What problem does it solve? Tracking long-running machine learning experiments across SSH servers, Vast.ai instances, and Modal apps is tedious and error-prone. This Skill checks what is running, collects screen output and JSON results, pulls Weights & Biases metrics, and finalizes per-run cost records so you always know experiment status and spend. ## Core Features & Use Cases - Status Checks: Inspect screen sessions over SSH, query Vast.ai instance state, and list Modal apps to determine whether runs are active, finished, or crashed. - Result Collection: Capture screen output, fetch JSON result files, and pull training curves and eval metrics from the W&B API for comparison against baselines. - Cost Finalization: Write canonical cost.json manifests with wall-clock time, GPU hours, and run status for downstream budget tracking, plus cost reminders for Vast.ai and Modal. - Use Case: You launched three fine-tuning runs on a remote GPU server last night. Ask to check results, and get a comparison table of metrics versus baseline, W&B dashboard links, and a reminder to destroy idle Vast.ai instances. ## Quick Start Check the results of my running experiments on the remote server and summarize the metrics compared to baseline.
