By retaining only the parameters with the largest change in magnitude (e.g., the top-k percentile), this process filters out redundant or less impactful parameters

Trim:This step zeroes out parameters that underwent minimal change during the fine-tuning of each LoRA module