The Marin Project

Review a Marin access-change pull request by decrypting user identities and explaining their roles and resources. After explicit human confirmation, merge the request and apply only the verified permission changes through Pulumi.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Checking who receives access in an encrypted grant request
• Reviewing roles and resources before approving access
• Applying confirmed permission additions or revocations

Investigate Finelog logs and runtime measurements for Iris tasks, training, and vLLM. Discover available fields, write time-bounded SQL, summarize resource usage, compare central and regional records, and interpret counters correctly.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Summarizing Iris task memory and CPU usage
• Comparing central and regional logs for forwarding delays
• Aggregating vLLM counters with reset handling

Review GitHub experiment issues and update concise summaries that explain the setup, investigation, and current conclusion. Add useful supporting links and maintain labels so newcomers can understand both ongoing and completed work.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Add summaries to experiment issues
• Update stale conclusions or unresolved blockers
• Make closed experiment issues easier to understand

Operate Iris jobs and tasks, from submission and monitoring to debugging and recovery. The skill also guides scheduling checks, coordination between clusters, controller deployments, development GPU or TPU reservations, and investigation of stuck CoreWeave pods.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Monitoring Iris jobs and investigating failures
• Diagnosing scheduling delays and cross-cluster coordination
• Deploying or rolling back controllers

Search Echo for previous Marin work and cite the original evidence during debugging or research. Reuse relevant findings throughout an investigation, and create or update concise wiki notes that guide future decisions and troubleshooting.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Finding previous work before an investigation
• Searching past error discussions
• Recording reusable troubleshooting guidance

Run a repository's lint review, an automated check of pull request changes, and report every finding without editing code. Findings can be printed locally or posted as individual GitHub comments, with a fallback for lines outside the diff.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Run repository lint reviews in CI
• Post findings as comments on pull request lines
• Report findings that cannot be attached to changed lines

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.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Deploy backend or optimizer changes to a live training run
• Compare loss and throughput across matching training steps
• Rehearse checkpoint handoff and rollback procedures

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.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Check training health with an unchanged canary run
• Test one or two configuration changes in a daily run
• Monitor a ferry job through its terminal state

Check documentation and operator scripts against current code and command conventions. Correct confirmed mismatches in small, focused changes, publish them through a pull request, and report what was inspected when no update is needed.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Check README instructions against current behavior
• Update outdated command examples
• Review documentation and code consistency on a schedule

Build, validate, and release Marina applications and reports, then load or query their Postgres data. Manage application revisions, preview changes locally, roll back to retained packages, or archive applications through the Marina command line.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Build and publish a Marina report or application
• Preview application changes with local Postgres
• Load tables and inspect data with SQL

Collect experiment reports from issues and organize them in an existing Markdown index. It places entries in relevant sections, merges WandB experiment links into existing records, and checks for duplicate entries and formatting inconsistencies.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Adding new experiment reports to an index
• Sorting uncategorized experiments
• Merging WandB links into existing entries

Analyze oversized logs with natural-language queries using Gemini. Extract matching original lines with line numbers, produce a Markdown summary, or combine both steps to focus a report on relevant errors and events.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Find relevant lines in an oversized log
• Summarize log events around a specific question
• Extract errors before generating a focused report

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.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Compare Zephyr performance before and after a PR
• Assess speed gains against additional CPU usage
• Check peak memory and out-of-memory regressions

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.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Launch a specified Zephyr pipeline and check startup
• Monitor stage progress and worker health over time
• Investigate stalled stages or repeated worker failures

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.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Investigate and fix a specified GitHub issue
• Check for overlapping issues and pull requests
• Add a regression test for a reported bug

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.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Define a reinforcement learning experiment
• Check resource allocation and role configuration before launch
• Verify training with a small trial run

Profile JAX, Levanter, or Marin training to investigate slow startup, compilation, initialization, or throughput. Turn execution traces, recorded timing events, into structured reports, identify costly operations, and compare performance before and after a bounded change.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Investigate slow training startup or compilation
• Identify costly operations and communication delays
• Compare training performance before and after a change

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.

한국어 요약은 준비 중입니다. 영어 요약을 표시하고 있습니다.

• Investigate GPU or TPU canary failures triggered by CI
• Distinguish infrastructure failures from training or metric problems
• File a GitHub issue with logs and run context