How do I use natural language to build a machine learning model?

Summary

• Genie Code takes natural language requests and automatically generates complete, multi-notebook ML workflows including featurization, model training, hyperparameter tuning, evaluation, and deployment.
• You describe your ML task in plain English (e.g., "train a classification model to predict customer churn"), and Genie Code writes the engineering code while you stay in control and decide what to keep.
• Genie Code integrates deeply with Databricks MLflow for experiment tracking, Model Serving for deployment readiness checks, AI Runtime for GPU training, and Genie Ontology to learn your team's patterns.
• Multi-model routing ensures each part of your ML pipeline uses the best specialized model for that task, achieving higher quality than a single general-purpose coding agent.
• Genie Code is available directly in Databricks notebooks, SQL editor, and other Databricks surfaces, making it accessible wherever you already work.

How do I use natural language to build a machine learning model?
Databricks Genie Code lets you build machine learning models by describing your task in natural language instead of writing engineering code. It generates complete, production-ready multi-notebook workflows that handle the complex ML engineering work, freeing data scientists to focus on model strategy and interpretation.

What is natural language ML development?

Traditionally, building an ML model requires writing code for data loading, feature engineering, model selection, hyperparameter tuning, evaluation, and deployment. Each step involves boilerplate code, manual validation, and coordination across notebooks. Natural language ML development flips this: you describe your goal in plain English, and the system generates the full workflow. You review the generated code, make adjustments, and trust that best practices are applied automatically.

Key capabilities to evaluate

When choosing a natural language ML tool, consider these capabilities:

Capability	Why it matters
Full lifecycle automation	Featurization, training, tuning, evaluation, and deployment handled in coordinated steps, not just code snippet generation.
Context awareness of your data	The tool should discover your Unity Catalog data and understand table schemas without manual exploration.
ML best practices	Automatic detection of class imbalance, feature quality checks, and appropriate model selection for your problem type.
Integration with ML infrastructure	Reads MLflow run history, checks Model Serving endpoint health, and routes to AI Runtime for GPU training when needed.
Organizational learning	Learns your team's patterns for feature engineering, model selection, and evaluation to generate code consistent with your practices.
Multi-notebook coordination	Edits across multiple files are coordinated, so generated workflows stay coherent and runnable.

How Databricks approaches natural language ML

Genie Code generates full ML workflows from natural language. You provide a natural language request like "Create a classification model to predict customer churn using these datasets" along with table references. Genie Code then generates complete multi-notebook workflows including data loading, featurization, model training, hyperparameter tuning, evaluation, and preparation for deployment. The code is human-readable and you can edit any part.
Deep integration with your ML stack. Genie Code reads your experimentation data from MLflow (runs, artifacts, model lineage, quality metrics), inspects Model Serving endpoint health to understand deployment constraints, and automatically moves to AI Runtime when GPU training is needed for deep learning tasks. This integration means Genie Code understands your team's ML infrastructure and works within it.
Learning from your team's patterns. Through Genie Ontology, Genie Code learns how your team builds features, selects models, and evaluates candidates. This organizational context means generated code follows your established patterns and conventions, making it easier for your team to adopt and modify.
Intelligent routing, not a single model. Genie Code is not powered by a single large model. Instead, it routes different tasks (feature engineering, model selection, evaluation) to specialized models and tools. This multi-model approach ensures each part of your workflow uses the best-suited agent for that specific job.
You stay in control. Generated notebooks are presented for your review. You can inspect the generated code, make changes, and decide what to keep. Genie Code stays transparent about what it's doing.

Use cases

Rapid model prototyping. A data scientist exploring a new business problem can describe the task to Genie Code and get a working baseline model in minutes instead of spending hours writing featurization and training code. This accelerates experimentation cycles.
Scaling model development across teams. When multiple teams need to build similar models (e.g., churn prediction for different business lines), Genie Code can generate models for each team, adapting to each team's data schema while maintaining consistent ML practices.
Streamlining model refreshes. When you need to retrain a model on new data, describe the retraining task to Genie Code, and it generates updated workflows. No need to manually refresh old notebook code.
Reducing model engineering burden. In organizations where a few ML engineers support many analysts, Genie Code lets analysts generate their own models without waiting for engineering support, freeing engineers for higher-value tasks.

FAQs

How do I know what to ask Genie Code?

Describe your ML task like you would to a colleague: what kind of data, what you're predicting, and any constraints (e.g., "use only features from these tables"). More specific requests (including data locations, target variable, problem type) produce better results than vague ones.

Can Genie Code handle custom model types or specialized frameworks?

Genie Code works with standard ML frameworks like scikit-learn, XGBoost, and TensorFlow. For highly specialized models or novel architectures, you may need to edit the generated code or write custom training logic alongside Genie Code output.

What if I don't like the generated code?

You can edit the notebooks directly. Genie Code stays available if you want to generate new sections or refine parts of the workflow. You maintain full control over what runs.

Does Genie Code handle time-series forecasting?

Yes. Describe a time-series forecasting task, and Genie Code generates appropriate feature engineering and model selection for forecasting problems, including lag features and seasonal decomposition when relevant.

How does Genie Code handle imbalanced datasets?

Genie Code automatically detects class imbalance in your target variable and applies appropriate techniques like stratified splitting, resampling, or class weighting. It includes these decisions in the generated code so you can see and modify them.