What problem does it solve? Fine-tuning runs often adopt a learning rate copied from a paper or hard-coded in a plan, then get judged by healthy-looking loss or reward curves that prove nothing. This Skill enforces a discipline where the learning rate is always swept over a wide range and judged only against explicit pass criteria. ## Core Features & Use Cases - Mandatory LR Sweeping: Treats any fixed or paper-copied learning rate as a single grid point and requires sweeping a wide range for full fine-tuning, LoRA, QLoRA, DoRA, and other PEFT adapters. - Criteria-Based Acceptance: Accepts a learning rate only when the run meets the pass criteria in task.md or the milestone's claim criteria, rejecting loss, grad norm, and reward curves as evidence. - Iterative Re-Run Protocol: When criteria are not met, changes the learning rate and re-runs until an LR passes, covering SFT, DPO, GRPO, PPO, and RLHF objectives. - Use Case: While running an SFTTrainer job that converges smoothly but fails the task's evaluation criteria, apply this Skill to reject the run, sweep the learning rate across a wide range, and re-run until a configuration passes. ## Quick Start Sweep the learning rate over a wide range for this LoRA fine-tune and re-run until the task.md pass criteria are met.
