Fill in typical peak LRs for full FT, LoRA, and DPO

Learning rate controls how big a step the optimizer takes on each update. Set it too high and training diverges or the model forgets its base skills. Set it too low and it barely learns. The right value depends on which kind of tuning you are doing. Full fine-tuning moves every weight in the model, so you keep the steps small. For a 7B model, peak learning rate usually lands between **1e-5** and 5e-5. Bigger steps risk wrecking the pretrained knowledge. LoRA freezes the base model and trains only small adapter matrices that start at zero. Because so little is moving, you can afford roughly ten times the full FT rate, so peak around 1e-4 to 5e-4 is normal. DPO teaches the model which answer to prefer, and that signal is fragile. Too large a step and the model collapses into repetitive or broken text. So DPO uses a much lower rate, around 5e-7 to 5e-6. In every case you ramp up slowly with a short warmup, then decay along a cosine curve toward zero.

Learning rate scales with how much of the model you are moving and how fragile the training signal is. The standard pattern is a cosine schedule with a short linear warmup. Full fine-tuning of a 7B model touches every weight, so peak learning rate stays small, roughly 1e-5 to 5e-5. LoRA freezes the base and trains tiny adapters that initialize to zero. That isolation lets it tolerate about ten times more, so peak typically sits between 1e-4 and 5e-4 for rank 16. DPO is the outlier. It optimizes a preference objective from an already-tuned checkpoint, and the gradient is sharp near the reference policy. Overshoot and you get mode collapse or degenerate text. So DPO runs an order or two lower than SFT, around 5e-7 to 5e-6. Warmup matters as much as the peak. Most recipes warm up over 3 to 10 percent of steps, then decay along cosine to near zero. Warmup stabilizes the early high-variance updates; cosine decay lets the model settle into a sharp minimum at the end.