Rock particle segmentation is essential for identifying thin sections accurately, yet tuning deep learning models for this task is challenging. Existing studies have attempted auto-tuning techniques but face three limitations, including limited search space, inefficient search, and high demand for storage space. In this paper, we introduce HyperRPS, a novel multi-fidelity hyper-parameter tuning framework tailored for rock particle segmentation, which addresses these issues simultaneously. First, by augmenting the search space with diverse hyper-parameters, we unleash the potential of the segmentation model. Second, to accelerate the hyper-parameter search, we propose a tuning algorithm leveraging multi-fidelity surrogate fitting and time constraint modeling. Third, we develop a storage management technique to reduce the space requirements for asynchronous scheduling. Experimental results on both public and rock particle segmentation datasets demonstrate HyperRPS’s superior performance. Specifically, on the rock particle segmentation dataset, HyperRPS achieves notable improvements of 8.26% and 2.78% on average precision when tuning Mask2Former (Swin-B) and Mask2Former (Swin-L), respectively, outperforming existing state-of-the-art methods.

Large language models (LLMs) have demonstrated remarkable generalization capabilities across diverse domains, largely attributed to the availability of massive amounts of high-quality training data. Recently, the development paradigm of LLMs has been shifting from a model-centric to a data-centric perspective. In this paper, we provide a comprehensive survey of data preparation algorithms and workflows for LLMs, categorized into three stages: pre-training, continual pre-training, and post-training. We further summarize widely used datasets along with their associated data preparation method, offering a practical reference for researchers who may lack extensive experience in the field of data preparation. Finally, we outline potential directions for future work, highlighting open challenges and opportunities in advancing data preparation for LLMs.

Generative models have achieved significant success in image and audio tasks and have recently been adapted to handle textual data by learning discrete data distributions. However, their application to trust relationship generation remains largely unexplored. This paper addresses this gap by investigating how online social networks structures influence the prediction of trust relationships through structural analysis. We propose DiffTrust++, a novel self-conditioned diffusion model tailored for generating trust relationships in social environments. First, we learn and encode structural equivalence in online social networks and user representations. Second, recognizing that online social networks often follow power-law distributions, we incorporate diffusion degrees to model varying densities under different noise distributions. Finally, asymmetric trust relationships among social entities are generated and evaluated based on their structural features. To evaluate the effectiveness of DiffTrust++, we conducted experiments using real-world datasets and compared our model with benchmark approaches. The experimental results clearly show that our model excels in generating trust relationships, outperforming the alternatives.

Most traditional distributed deep learning training systems have been based on parameter servers which have centralized communication architectures that face serious communication bottlenecks due to the large amounts of communications and AllReduce communication frameworks which have decentralized communication architectures that cannot store the entire model due to the large number of parameters. This paper presents PS-Hybrid, a hybrid communication framework, for large deep learning recommendation model training which decouples the communication logic from the embedded parameters and other parameters. Tests show that this prototype system achieves better performance than previous parameter servers for recommendation model training. The system is 48% faster than TensorFlow-PS with 16 computing nodes.