marin-communityOfficial·51 Agent Skills Included

marin

Train, debug, and operate large language models end to end

Runs the full foundation model workflow: dataset registration, tokenization, distributed training on TPUs and GPUs, evaluation, and cluster operations. Removes manual job babysitting, lost experiment records, and fragile fork maintenance through scripted agent playbooks. Guides agents through commits, PR reviews, incident logs, and profiling so research runs stay reproducible and auditable.

3,245|266

↗

npx skills add marin-community/marin --all -g -y

Available:

Tells the AI agent how to route tasks to the right sub-project manuals and skills, enforce Marin's coding and commit rules, search prior work in Echo, and safely operate training clusters without unauthorized restarts.

All Skills in This Repository (51)

Pure Emerald Level Indicators

📦 In Repo

marin-community

zephyr-debugger

Diagnose Zephyr pipeline issues by analyzing logs, profiling, and resource data.

iris-controller-debug

Analyze offline checkpoint snapshots and profile Iris controller processes via RPC.

design-doc

Generate concise design documents for proposed technical changes.

ferries

Manage and monitor automated integration test runs for the Marin framework.

agent-research

Manages exploratory research workflows with logs, branches, and benchmarks.

scrub-experiment-issue-tldrs

Update TL;DR summaries and links on experiment GitHub issues.

github-pr-review

Review GitHub pull requests for correctness and behavioral regressions.

handle-gh-scrub

Automate scheduled GitHub issue and pull request triage scrub runs.

file-issue

Create GitHub issues in the marin-community/marin repository using predefined templates.

change-grug

Manage Grug and Grugformer experiment modifications with a template-first workflow.

pull-request

Standardize GitHub pull request authoring and review within the Marin framework.

add-dataset

Inspect Hugging Face dataset schemas and register them as Marin ExecutorSteps.

Official

Intermediate

Frequently Asked Questions

FAQPage Schema

How to install Marin?▼

Run `npx skills add marin-community/marin --all -g -y` in your terminal to install all agent skills in this suite globally.

What does Marin do?▼

Marin is an open platform for training large language models, covering data curation, tokenization, pretraining, post-training, and evaluation. Its agent skills automate the surrounding workflows like job monitoring, PR reviews, and experiment tracking.

Can AI agents manage training runs with Marin?▼

Yes. Skills like manage-hero-run, use-iris, and babysit-zephyr let an agent launch, monitor, debug, and seal large training jobs on TPU and GPU clusters with strict safety guardrails.

Does Marin work with Claude Code and other agents?▼

Yes. All skills follow the standard SKILL.md format under .agents/skills and .claude/skills, so they load in Claude Code, Codex, and compatible agent runtimes.

Do I need ML expertise to use these skills?▼

Basic familiarity helps, but the skills encode Marin's operational playbooks step by step, so an agent can follow them from plain-English requests without you knowing every command.