From Parameters to Data: A Task-Parameter-Guided Fine-Tuning Pipeline for Efficient LLM Alignment

In three linesP2D, an LLM alignment framework, couples data selection with parameter-efficient fine-tuning by identifying task-critical attention heads. It mines high-affinity data and prunes 90% of parameters using these heads as a functional filter. Result: +8.3pp performance gain and 7.0x end-to-end speedup using only 10% of data and 10% of heads.

Summary generated by Claude — human-verified