---

name: modeling-data description: "Build, mutate, inspect, and transform PyMC models with Model contexts, data containers, dimensions, logp/debug tools, and do/observe workflows." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

PyMC modeling-data

Use this sub-skill when a task needs a PyMC model graph to be built, mutated, inspected, or conditionally transformed before inference. It covers pm.Model, named variables, observed data, coords/dims, pm.Data, pm.set_data, pm.Deterministic, pm.Potential, initial points, practical logp/debug helpers, graph visualization, pm.do, pm.observe, and the experimental pymc.dims surface.

Read or run these bundled resources

• Read references/modeling-and-data-workflows.md for context-managed model construction, data updates, dimensions, prediction-data resizing, logp/debug inspection, graph visualization, pm.do, pm.observe, or pymc.dims routing.
• Read references/api-reference.md for verified signatures, parameter meanings, and model-method call patterns.
• Read references/troubleshooting.md for shape/coordinate errors, invalid observed data, duplicate names, missing-value imputation, transform/value-variable confusion, or Graphviz issues.
• Run scripts/model_data_smoke.py for a tiny CPU-only check covering model creation, logp, pm.Data, pm.set_data, pm.do, pm.observe, and optional posterior-predictive shape changes.

Route here

• Creating a model with with pm.Model(...) as model: or object-style model=... registration.
• Inspecting model.named_vars, model.free_RVs, model.observed_RVs, model.data_vars, model.deterministics, model.coords, and model.named_vars_to_dims.
• Adding observed data and understanding missing-data/imputation caveats.
• Using coords, dims, pm.Data, model.set_data, and pm.set_data for reusable models and prediction-data resizing.
• Recording derived quantities or log-probability factors with pm.Deterministic and pm.Potential.
• Evaluating initial points, compiled logp functions, simple compiled expressions, model.debug, and model.to_graphviz.
• Transforming models with pm.do interventions or pm.observe conditioning.

Route elsewhere

• Distribution catalogs, .dist() component distributions, custom distributions, custom logp/logcdf, support-point, and distribution transform internals: use ../distributions-logprob/SKILL.md.
• MCMC configuration, prior/posterior predictive execution, DataTree group interpretation, convergence diagnostics, and output backends: use ../inference-predictive/SKILL.md.
• Gaussian processes, ODE models, variational inference, and minibatches as an inference strategy: use ../advanced-workflows/SKILL.md.

Operating guardrails

1. Prefer a context manager and keep variable creation inside it unless you explicitly pass model=model.
2. Treat regular coords/dims as shape and output-label metadata, not xarray-style label alignment.
3. Use pm.Data for inputs you plan to replace; pm.set_data can change values and shape but not rank, and coordinate lengths must stay synchronized.
4. Use model.initial_point() as the safest starting point for compiled logp checks because it uses value-variable names, including transformed names such as sigma_log__.
5. pm.do and pm.observe return new models. Do not assume the original model was mutated.