ko-dialect-converter

Get Stage-3 GRPO running and converging on a small GPU. Use when GRPO "won't run", crashes on launch, OOMs, is impossibly slow, throws bitsandbytes/libnvJitLink errors, "Attempting to unscale FP16 gradients", TRL "unexpected keyword argument 'max_prompt_length'", or when training runs but reward-hacks / doesn't converge (completions pinned at max length, grad_norm nan). Trigger on "GRPO 안 돌아가", "GRPO 터진다", "reward hacking", "max_prompt_length 에러", "fp16 에러", "4bit 너무 느려".

Diagnose "metrics not showing up in Weights & Biases" problems — when a run appears in the dashboard with config/summary but the history panels say "There's no data for the selected runs", or charts are empty. Separates a client/code bug from a server-side issue (ingestion incident, quota, auth) using the scripts/wandb_doctor.py funnel. Trigger when someone says wandb logging "isn't working", metrics/charts are missing, "summary만 보인다", "history가 안 올라간다", or a training run logs but panels stay empty.