A QLoRA run matches full fine-tuning on style but fails...

The failure to recall product codes is a symptom of the difference between style transfer and knowledge acquisition. LoRA is effective for adapting the model’s tone or format, but it is not efficient for memorizing thousands of arbitrary, disconnected strings. Increasing the rank might help at the margin but risks losing the forgetting resistance gained by using low-rank adaptation. The optimal solution is to treat product codes as data to be retrieved, not as parameters to be learned.