Design or revise Marin experiment pipelines by connecting reusable tasks and their outputs in a dependency graph. Organize shared implementations separately from experiment settings, then check that the workflow connects the intended inputs and stages.
Design or revise a Marin experiment pipeline that binds reusable work into an artifact graph. Use for new workflow entrypoints, stage wiring, and pipeline placement decisions; use a narrower domain skill when it owns the launch.
Review the instructions the agent will receive before you install.
Keep the boundary between work and workflow explicit:
lib/** owns concrete implementations intended for reuse: functions, configurations, and produced data types. New library code must not compose a pipeline, construct an ArtifactStep or StepSpec, or reference a particular artifact handle, name, version, or output. The execution framework's Artifact and ArtifactStep definitions are infrastructure, not a model for new library workflow builders.experiments/** owns artifact graph construction and one-off binding code: choosing inputs, names, versions, dependencies, resources, and which reusable functions to run. Keep a transformation here only when it is specific to that experiment. Move work intended for reuse into a concrete lib/** implementation, then bind it in experiments/**.tests/** or lib/**. Keep tests of reusable library behavior in the library's normal test location. Follow root TESTING.md and the relevant module instructions.Inspect nearby current pipelines before editing. Read layout and examples for placement decisions and artifact binding when working with ArtifactStep, apply, paths, or caching. Existing library builders that return steps are exceptions, not precedent for adding more pipeline composition to lib/**.
For a change, verify that the experiment builds the intended dependencies and that reusable work can run without importing experiments/**. Dry-run the graph where the entrypoint supports it; run focused behavior tests without submitting a live job unless requested. Update user-facing docs when the workflow or API changes. Use the manage-pr skill if asked to publish the result.
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Compare Zephyr changes using the same prepared dataset and matching execution settings. Analyze Finelog metrics for each processing stage to distinguish CPU efficiency, elapsed time, and peak memory changes, then report improvements, regressions, or tradeoffs.
Register Hugging Face datasets for Marin by inspecting their structure without downloading the full dataset. Identify text fields, configurations, and data splits, then add a dataset module and validate the mapping with a sample.
Turn a specified GCP access request into a reviewable Pulumi change and pull request. Select the narrowest suitable role and encrypt personal email identities, covering resource permissions and Cloud Run viewer access through IAP, an access gateway.
Add or update Pallas/Mosaic kernels, specialized computation routines for JAX. The skill covers reference implementations, output and gradient checks, public wrappers, fallback behavior, performance benchmarks, and hardware-specific tuning when requested.
Try model or training changes in a copied Grug or Grugformer experiment, document the differences, and bring successful changes back into the base template while removing superseded experiment code and checking behavior.
Organize Marin experiment reports into a maintained index, or archive explicitly named older experiments using lasting tags and issue references. Existing conclusions and canonical links are preserved so experiment records remain traceable.