PASER introduces a targeted approach to recover pruned language model capabilities with minimal computational cost through selective data use.

• Addresses uneven capability degradation after model pruning
• Uses importance sampling to select optimal recovery data
• Achieves 94% of full fine-tuning performance using only 25% of the data
• Reduces negative transfer that can harm recovery efforts

This research enables more efficient deployment of compressed models in resource-constrained environments, offering practical solutions for engineering teams looking to balance model size and performance.

PASER: Post-Training Data Selection for Efficient Pruned Large Language Model Recovery