Launch, monitor, and seal Marin canary and daily ferry training runs.
npx skills add https://github.com/marin-community/marin --skill run-ferries
Please help me install this Agent Skill. Skill: run-ferries Source: https://github.com/marin-community/marin/tree/main/.agents/skills/run-ferries Command: npx skills add https://github.com/marin-community/marin --skill run-ferries
Running recurring canary and daily ferry training jobs in Marin requires strict operational discipline: bounded config changes, explicit launch approval, hours-long monitoring, and consistent run sealing. This Skill encodes that entire operating procedure so ferry runs stay stable, auditable, and reproducible.
experiments/ferries/daily.py, record the proposal in the run issue, launch, and monitor 4-5 hours to completion.scripts/ferries/daily_analysis.py, seal the run with a git tag (ferry/daily/YYYYMMDD/<run_slug>), and open a log-only PR updating docs/experiments/daily-ferry-log.md.Ask the agent to launch and monitor the Marin canary ferry, or to prepare and run today's daily ferry with a bounded config change.
High-intent search queries and answers about installing and using this skill.
Edit experiments/ferries/daily.py with 1-2 bounded changes, record the proposal in the run issue, push the launch commit, then run the job via Iris with `uv run iris --cluster=marin job run -- python -m experiments.ferries.daily`. Explicit requester approval is required before launch.
The canary is a stable, low-cost MoE health check that runs unchanged and requires no PR or sealing tag. The daily ferry exercises a larger envelope with one or two intentional config changes and requires a proposal record, sealing tag, and log-only closure PR.
Yes, the canary supports GPU execution on CoreWeave by setting CANARY_ACCELERATOR=gpu and MARIN_PREFIX to the S3 path, launching through the cw-us-east-02a cluster with the Iris job runner.
Ferry monitoring typically takes 4-5 hours and must continue until the job reaches a terminal state: SUCCEEDED, FAILED, or STOPPED. Do not stop the monitoring loop early; escalate non-trivial failures to humans.
Debug with one bounded fix attempt, then escalate to humans if the failure persists. Canary failures are treated as urgent infrastructure or training-health triage, with a focused PR only if a script or config change is necessary.
Create and push a git tag in the format ferry/daily/YYYYMMDD/<run_slug> pointing to the exact launch commit, then open a PR that updates only docs/experiments/daily-ferry-log.md with labels ferry, ferry-daily, ferry-log-only, and ferry-sealed.
skywalkjianGenerate structured Markdown experiment reports from training scripts and logs.
box755Standardize LLM evaluation across 60+ academic benchmarks using lm-evaluation-harness.
NVIDIAMonitor submitted compute jobs and report their execution status changes.
nishide-devProfile PyTorch Lightning training to diagnose data loading, computation, and GPU memory bottlenecks.