Beware of Calibration Data for Pruning Large Language Models

Thanks for your time and effort in reviewing our paper. We wonder if there might be some misunderstandings about our work, and we hope our response will clarify these points.

• Lack of Novel Contribution

Thanks for your feedback. We try to look for Williams & Altetras (EMNLP2024) as you mentioned, but Williams & Altetras do not publish in EMNLP 2024. We speculate that you might be referring to the work of Williams & Aletras (ACL 2024)[1]. In the related work section (Lines 133-137), we briefly discussed the differences between Williams & Aletras' work. Here, we provide a more detailed explanation:

(1) Williams & Aletras aimed to explore the impact of the source and amount of calibration data on LLM compression but did not provide specific guidance on how to select suitable calibration data. In contrast, our work further addresses which calibration data is suitable for LLM pruning and provides a practical and effective method.

(2) Williams & Aletras first highlighted the importance of calibration data for LLM compression, particularly in post-training pruning methods. However, they did not quantify its impact. Through extensive experiments, we explore the impact of calibration data under varying sparsity ratios and types. Our results show that as the sparsity ratio increases and the sparsity type becomes more structural, the choice of calibration data becomes increasingly critical. Additionally, we compare several advanced pruning algorithms and find that the impact of calibration data surpasses the performance gains from improvements in pruning algorithms.

(3) Williams & Aletras explore the impact of calibration data amount on model compression, but they were limited to the common-used C4 calibration data. In contrast, our research compares the performance of pruned models with calibration data from different sources as the data amount increases. We find that increasing the amount of calibration data from different sources does not improve pruning performance, indicating that the performance differences between different calibration datasets cannot be compensated for by simply increasing the data amount.

If you are not referring to the aforementioned paper, we sincerely hope you can provide a more specific title so that we can address your concerns.

• Lack of downstream tasks experiments

In our experiments, we evaluate the performance of LLMs on various downstream tasks (see Lines 189-196), including BoolQ, Winograde, PIQA, Hellaswag, ARC-e, ARC-c, and MMLU. The table below lists the evaluation datasets used in 6 papers. Most works evaluate performance on 6-7 tasks. We use the six most commonly used datasets and additionally evaluate the MMLU dataset. MMLU alone includes 57 different tasks, so we think our evaluation on downstream tasks is also sufficiently solid.

To address your concerns, we report the results of Wanda for the remaining four tasks on DCLM-7B in the below table, showing that pruned models using pretraining data or synthetic data as calibration significantly outperform those using other calibration data.

• No explanation on pruning performance

We provide a detailed description of our pruning performance measure in lines 189-196. It measures the average accuracy of the pruning model on multiple commonsense reasoning tasks. SparseGPT, Wanda, DsnoT, OWL, and Williams & Aletras (2024 ACL) all evaluate pruning model performance based on commonsense reasoning tasks.

[1] Miles Williams and Nikolaos Aletras. On the impact of calibration data in post-training quantization and pruning. 2024 ACL

Thanks for your careful review and insightful advice. The acknowledgment of our paper's contributions and the accurate summation of its highlights give us the confidence to continue improving it.

• Impact of synthetic calibration data on structured pruning

We evaluate the effectiveness of our synthetic calibration data on LLM-Pruner. The results are as follows:

Based on these results, we believe that synthetic calibration data can further improve the performance of structured pruning.

• The main differences between this work and Shin et al. (2024.6 arxiv)

Thank you for mentioning the work of Shin et al. We apologize for overlooking this important related work in our paper. We will include a citation and discussion of the differences in the revised version. Shin et al. primarily focused on optimizing the objective for post-training pruning and proposed various strategies to reduce reconstruction error. They observed that MSE objectives tend to overfit calibration data and suggested using self-generated calibration data to mitigate this issue. However, their evaluation was limited to improvements in reconstruction error and did not report performance on downstream tasks. In contrast, our work focuses on a comprehensive investigation of the impact of calibration data on pruned model performance, addressing key questions such as when calibration data matters, whether performance gaps between different calibration datasets can be narrowed, and how to construct appropriate calibration data. Additionally, the details of our synthetic calibration data generation method differ from theirs.

• The effectiveness of synthetic data generated from other sources

We use C4, Slimpajama, Wikipedia,DCLM to generate synthetic calibration data. The results are shown below:

We find that regardless of the source of the synthetic data, the performance of the pruned model shows minimal differences and significantly outperforms the original data (except for DCLM, which is the part of the training data).

• Explain the Figure 6

We apologize for our negligence. The gray dashed lines in Figure 6, 7, 8 represent magnitude-based pruning.

• Apply to LLM quantization

Williams & Aletras[1] find that existing quantization methods are very robust to calibration data, so our paper focuses solely on LLM pruning. We reasonably speculate that synthetic calibration data will not lead to significant performance improvements for LLM quantization.

[1] Miles Williams and Nikolaos Aletras. On the impact of calibration data in post-training quantization and pruning. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 10100–10118, Bangkok, Thailand, August 2024. Association for Computational Linguistics. URL https://aclanthology.org/2024.acl-long.544.

Thanks for the response. After reading carefully, I am satified with the work and will keep the score.

BTW, I can't entirely agree with reviewer 4fLd's point of overlap with EMNLP-2024, as no evidence has been provided to support this issue. In addition, the EMNLP-2024 conference has just been held, and the submission time of ICLR is much earlier than the EMNLP-2024 accepted notification. It is unclear what the Limited Novelty will be.

There was no intention of offense. It was just my personal opinion.

Thanks for your time and effort in providing valuable feedbacks, which is crucial in improving our work.

• Decompose or improve the table to highlight matching or exceeding the performance of using calibration data from the actual training set and exceeding the performance compared to calibration datasets belonging to other distributions.

Following the above suggestion, we have made the following modifications in the updated PDF file:

1. In Table 2, we clearly distinguish the relationship between calibration data and pretraining data. For DCLM-7B, DCLM is its open-source pretraining data, Wikipedia does not belong to pretraining data, while C4 and Slimpajama partially overlap with the pretraining data.

2. To clearly demonstrate the performance improvements of synthetic data, we add results using synthetic data generated from C4, Slimpajama, and DCLM. Regardless of the data source, models pruned with self-generated calibration data consistently surpass those using the original calibration data by up to 2.68% of averaged performance across different tasks.

3. Since we add more experiments with synthetic data, we move the results for LLaMA 2-7B to the Appendix to avoid Table 2 taking up too much space.

4. We revise the description of the experimental results in Section 5.2 based on the updated main results.

• Avoid redundancy in repeating the literature review and summarize the questions in the introduction.

1. We remove repeated literature review either from introduction or background for conciseness.

2. We also improve the description of related work by mentioning the name of the technique corresponding to each citation.

3. Additionally, we have summarized the questions in the introduction.

We summarize "However, many open questions regarding calibration data remain under-explored. For example, how does the impact of calibration data change with increased sparsity and structure of pruning? Can increasing the amount of calibration data narrow the performance gap between various datasets? What type of data is suitable for calibration? How do we select the appropriate calibration data in practice?" as "To learn more about calibration data, we design experiments to explore (1) the impact of calibration data with increased sparsity and varied pruning types, (2) the influence of the amount of calibration data, and (3) the selection strategy of calibration data."

• Improve the presentation of abstract to better reflect the outcomes of the paper and be easier to read.

1. We have improved sentences to avoid inadequate wording. We change "We fill this blank and surprisingly observe that the effects of calibration data even value more than designing advanced pruning strategies." to "We fill this blank and surprisingly observe that calibration data is also crucial to post-training pruning, especially for high sparsity. "

2. We have simplified complex sentences in the abstract. We replace "Previous research has primarily focused on designing advanced pruning methods, while different calibration data’s impact on pruning performance still lacks systematical exploration." with "Recent research has enhanced post-training pruning from different aspects but few of them systematically explore the effects of calibration data, and it is unclear if there exist better calibration data construction strategies."

3. We highlight the outcomes of our paper in the abstract. "Through controlled experiments on important influence factors of calibration data, including the pruning settings, the amount of data, and its similarity with pre-training data, we observe that a small size of data is adequate, and more similar data to its pre-training stage can yield better performance."

• Mention somewhere that the paper will first proceed by answering the calibration data related questions and then propose a new a novel technique for its generations.

Follow the suggestions, we have clearly sumamrize the main novel contributon of this work in Absratct and compare this works with the most relevant work in Background.

Thank you once again for your detailed and valuable suggestions to improve the presentation of our paper. We have revised the paper with all changes highlighted in blue. We look forward to your inspiring response.

Thanks a lot for your valuable comments. We hope the following responses can address your concern:

• Explain the correlation between training data similarity and pruning performance

We briefly analyze this issue in lines 379-383 of the paper. To address your concerns, we provide a more detailed explanation here. Based on the optimization objective of post-training pruning:

\underset{\hat{\boldsymbol{W}}_l}{\text{min}} \left\\| \boldsymbol{W}_l \boldsymbol{X}_l - \hat{\boldsymbol{W}}_l \boldsymbol{X}_l \right\\|_F,

calibration data affects $X_l$ in the formula. If the calibration data is not representative, the $X_l$ may not adequately reflect the LLM's broad general capabilities. This underrepresentation can lead to biased measurements of the importance of $W_l$ parameters during pruning, affecting performance. Through LLM pretraining, the training data is typically well represented by the LLM and captures most of its general abilities. Therefore, we believe using training data or data similar to it as calibration data defines a better pruning objective, enabling a more accurate estimation of LLM parameter importance while preserving the LLM's general capabilities.

• The fine-grained impact of calibration data on different modules.

Thanks for your valuable suggestion, which is highly insightful for understanding the impact of calibration data on pruning. Based on your suggestion, we conduct preliminary experiments to explore the impact of calibration data on sparsity ratio allocation across layers. We take OWL as an example and report the layer-wise pruning ratio allocation under the 60% sparsity ratio in the table below:

Comparing the two allocation results, we find that when using synthetic data as calibration data, OWL automatically assigns higher pruning ratios to the middle layers (layers 20–30) while assigning lower ratios to other layers. This is consistent with the conclusion in ShortGPT[1] that middle-to-late layers are less important, while the layers at both ends are more critical. Thus, we think suitable calibration data not only improves the accuracy of parameter importance estimation but also optimizes layer-wise importance estimation, enabling more reasonable pruning ratio allocation.

As for more fine-grained studies, such as specific attention heads, frankly, due to our limited understanding of how different attention heads affect model performance, we are currently unable to draw unified conclusions based on pruning results from different calibration data. Therefore, we have to leave this insightful topic for future research.

• Explain the Figure 5

Min-K%++ is a metric for pretraining data detection, where higher values indicate greater similarity to pretraining data and a higher likelihood of being part of it [2]. Figure 5 uses kernel density estimation to show the distribution of Min-K%++ values for C4, SlimPajama, Wikipedia, and self-generated synthetic data. The distribution curve for synthetic data is noticeably right-skewed, indicating that its Min-K%++ values are generally higher than the other three calibration data.

[1] Men, Xin, et al. "Shortgpt: Layers in large language models are more redundant than you expect." arXiv preprint arXiv:2403.03853 (2024).

[2] Zhang, Jingyang, et al. "Min-k%++: Improved baseline for detecting pre-training data from large language models." arXiv preprint arXiv:2404.02936 (2024).