Neural network pruning, the process of reducing network size by removing parameters, has seen significant interest in recent years. This paper provides a meta-analysis of 81 papers on pruning, highlighting consistent findings and the lack of standardized benchmarks and metrics. The authors identify issues with current practices, suggest best practices, and introduce ShrinkBench, an open-source framework for standardized evaluations of pruning methods. Key findings include the effectiveness of pruning in reducing model size without significant loss of accuracy, the superiority of structured pruning over unstructured pruning, and the importance of standardized datasets, networks, metrics, and experimental practices. The paper also discusses the challenges of comparing pruning methods due to fragmentation in reported results and confounding variables, and provides recommendations for future research to improve the reproducibility and comparability of pruning studies.Neural network pruning, the process of reducing network size by removing parameters, has seen significant interest in recent years. This paper provides a meta-analysis of 81 papers on pruning, highlighting consistent findings and the lack of standardized benchmarks and metrics. The authors identify issues with current practices, suggest best practices, and introduce ShrinkBench, an open-source framework for standardized evaluations of pruning methods. Key findings include the effectiveness of pruning in reducing model size without significant loss of accuracy, the superiority of structured pruning over unstructured pruning, and the importance of standardized datasets, networks, metrics, and experimental practices. The paper also discusses the challenges of comparing pruning methods due to fragmentation in reported results and confounding variables, and provides recommendations for future research to improve the reproducibility and comparability of pruning studies.