What problem does it solve? Running LLM training experiments produces scattered evidence across W&B offline runs, remote server logs, and Prometheus metrics, making it hard to reproduce results or build auditable experiment records. This Skill consolidates all of that into immutable local-first run bundles with manifests and config locks. ## Core Features & Use Cases - Minimal Experiment Execution: Runs trainer.train() for 1-2 steps via the minimal runner to generate smoke-test evidence quickly. - W&B Offline Sync: Syncs offline W&B runs locally with semantic project/run naming (e.g., verl project, run names encoding model, scale, algorithm, and sequence length), and saves runs.json plus rebuild-wandb.sh so the data bundle can rebuild the same local W&B view. - Remote Log & Prometheus Collection: Pulls remote training logs back to local storage and pushes NPU metrics to a local Pushgateway, explicitly recording which metrics (e.g., HBM/Core utilization) are missing. - Use Case: After running a formal Verl GRPO case on a remote NPU server, use this Skill to produce a self-contained evidence bundle with logs/, wandb/, prom/, and manifest.json that supports bundle-local fallback loading without relying on old absolute paths. ## Quick Start Run the data collection script with your config file, target server, and a run ID to execute a minimal experiment and gather all evidence into a local bundle.
