lora merge destroyed performance, only trained for 200 steps

trained a lora on llama 3.1 8b for a specific classification task, 200 steps, loss looked good (dropped from 1.8 to 0.3). merged it back into the base model and now the model is worse than baseline on everything including the task i trained it on base model: 71% on my eval fine-tuned: 58% i must be doing something wrong with the merge but i just used the standard peft merge script. is 200 steps too few or did i wreck the model some other way