triage-sft.yaml
Every run trains with AdamW, a linear warmup over the first 3% of steps, then cosine decay, and weight decay 0.0. A
lora block applies the adapter to all linear layers, with no dropout. No recipe key changes these settings.
Each method adds its own keys under its name, such as grpo.group_size. The method pages list them: SFT, GRPO, DPO, and Distillation.
--recipe triage-sft finds recipes/triage-sft.yaml, and --recipe also takes a path. Each run records the hash of the file, so an edit makes a new version of the recipe. --model replaces model for one run.
