FLoRA: Federated Fine-Tuning Large Language Models with Heterogeneous Low-Rank Adaptations

The rapid development of Large Language Models (LLMs) has been pivotal in advancing AI, with pre-trained LLMs being adaptable to diverse downstream tasks through fine-tuning. Federated learning (FL) further enhances fine-tuning in a privacy-aware manner by utilizing clients' local data through in-situ computation, eliminating the need for data movement. However, fine-tuning LLMs, given their massive scale of parameters, poses challenges for clients with constrained and heterogeneous resources in FL. Previous methods employed low-rank adaptation (LoRA) for efficient federated fine-tuning but utilized traditional FL aggregation strategies on LoRA adapters. These approaches led to mathematically inaccurate aggregation noise, reducing fine-tuning effectiveness and failing to address heterogeneous LoRAs. In this work, we first highlight the mathematical incorrectness of LoRA aggregation in existing federated fine-tuning methods. We introduce a new approach called FLORA that enables federated fine-tuning on heterogeneous LoRA adapters across clients through a novel stacking-based aggregation method. Our approach is noise-free and seamlessly supports heterogeneous LoRA adapters. Extensive experiments demonstrate FLORA' s superior performance in both homogeneous and heterogeneous settings, surpassing state-of-the-art methods. We envision this work as a milestone for efficient, privacy-preserving, and accurate federated fine-tuning of LLMs. Our code is available at https://github.com/ATP-1010/FederatedLLM.

SHED: Shapley-Based Automated Dataset Refinement for Instruction Fine-Tuning

The pre-trained Large Language Models (LLMs) can be adapted for many downstream tasks and tailored to align with human preferences through fine-tuning. Recent studies have discovered that LLMs can achieve desirable performance with only a small amount of high-quality data, suggesting that a large amount of the data in these extensive datasets is redundant or even harmful. Identifying high-quality data from vast datasets to curate small yet effective datasets has emerged as a critical challenge. In this paper, we introduce SHED, an automated dataset refinement framework based on Shapley value for instruction fine-tuning. SHED eliminates the need for human intervention or the use of commercial LLMs. Moreover, the datasets curated through SHED exhibit transferability, indicating they can be reused across different LLMs with consistently high performance. We conduct extensive experiments to evaluate the datasets curated by SHED. The results demonstrate SHED's superiority over state-of-the-art methods across various tasks and LLMs; notably, datasets comprising only 10% of the original data selected by SHED achieve performance comparable to or surpassing that of the full datasets.

FedHyper: A Universal and Robust Learning Rate Scheduler for Federated Learning with Hypergradient Descent

The theoretical landscape of federated learning (FL) undergoes rapid evolution, but its practical application encounters a series of intricate challenges, and hyperparameter optimization is one of these critical challenges. Amongst the diverse adjustments in hyperparameters, the adaptation of the learning rate emerges as a crucial component, holding the promise of significantly enhancing the efficacy of FL systems. In response to this critical need, this paper presents FedHyper, a novel hypergradient-based learning rate adaptation algorithm specifically designed for FL. FedHyper serves as a universal learning rate scheduler that can adapt both global and local rates as the training progresses. In addition, FedHyper not only showcases unparalleled robustness to a spectrum of initial learning rate configurations but also significantly alleviates the necessity for laborious empirical learning rate adjustments. We provide a comprehensive theoretical analysis of FedHyper's convergence rate and conduct extensive experiments on vision and language benchmark datasets. The results demonstrate that FEDHYPER consistently converges 1.1-3x faster than FedAvg and the competing baselines while achieving superior final accuracy. Moreover, FedHyper catalyzes a remarkable surge in accuracy, augmenting it by up to 15% compared to FedAvg under suboptimal initial learning rate settings.

A Method to Judge the Style of Classical Poetry Based on Pre-trained Model

One of the important topics in the research field of Chinese classical poetry is to analyze the poetic style. By examining the relevant works of previous dynasties, researchers judge a poetic style mostly by their subjective feelings, and refer to the previous evaluations that have become a certain conclusion. Although this judgment method is often effective, there may be some errors. This paper builds the most perfect data set of Chinese classical poetry at present, trains a BART-poem pre -trained model on this data set, and puts forward a generally applicable poetry style judgment method based on this BART-poem model, innovatively introduces in-depth learning into the field of computational stylistics, and provides a new research method for the study of classical poetry. This paper attempts to use this method to solve the problem of poetry style identification in the Tang and Song Dynasties, and takes the poetry schools that are considered to have a relatively clear and consistent poetic style, such as the Hongzheng Qizi and Jiajing Qizi, Jiangxi poetic school and Tongguang poetic school, as the research object, and takes the poems of their representative poets for testing. Experiments show that the judgment results of the tested poetry work made by the model are basically consistent with the conclusions given by critics of previous dynasties, verify some avant-garde judgments of Mr. Qian Zhongshu, and better solve the task of poetry style recognition in the Tang and Song dynasties.

Generation of Chinese classical poetry based on pre-trained model

In order to test whether artificial intelligence can create qualified classical poetry like humans, the author proposes a study of Chinese classical poetry generation based on a pre-trained model. This paper mainly tries to use BART and other pre training models, proposes FS2TEXT and RR2TEXT to generate metrical poetry text and even specific style poetry text, and solves the problem that the user's writing intention gradually reduces the relevance of the generated poetry text. In order to test the model's results, the authors selected ancient poets, by combining it with BART's poetic model work, developed a set of AI poetry Turing problems, it was reviewed by a group of poets and poetry writing researchers. There were more than 600 participants, and the final results showed that, high-level poetry lovers can't distinguish between AI activity and human activity, this indicates that the author's working methods are not significantly different from human activities. The model of poetry generation studied by the author generalizes works that cannot be distinguished from those of advanced scholars. The number of modern Chinese poets has reached 5 million. However, many modern Chinese poets lack language ability and skills as a result of their childhood learning. However, many modern poets have no creative inspiration, and the author's model can help them. They can look at this model when they choose words and phrases and they can write works based on the poems they already have, and they can write their own poems. The importance of poetry lies in the author's thoughts and reflections. It doesn't matter how good AI poetry is. The only thing that matters is for people to see and inspire them.

A General Framework for Defending Against Backdoor Attacks via Influence Graph

In this work, we propose a new and general framework to defend against backdoor attacks, inspired by the fact that attack triggers usually follow a \textsc{specific} type of attacking pattern, and therefore, poisoned training examples have greater impacts on each other during training. We introduce the notion of the {\it influence graph}, which consists of nodes and edges respectively representative of individual training points and associated pair-wise influences. The influence between a pair of training points represents the impact of removing one training point on the prediction of another, approximated by the influence function \citep{koh2017understanding}. Malicious training points are extracted by finding the maximum average sub-graph subject to a particular size. Extensive experiments on computer vision and natural language processing tasks demonstrate the effectiveness and generality of the proposed framework.

Learning from Incomplete Data by Simultaneous Training of Neural Networks and Sparse Coding

Handling correctly incomplete datasets in machine learning is a fundamental and classical challenge. In this paper, the problem of training a classifier on a dataset with missing features, and its application to a complete or incomplete test dataset, is addressed. A supervised learning method is developed to train a general classifier, such as a logistic regression or a deep neural network, using only a limited number of features per sample, while assuming sparse representations of data vectors on an unknown dictionary. The pattern of missing features is allowed to be different for each input data instance and can be either random or structured. The proposed method simultaneously learns the classifier, the dictionary and the corresponding sparse representation of each input data sample. A theoretical analysis is provided, comparing this method with the standard imputation approach, which consists of performing data completion followed by training the classifier with those reconstructions. Sufficient conditions are identified such that, if it is possible to train a classifier on incomplete observations so that their reconstructions are well separated by a hyperplane, then the same classifier also correctly separates the original (unobserved) data samples. Extensive simulation results on synthetic and well-known reference datasets are presented that validate our theoretical findings and demonstrate the effectiveness of the proposed method compared to traditional data imputation approaches and one state of the art algorithm.