Large language models (LLMs) are impressive, but their massive size makes them resource-intensive. Pruning, a technique that trims unnecessary parts of the model, helps make LLMs smaller and more efficient. However, pruning often comes at the cost of reduced performance. To recover this lost performance, researchers use post-training. But how much post-training data is enough? Researchers explored this question and discovered a fascinating relationship they've dubbed the "P² Law" (Post-training after model Pruning Law). This law reveals how four key factors interact to predict the performance of a pruned LLM after post-training: the original model size, the amount of post-training data, the aggressiveness of the pruning (pruning rate), and the model's initial performance before pruning. Through experiments on popular LLMs like Llama-3 and Qwen-2.5, the researchers demonstrated that the P² Law accurately predicts how much post-training data is needed to recover performance after pruning. This is a significant step towards making LLMs more accessible and less computationally expensive. The law even generalizes across different datasets, model sizes, and pruning rates, suggesting its broad applicability. This opens up exciting possibilities for optimizing LLMs and deploying them in real-world applications where resources are limited. While the current research primarily focuses on specific LLM architectures, future work aims to extend the P² Law to other architectures like Mixture of Experts (MoE), further expanding its usefulness in the ever-evolving landscape of AI.