mlflow2rdf

A Python package to convert MLflow tracking data into MLSO-aligned RDF Knowledge Graphs.

Overview

mlflow2rdf transforms your MLflow experiment metadata (parameters, metrics, tags) into semantic RDF triples that conform to the MLSO (Machine Learning Sailor Ontology) standard. This enables:

• Semantic querying via SPARQL instead of imperative MLflow API loops
• Cross-platform interoperability with public ML knowledge graphs like MLSea
• FAIR principles for ML experiments: Findable, Accessible, Interoperable, Reusable

Installation

pip install mlflow2rdf

Quick Start

Command Line Interface

# Convert MLflow runs to RDF
mlflow2rdf --uri /path/to/mlruns --out output.ttl

# With SHACL validation
mlflow2rdf --uri /path/to/mlruns --out output.ttl --validate

Flags are --uri / --out (not --mlruns / --output).

Python API

from mlflow2rdf import SemanticTracker, export_and_validate

# Convert to RDF
tracker = SemanticTracker(
    mlflow_uri="/path/to/mlruns",
    output_graph_uri="http://w3id.org/mlsea/mlflow/kg",
    default_modality="tabular",
)
graph = tracker.export(experiment_ids=None, exclude_failed_runs=False)

# Serialize to Turtle format
graph.serialize("output.ttl", format="turtle")

# Or export + validate in one call
export_and_validate("/path/to/mlruns", "output.ttl", inference="owlrl")

The previously documented MLflow2RDFConverter class does not exist. The real entry points are SemanticTracker and export_and_validate.

Features

• Multi-modal support: Tabular, image, text, time-series, and multi-modal data
• Paradigm-aware routing: Automatic parameter routing based on learning paradigm
• Pipeline relationship inference: Detects knowledge distillation, LoRA adapters, self-supervised learning chains
• SHACL validation: Comprehensive shape validation against MLSO constraints
• Blind spot analysis: Completeness verification of metadata extraction

Supported Learning Paradigms

Output Format

The package generates RDF triples in Turtle format, using MLSO/MLST vocabulary:

@prefix mls: <http://www.w3.org/ns/mls#> .
@prefix mlso: <http://w3id.org/mlso/> .

<run/abc123> a mls:Run ;
    mls:hasInput <dataset/cifar10> ;
    mls:hasOutput <evaluation/acc_0.95> ;
    mlso:hasParadigm "Supervised Classification" .

Requirements

• Python >= 3.8
• MLflow >= 2.0.0
• RDFLib >= 6.0.0
• pySHACL >= 0.25.0

License

MIT License

Citation

If you use this package in your research, please cite:

@mastersthesis{jia2026mlflow2rdf,
  author = {Jia, Sijie},
  title = {Bridging ML Tracking and Semantic Interoperability: Transforming MLflow Experiment Metadata to MLSO-Aligned RDF Knowledge Graphs},
  school = {KU Leuven},
  year = {2026}
}

Links

• MLSO Ontology
• MLSea Knowledge Graph
• Issue Tracker