Run Experiment

What problem does it solve?

Manually submitting multiple SLURM jobs and constantly monitoring their status is tedious and inefficient. This Skill automates the submission of all fine-tuning jobs and provides continuous, automatic monitoring, freeing you from constant oversight and letting AI handle the heavy lifting.

Core Features & Use Cases

• Automated Job Submission: Submits all finetune.slurm jobs with appropriate staggering to prevent cache issues.
• Real-time Status Monitoring: Polls SLURM (squeue, sacct) to track job states (PENDING, RUNNING, COMPLETED, FAILED) for all your runs.
• Automatic Status Updates: Keeps your experiment_summary.md updated with job IDs, start/completion times, and current status for each run.
• Use Case: After scaffolding an experiment with 10 fine-tuning runs, use this skill to submit all jobs to the cluster and then walk away. The skill will monitor their progress, update your experiment plan, and notify you when everything is complete, allowing you to focus on analysis.

Quick Start

Assuming an experiment has been scaffolded and finetune.slurm files exist:

claude run-experiment

The skill will submit jobs and monitor their progress,

updating experiment_summary.md automatically until all jobs are complete.