Search papers, labs, and topics across Lattice.
This paper introduces ZipCal, a model-agnostic data curation strategy for post-training compression of LLMs that selects calibration data based on Zipfian power laws to maximize lexical diversity. ZipCal aims to improve pruning and quantization performance by focusing on intrinsic data properties rather than model-specific signals. Experiments show ZipCal outperforms uniform random sampling and achieves comparable downstream performance to perplexity-based methods, while being significantly faster (240x).
Forget perplexity – ZipCal uses Zipf's law to curate calibration data for LLM compression, matching state-of-the-art performance at 240x the speed.
Post-training model compression is essential for enhancing the portability of Large Language Models (LLMs) while preserving their performance. While several compression approaches have been proposed, less emphasis has been placed on selecting the most suitable set of data (the so-called \emph{calibration data}) for finding the compressed model configuration. The choice of calibration data is a critical step in preserving model capabilities both intra- and inter-tasks. In this work, we address the challenge of identifying high-performance calibration sets for both pruning and quantization by analyzing intrinsic data properties rather than model-specific signals. We introduce \texttt{\textbf{ZipCal}}, a model-agnostic data curation strategy that maximizes lexical diversity based on Zipfian power laws. Experiments demonstrate that our method consistently outperforms standard uniform random sampling across various pruning benchmarks. Notably, it also performs on par, in terms of downstream performance, with a state-of-the-art method that relies on model perplexity. The latter becomes prohibitively expensive at large-scale models and datasets, while \texttt{\textbf{ZipCal}} is on average $\sim$240$\times$ faster due to its tractable linear complexity\footnote{We make the code and the experiments available at https://anonymous.4open.science/r/zipcal-71CD/.}.