracking history.
Ensure the [dashboard] extra is installed for HTML dashboard generation within the MLflow artifact upload.