What problem does it solve?
Running ML experiments on remote servers makes it hard to know whether training is progressing, finished, or failed without manually SSHing in and digging through logs. This Skill automates checking running screens, collecting output, parsing JSON results, and pulling W&B metrics so you get a clear status summary.
Core Features & Use Cases
- Remote Session Inspection: Lists active screen sessions over SSH on standard servers and vast.ai instances, and checks Modal app status for cloud runs.
- Result Collection: Captures screen output, fetches JSON result files, and pulls training curves and metrics from Weights & Biases via its Python API.
- Comparison & Interpretation: Summarizes metrics in a baseline comparison table, flags anomalies like NaN or divergence, and reports running costs for vast.ai and Modal.
- Use Case: You launched three fine-tuning runs on a vast.ai GPU instance last night. Ask the assistant to check results, and it returns a table comparing each run's eval loss against the baseline, plus the W&B dashboard link and a reminder to destroy the instance to stop billing.
Quick Start
Ask the assistant to check the results of the experiment running on your GPU server and summarize how it compares to the baseline.
