L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Design et création · SKILL.md

generation-runtime

generation-runtime

Run DataDesigner through the Python API for validate, check_models, preview, create, workflow chaining, resume, export, and runtime troubleshooting.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: generation-runtime description: "Run DataDesigner through the Python API for validate, check_models, preview, create, workflow chaining, resume, export, and runtime troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Generation Runtime

Use this sub-skill when you need to execute DataDesigner from Python rather than author column configs or map CLI commands.

Start here

  • API Reference — public DataDesigner, DatasetCreationResults, PreviewResults, RunConfig, and workflow signatures.
  • Runtime Workflows — recommended validate → readiness → preview → create flow, async usage, and workflow chaining.
  • Artifacts, Results, and Resume — artifact tree, result semantics, resume modes, and export / push-to-hub behavior.
  • Troubleshooting — model and MCP readiness failures, early shutdown, resume incompatibility, export issues, and TTY logging differences.
  • Sampler Smoke Script — deterministic local smoke that avoids remote APIs.

Use this sub-skill for

  • Running DataDesigner.validate, preview, create, acreate, check_models, set_run_config, and compose_workflow.
  • Inspecting PreviewResults and DatasetCreationResults outputs, artifact paths, processor outputs, export files, and resume checkpoints.
  • Diagnosing model alias, MCP tool alias, or async/runtime issues during local generation.
  • Understanding when validate still works even when no usable model aliases exist.

Route away

Safety defaults

  • Prefer sampler-only smoke runs first when model aliases are missing or untrusted.
  • Use validate before check_models; validate only checks the configuration shape and can succeed without usable model aliases.
  • Set RunConfig(display_tui=False, otel_metrics_port=None) when you want quiet local generation and no terminal UI.
  • Keep buffer_size, preserve_dropped_columns, and the workflow stage layout unchanged when resuming a run.
  • Do not call push_to_hub from the smoke helper; it requires network access and a valid Hugging Face token.
Skills enregistrées