shaojieMike

You can run Verl GRPO training cases, probe remote GPU/NPU hardware, health-check training stacks and conda environments, collect experiment evidence from W&B and Prometheus, render HTML experiment reports, manage model and dataset assets, clean empty W&B runs, and produce RMB token cost reports.

ML engineers and researchers operating reinforcement learning training pipelines, particularly teams using Verl with Qwen models and geo3k datasets on remote GPU or Huawei Ascend NPU servers who need structured experiment evidence, loop constraints, and cost accountability.

Loop-constraints loads enforcement rules before triage, edits, or remote work; loop-budget enforces cadence, attempt, token, time, sub-agent, and kill-switch limits; loop-triage reads run evidence and loop state to propose one bounded next experiment or escalate, with L1 being report-only.

Skills assume SSH access to remote training servers with npu-smi or nvidia-smi, local services including W&B, Prometheus, Grafana, and Pushgateway, conda environments with verl or veomni installed, Pydantic-based config files, and optional proxy fallback at 127.0.0.1:7890 for restricted networks.

It scans Verl example scripts and YAML files to extract all user-adjustable parameters and Hydra overrides, creates per-parameter Chinese markdown explanations under docs/verl/features/lists, and produces an Excel ledger with parameter category, common values, performance impact, accuracy impact, and example counts.