Мониторинг ML-экспериментов на удалённых GPU

★ 7.1 · devops

monitor-experiment is a Claude Code skill that tracks the progress of running experiments on remote servers, including Vast.ai and Modal instances, and aggregates their outputs. It operates via SSH and Bash, checking active screen sessions, reading JSON result files, and — when `wandb: true` is set in CLAUDE.md — pulling training curves and metrics (train/loss, eval/loss, eval/ppl, learning rate) through the Weights & Biases Python API. Results are presented as a comparison table with deltas against a baseline, flagging divergence or NaN values and suggesting next steps. The skill tracks Vast.ai billing and prompts the user to destroy idle instances, while Modal apps are noted as auto-terminating with no cleanup required. Feishu notifications are sent after collection if configured. Designed for ML engineers who need a single command to audit GPU training runs across multiple providers.