Update external forks used by Marin to newer upstream code, review which custom patches to retain, update pinned revisions, and run required tests before preparing a draft pull request for review.
Review recent issues, pull requests, and recurring operational problems to choose a concrete repository improvement. Check for duplicate work, implement a suitable change, or document a plan and follow-up when implementation is blocked.
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.
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.
Launch and continuously monitor a specified Zephyr pipeline on Iris when asked. It checks coordinator and worker health, tracks stage progress, identifies stalled or repeated failures, and requires explicit approval before stopping or restarting jobs.
Deploy code changes to a live hero training run by starting a new run from a permanent checkpoint, a saved training state. Compare matching training steps against agreed criteria and roll back if the trial fails.
Define and launch Marin SkyRL reinforcement learning experiments on Iris. It guides configuration checks, model and data dependencies, small trial runs, and progress verification while keeping artifact versions and runtime revisions explicit.
Publish concise infrastructure incident records in Echo, with symptoms, response steps, causes, and evidence links. Search for existing records first, update ongoing incidents, and connect the published record to its related pull request or issue.
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.
Preserve old experiment scripts with a dated Git tag, a named snapshot of the code, and add comments linking related GitHub issues to the archived files and cleanup pull request.
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experiment, document the differences, and bring successful changes back into the base template while removing superseded experiment code and checking behavior.
Prepare completed changes for commit and pull request review. It cleans the diff, runs repository checks, commits and pushes changes, prepares the pull request, and monitors automated checks and review feedback after publication.
Resolve an identified GitHub issue through investigation, duplicate-work checks, implementation, and regression testing. Open a pull request, address CI and review feedback, and track the outcome while keeping the issue updated with concise progress comments.
Run Marin canary checks for training health or daily experiments with small configuration changes. Monitor each job to completion, record evaluation results, and close daily runs with a version tag and log update.
Diagnose failed canary ferry runs, automated checks for failures or regressions, when invoked by CI with the required context. It gathers logs, classifies the cause, and files a GitHub issue with supporting evidence.
Upload model weights to Hugging Face Hub with an OpenMDW 1.1 license and a required research warning. Check the published README and license file after uploading or updating to confirm that the required information is present.
Investigate code failures, startup problems, and performance regressions involving JAX, Iris, Zephyr, or TPU hardware. It guides failure reproduction, focused hypothesis testing, and infrastructure diagnosis using the relevant operational procedures.
Manage files in GCS, CoreWeave S3, R2, and local storage through Marin's fsutil command-line tool. List, inspect, copy, move, or synchronize data using bucket connections defined in the checkout's configuration.
Is this your account? Link GitHub to bring it onto your creator page (Japanese)