Tutorials

These tutorials are task-oriented walkthroughs. Each one names the model shape, the inference route, and the point where the result should be inspected or validated.

How to Use These Tutorials

Each tutorial is written as a workflow, not as a benchmark claim. The code blocks show the smallest useful path through the API; the surrounding text names the checks that make the result credible in a real project.

For exploratory work, start with the model shape and verify that fitting, scoring, and sampling run without warnings. For production-facing work, keep a separate validation split, record the estimator or program configuration, and save the diagnostic output that would explain a later promotion decision.

The tutorials intentionally point back to the reference guides. Use the walkthroughs to assemble a workflow, then use the guides to check edge cases, capability contracts, optional dependencies, and release-readiness expectations.

Choose by Problem

If you need to	Walkthrough	Surface
model records with mixed field types	Fitting Heterogeneous Records	Stable core
express a model with the PPL layer	PPL Mixture Workflow	Active development
enumerate top-k support values	Enumeration and Ranking	Stable/evolving core
save, serve, and monitor model artifacts	Production Artifacts	Practical production helpers
replace repeated LLM calls with a calibrated local model	LLM Distillation Cascade	Active task workflow
decide when an LLM should abstain	LLM Uncertainty	Active reasoning workflow
build shared representations for multiple modalities	Representation and Model Families	Active representation workflow
combine distribution operations with structured decisions	Relations and Operations	Stable/evolving core
run an auditable model improvement loop	Evolution and Analysis	Active design/evolution workflow

What Each Tutorial Demonstrates

Tutorial	Main idea	Related guides
Fitting Heterogeneous Records	A tuple-shaped row becomes a composite estimator, and a mixture adds a latent cluster over the whole record.	Core Concepts, Structured Statistical Families
PPL Mixture Workflow	free parameters and Mix lower to the same estimator/distribution contract as the core API.	Probabilistic Programming, Automatic Inference
Enumeration and Ranking	A fitted model can expose ranked support traversal when the capability is available.	Enumeration and Ranking, Capabilities and Contracts
Production Artifacts	Fitted models need provenance, registry metadata, serving wrappers, and drift checks.	Production Workflows, Model Lifecycle
LLM Distillation Cascade	A teacher labels examples, a local model learns the task, and calibrated confidence decides whether to answer or escalate.	Task Distillation, Task Serving, Routing, and Edge Deployment
LLM Uncertainty	Repeated LLM samples become semantic entropy, answer confidence, and abstention decisions.	Uncertainty, Reasoning Systems
Representation and Model Families	Segmenters, embeddings, and vector quantizers turn heterogeneous modalities into a shared modeling stream.	Representation Layer, Model Families
Relations and Operations	Distribution operations and structured feasible-set solvers solve different parts of a decision workflow.	Operations, Relations
Evolution and Analysis	Diagnostics and objective-led search promote challengers only when they pass a verification gate.	Analysis Utilities, Evolution and Search

• Fitting Heterogeneous Records
• PPL Mixture Workflow
• Enumeration and Ranking
• Production Artifacts
• LLM Distillation Cascade
• LLM Uncertainty
• Representation and Model Families
• Relations and Operations
• Evolution and Analysis