Low-rank adaptation loar-based large language model federated fine-tuning method and system

Abstract

This invention discloses a federated fine-tuning method and system for large language models based on low-rank adaptation LoRA. The method includes: a server initializing K sets of LoRA parameters for an L-layer large language model; the server distributing LoRA parameters to K selected clients, ensuring each client receives the corresponding LoRA parameters while maintaining the weights of the pre-trained model; performing LoRA-based large language model fine-tuning training on a local dataset to update the LoRA parameters and submitting the updated LoRA parameters to the server; the server performing adaptive cross-aggregation based on the locally updated LoRA parameters submitted by the K clients; and the server updating the model parameters of the large language model based on the K sets of cross-aggregated LoRA parameters. This invention aims to improve the adaptability and robustness of federated low-rank adaptation fine-tuning of large language models under heterogeneous conditions, thereby enhancing the natural language processing performance of large language models.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method and system for federated fine-tuning of large language models based on low-rank adaptation LoRA. Background Technology

[0002] The rise of the pre-training and fine-tuning paradigm has reshaped traditional training methods for deep learning models, especially in natural language processing tasks, where large language models (LMs) have shown unprecedented performance improvements. However, the significant increase in model size has also brought high training costs and inference latency, posing a serious challenge to practical deployment. To address these issues, the Parameter-Efficient Fine-Tuning (PEFT) method has been proposed and widely applied in various downstream tasks. PEFT aims to update only a small subset of trainable parameters in the pre-trained model, while keeping the remaining parameters unchanged, thereby achieving near-full-scale fine-tuning performance with extremely low computational and storage overhead.

[0003] Typical efficient parameter fine-tuning methods include adapter-based methods, prefix-tuning, and low-rank adaptation (LoRA). These methods significantly improve the flexibility and efficiency of model adaptation, enabling large models to be deployed in resource-constrained environments such as edge devices and federated learning. Among them, low-rank adaptation has attracted widespread attention due to its simple structure and excellent performance. The core idea of ​​low-rank adaptation is to freeze the original model weights during the fine-tuning process and introduce trainable low-rank matrices only in the update directions of certain weight matrices: the dimension-reduced projection matrix A and the dimension-up reconstruction matrix B, thereby significantly reducing the size of training parameters and lowering the computational resource requirements. However, existing federated learning methods based on low-rank adaptation usually adopt a uniform aggregation strategy (such as FedAvg), failing to fully consider the heterogeneity of data distribution among clients and the performance differences of low-rank adaptation parameters across different model layers. In non-independent and identically distributed (non-IID) environments, this consistent aggregation strategy often leads to performance degradation and training instability. Furthermore, most current methods assign the same or static aggregation weights to all clients, failing to dynamically model the differences in parameter evolution and semantic deviations among participants, thus limiting the model's ability to balance generalization and personalization. Uniform or static aggregation strategies cannot capture these fine-grained differences, potentially leading to performance degradation. Summary of the Invention

[0004] The technical problem to be solved by this invention is as follows: Addressing the aforementioned problems in existing technologies, this invention provides a federated fine-tuning method and system for large language models based on Low-Rank Adaptation (LoRA). This invention aims to address the issues of inconsistent client-side optimization directions caused by efficient parameter-based federated fine-tuning scenarios and extreme data heterogeneity, the difficulty of adapting fixed aggregation strategies to multi-task scenarios, and the failure of unified aggregation strategies due to dynamic changes in LoRA parameters between layers. It achieves efficient fine-tuning of large language models under non-independent and identically distributed data, improves the adaptability and robustness of federated low-rank adaptation (LoRA) fine-tuning under heterogeneous conditions, and enhances the natural language processing performance of large language models.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A federated fine-tuning method for large language models based on low-rank adaptation LoRA includes the following steps: S1, The server initializes K sets of LoRA parameters for the L-layer large language model. , }, For the number of clients, ~ These are the LoRA parameters for groups 1 to K; S2, the server sends LoRA parameters to the selected K clients respectively. This enables various clients Receive the corresponding LoRA parameters Keeping the weights of the pre-trained model unchanged, for the local dataset Perform fine-tuning training of a large language model based on LoRA to update LoRA parameters and obtain locally updated LoRA parameters, which are then submitted to the server. S3, the server performs adaptive cross-aggregation based on the locally updated LoRA parameters submitted by K clients, including: for any client Specifically targeting the computational clients of each layer of the large language model. any of the first The LoRA parameter of the layer, and the first digit of any client among the remaining K-1 clients. Calculate the similarity between LoRA parameters of each layer and find the client with the lowest similarity. , will client The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. Cross-aggregation of LoRA parameters in the same layer; S4, the server updates the model parameters of the large language model based on the LoRA parameters obtained from the K groups of cross-aggregation. The updated large language model is then tested to obtain the prediction accuracy. If the prediction accuracy does not meet the requirements, the LoRA parameters obtained from the K groups of client cross-aggregation are used as the new LoRA parameters. If the condition is met, proceed to step S2; otherwise, end and exit.

[0006] Optionally, the LoRA parameter any of the first Group LoRA parameters All include L layers of LoRA parameters, and each layer of LoRA parameters includes a dimension-reduced projection matrix A and a dimension-up reconstruction matrix B.

[0007] Optionally, in step S3, the computation client is calculated for each layer of the large language model. any of the first The LoRA parameter of the layer, and the first digit of any client among the remaining K-1 clients. When calculating the similarity between LoRA parameters of layers, the expression for the similarity calculation function is: ; in, For the client Client The Similarity between LoRA parameters of layers For clients Client The The dimension reduction projection matrix A in the LoRA parameters of the layer, For clients Client The The dimension-upgrading reconstruction matrix B in the LoRA parameters of the layer.

[0008] Optionally, in step S3, the client The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. The function expression for cross-aggregation of LoRA parameters at the same level is: ; ; in, and Clients respectively After cross-aggregation, the first The dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the LoRA parameters of the layer. and For clients The local update The dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the LoRA parameters of the layer. and For clients The local update The weights of the dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the layer LoRA parameters.

[0009] Optionally, in step S3, the client The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. Before cross-aggregating LoRA parameters at the same level, a dynamic calculation client is also included. The local update The weights of the dimension-reduced projection matrix A and the dimension-upgraded reconstruction matrix B in the layer LoRA parameters are: ; in, For the client The local update The weights of the dimension-reduced projection matrix A or the dimension-upgraded reconstruction matrix B in the layer LoRA parameters. To find the maximum value, This is the lower bound of the aggregation weight. and For aggregate weight pairs and training progress factor Sensitivity For the client and client Similarity factor of LoRA parameters in the same layer This represents the number of training rounds.

[0010] Optionally, the client and client The functional expression for the similarity factor of the LoRA parameter in the same layer is: ; in, For the client and client The Similarity factor of layer LoRA parameters This is the similarity calculation function.

[0011] Furthermore, the present invention also provides a large language model federated fine-tuning system based on low-rank adaptation LoRA, including interconnected microprocessors and memory, wherein the microprocessors are programmed or configured to execute the large language model federated fine-tuning method based on low-rank adaptation LoRA.

[0012] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the large language model federated fine-tuning method based on low-rank adaptation LoRA by a processor.

[0013] Furthermore, the present invention also provides a computer program product, including a computer program or instructions that are programmed or configured to execute the large language model federated fine-tuning method based on low-rank adaptation LoRA via a processor.

[0014] Compared with existing technologies, this invention mainly achieves the following beneficial effects: Addressing the problems of high communication overhead, insufficient model adaptability, and decreased convergence performance in LoRA fine-tuning of existing federated learning methods under non-independent and identically distributed (non-IID) data scenarios, this invention introduces a hierarchical cross-aggregation and adaptive weight allocation mechanism. It calculates the similarity and difference of client LoRA parameters at both the global and local levels, selects appropriate aggregation objects, and achieves efficient parameter interaction across clients. Simultaneously, it dynamically adjusts the aggregation weights based on the importance of different layers in model updates, balancing convergence speed and model performance. Experimental results show that this invention significantly outperforms existing methods on multiple benchmark datasets, achieving improvements in accuracy and convergence speed, and enhancing the model's robustness in heterogeneous data environments while reducing training costs. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0016] Figure 2 These are the data analysis results under independent identically distributed (IID) and non-independent identically distributed (non-IID) conditions in the embodiments of the present invention, where (a) is the data analysis result under independent identically distributed (IID) conditions and (b) is the data analysis result under non-independent identically distributed (non-IID) conditions. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0018] like Figure 1 As shown, the federated fine-tuning method for large language models based on low-rank adaptation LoRA in this embodiment includes the following steps: S1, The server initializes K sets of LoRA parameters for the L-layer large language model. , }, For the number of clients, ~ These are the LoRA parameters for groups 1 to K; S2, the server sends LoRA parameters to the selected K clients respectively. This enables various clients Receive the corresponding LoRA parameters Keeping the weights of the pre-trained model unchanged, for the local dataset Perform fine-tuning training of a large language model based on LoRA to update LoRA parameters and obtain locally updated LoRA parameters, which are then submitted to the server. S3, the server performs adaptive cross-aggregation based on the locally updated LoRA parameters submitted by K clients, including: for any client Specifically targeting the computational clients of each layer of the large language model. any of the first The LoRA parameter of the layer, and the first digit of any client among the remaining K-1 clients. Calculate the similarity between LoRA parameters of each layer and find the client with the lowest similarity. , will client The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. Cross-aggregation of LoRA parameters in the same layer; S4, the server updates the model parameters of the large language model based on the LoRA parameters obtained from the K groups of cross-aggregation. The updated large language model is then tested to obtain the prediction accuracy. If the prediction accuracy does not meet the requirements, the LoRA parameters obtained from the K groups of client cross-aggregation are used as the new LoRA parameters. If the condition is met, proceed to step S2; otherwise, end and exit.

[0019] LoRA parameters in this embodiment any of the first Group LoRA parameters All models include L layers of LoRA parameters, and each layer's LoRA parameters include a reduced-dimensional projection matrix A and an increased-dimensional reconstruction matrix B. This embodiment systematically analyzes the application of LoRA in federated learning, focusing on the differences between the low-rank adaptation (LoRA) parameters of each layer (especially the reduced-dimensional projection matrix A and the increased-dimensional reconstruction matrix B), and the impact of aggregated weights on model performance. The analysis reveals a key phenomenon: LoRA parameters exhibit significant heterogeneity across different layers; even within the same layer, the reduced-dimensional projection matrix A and the increased-dimensional reconstruction matrix B show different update patterns. Therefore, an average similarity analysis is performed on the LoRA parameters of different layers. Figure 2These are the data analysis results under independent and identically distributed (IID) and non-independent and identically distributed (non-IID) conditions in this embodiment, where (a) is the data analysis result under the independent and identically distributed condition, and (b) is the data analysis result under the non-independent and identically distributed condition. LoRA-A and LoRA-B represent the dimension reduction projection matrix A and the dimension increase reconstruction matrix B, respectively. Figure 2 It can be seen that the similarity of LoRA parameters at different layers, i.e., the reduced-dimensional projection matrix A and the increased-dimensional reconstruction matrix B, is different. The similarity of the reduced-dimensional projection matrix A at the same layer is much higher than that of the increased-dimensional reconstruction matrix B. It is precisely because of the significant heterogeneity of LoRA parameters at different layers that even the reduced-dimensional projection matrix A and the increased-dimensional reconstruction matrix B at the same layer exhibit different update patterns. This embodiment employs two parts: a hierarchical aggregation model selection mechanism and a similarity-based adaptive aggregation rate adjustment strategy. The server initializes K LoRA parameters as follows: During the initialization and distribution of global LoRA parameters, including at the beginning of each round, the server broadcasts the current global LoRA parameters to the selected K clients. Each client uses the received LoRA parameter Initialize its local LoRA module. When each client performs local LoRA fine-tuning, the client... Keeping the weights of the pre-trained model unchanged, for its local dataset Perform large model fine-tuning training based on LoRA and update LoRA parameters. Including A and B. After each client completes its local LoRA fine-tuning, it needs to upload its local LoRA data. This includes uploading the updated LoRA parameters to each client after completing the local LoRA module training. Upload to server, update Finally, the layer-by-layer adaptive cross-aggregation involves determining the most suitable LoRA module for each layer based on the layer-by-layer similarity of the client's LoRA parameters using the CollaboratorSelector module. The aggregator employs a similarity-guided adaptive adjustment strategy to dynamically set the aggregation rate. Finally, adaptive weights are used to aggregate the LoRA parameters layer-by-layer on the selected pairs to update the model list for the next round.

[0020] In step S3 of this embodiment, the computational clients for each layer of the large language model are respectively targeted. any of the first The LoRA parameter of the layer, and the first digit of any client among the remaining K-1 clients. When considering the similarity between LoRA parameters of the layers, for the first layer... The LoRA parameter is the first Layer parameters include and The cosine similarity is calculated pairwise for clients i and j based on cosine similarity. Specifically, the function expression for calculating the similarity is: ; in, For the client Client The Similarity between LoRA parameters of layers For clients Client The The dimension reduction projection matrix A in the LoRA parameters of the layer, For clients Client The The dimension-upgrading reconstruction matrix B in the LoRA parameters of the layer.

[0021] In step S3, the client The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. Cross-aggregation of LoRA parameters at the same layer can be expressed as: ; in, For the client and client The Similarity factor of layer LoRA parameters This is the similarity calculation function.

[0022] In step S3 of this embodiment, the client... The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. The function expression for cross-aggregation of LoRA parameters at the same level is: ; ; in, and For clients After cross-aggregation, the first The dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the LoRA parameters of the layer. and For clients The local update The dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the LoRA parameters of the layer. and For clients The local update The weights of the dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the layer LoRA parameters.

[0023] The similarity-guided adaptive aggregation rate adjustment strategy calculates dynamic aggregation weights, based on two parts: one is the similarity factor. First, a secondary penalty is applied to the similarity difference to prevent aggressive fusion; second, a training progress factor is used. It increases conservatism over time to achieve stable convergence. In step S3 of this embodiment, the client... The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. Before cross-aggregating LoRA parameters at the same level, a dynamic calculation client is also included. The local update The weights of the dimension-reduced projection matrix A and the dimension-upgraded reconstruction matrix B in the layer LoRA parameters are: ; in, For the client The local update The weights of the dimension-reduced projection matrix A or the dimension-upgraded reconstruction matrix B in the layer LoRA parameters. To find the maximum value, This is the lower bound of the aggregation weight. and For aggregate weight pairs and training progress factor Sensitivity For the client and client Similarity factor of LoRA parameters in the same layer This represents the number of training epochs. Based on the formula above, the following can be calculated respectively. and Dynamic weights and In this embodiment, the client and client The functional expression for the similarity factor of the LoRA parameter in the same layer is: ; in, For the client and client The Similarity factor of layer LoRA parameters This is the similarity calculation function.

[0024] To verify the effectiveness of the federated fine-tuning method for large language models based on low-rank adaptation LoRA in this embodiment, the following experiments were conducted in this embodiment.

[0025] Experiment 1: In this experiment, the accuracy performance of the method in this embodiment (FedLoCA) and different existing methods (FedIT, FLoRA, FlexLoRA, FedSA and LORA-FAIR), datasets (20news group and ARC-Easy), and different degrees of data heterogeneity were tested on the Roberta and Llama models. The results are shown in Table 1.

[0026] Table 1: Accuracy Performance Test Table.

[0027]

[0028] As shown in Table 1, the method in this embodiment (FedLoCA) and existing methods (FedIT, FLoRA, FlexLoRA, FedSA and LORA-FAIR) and datasets (20news group dataset and ARC-Easy dataset) can effectively improve the performance and accuracy of Roberta and Llama models.

[0029] Experiment 2: This experiment tests the number of training epochs required for the method in this embodiment (FedLoCA) and existing methods (FedIT, FLoRA, FlexLoRA, FedSA, and LORA-FAIR) to achieve the same target accuracy under the same model and dataset. The results are shown in Table 2.

[0030] Table 2: Experimental results on the number of training epochs required to achieve the same target accuracy for the same model and dataset.

[0031] As shown in Table 2, the method of this embodiment (FedLoCA) can effectively reduce training costs compared with existing methods (FedIT, FLoRA, FlexLoRA, FedSA and LORA-FAIR).

[0032] Experiment 3: To verify the effectiveness of the method in this embodiment, an ablation experiment was designed to evaluate the performance with and without a layered collaborator selection mechanism (Layer) and an adaptive aggregation rate adjustment (RA) strategy, comparing them individually and in combination. The configurations considered are as follows: "w / o Layer+RA": FedLoCA removes both the Layer and RA strategies; "w / o RA": FedLoCA removes the RA strategy; "w / o Layer": FedLoCA removes the Layer mechanism. In this experiment, 300 rounds of experiments were conducted on the 20newsgroup dataset under non-independent identically distributed (non-IID) conditions (β=0.5). The final ablation results for the optimal accuracy are shown in Table 3.

[0033] Table 3: Ablation Experiment Results

[0034] As shown in Table 3, the method in this embodiment (FedLoCA) achieved the best accuracy (74.47%) in only 206 rounds, indicating that it is superior to other configurations in terms of both effectiveness and efficiency. The results prove that the Layered Collaborator Selection Mechanism and the Adaptive Aggregation Rate Adjustment Strategy (RA) are both necessary and complementary.

[0035] In summary, this embodiment implements a hierarchical dynamic aggregation mechanism for low-rank adaptation (LoRA) based on the large language model federated fine-tuning method. In each round of communication, the model is matched layer by layer based on parameter similarity. From each layer, the low-rank adaptation (LoRA) parameters of the client least similar to the current client are selected for aggregation, thereby encouraging parameter diversity and enhancing the generalization ability of the local model. Simultaneously, this embodiment introduces a dynamic weight generation mechanism for the large language model federated fine-tuning method based on low-rank adaptation (LoRA). The aggregation weights are adaptively adjusted according to the consistency of parameter update direction and the magnitude of local parameter changes, effectively mitigating the overfitting and bias problems caused by static weight strategies. This embodiment implements a hierarchical-aware federated fine-tuning method for low-rank adaptation (LoRA) in large language models, namely FedLoCA. This method mainly includes dynamically selecting collaborating clients for each layer based on the similarity of LoRA parameters between clients; and dynamically adjusting the aggregation weights of the LoRA parameters (dimensionality reduction projection matrix A and dimension increase reconstruction matrix B) according to parameter similarity and the number of training rounds, thereby achieving stronger personalization and generalization capabilities.

[0036] Furthermore, this embodiment also provides a large language model federated fine-tuning system based on low-rank adaptation LoRA, including interconnected microprocessors and memory, wherein the microprocessors are programmed or configured to execute the large language model federated fine-tuning method based on low-rank adaptation LoRA.

[0037] Furthermore, this embodiment also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the large language model federated fine-tuning method based on low-rank adaptation LoRA by a processor.

[0038] Furthermore, this embodiment also provides a computer program product, including a computer program or instructions that are programmed or configured to execute the large language model federated fine-tuning method based on low-rank adaptation LoRA via a processor.

[0039] Those skilled in the art will understand that the technical solutions provided by this invention may take the form of a method, system, or computer program product. Therefore, this invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce an implementation of the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0040] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A federated fine-tuning method for large language models based on low-rank adaptation LoRA, characterized in that, Includes the following steps: S1, The server initializes K sets of LoRA parameters for the L-layer large language model. , }, For the number of clients, ~ These are the LoRA parameters for groups 1 to K; S2, the server sends LoRA parameters to the selected K clients respectively. This enables various clients Receive the corresponding LoRA parameters Keeping the weights of the pre-trained model unchanged, for the local dataset Perform fine-tuning training of a large language model based on LoRA to update LoRA parameters and obtain locally updated LoRA parameters, which are then submitted to the server. S3, the server performs adaptive cross-aggregation based on the locally updated LoRA parameters submitted by K clients, including: for any client Specifically targeting the computational clients of each layer of the large language model. any of the first The LoRA parameter of the layer, and the first digit of any client among the remaining K-1 clients. Calculate the similarity between LoRA parameters of each layer and find the client with the lowest similarity. , will client The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. Cross-aggregation of LoRA parameters in the same layer; S4, the server updates the model parameters of the large language model based on the LoRA parameters obtained from the K groups of cross-aggregation. The large language model with updated model parameters is then tested to obtain the prediction accuracy. If the prediction accuracy does not meet the requirements, the LoRA parameters obtained from the K groups of client cross-aggregation are used as the new LoRA parameters. If the condition is met, proceed to step S2; otherwise, end and exit.

2. The federated fine-tuning method for large language models based on low-rank adaptation LoRA according to claim 1, characterized in that, The LoRA parameter any of the first Group LoRA parameters All include L layers of LoRA parameters, and each layer of LoRA parameters includes a dimension-reduced projection matrix A and a dimension-up reconstruction matrix B.

3. The federated fine-tuning method for large language models based on low-rank adaptation LoRA according to claim 1, characterized in that, In step S3, the computation clients for each layer of the large language model are calculated respectively. any of the first The LoRA parameter of the layer, and the first digit of any client among the remaining K-1 clients. When calculating the similarity between LoRA parameters of layers, the expression for the similarity calculation function is: ； in, For the client Client The Similarity between LoRA parameters of layers For clients Client The The dimension reduction projection matrix A in the LoRA parameters of the layer, For clients Client The The dimension-upgrading reconstruction matrix B in the LoRA parameters of the layer.

4. The federated fine-tuning method for large language models based on low-rank adaptation LoRA according to claim 1, characterized in that, In step S3, the client The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. The function expression for cross-aggregation of LoRA parameters at the same level is: ； ； in, and For clients After cross-aggregation, the first The dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the LoRA parameters of the layer. and For clients The local update The dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the LoRA parameters of the layer. and For clients The local update The weights of the dimension reduction projection matrix A and the dimension increase reconstruction matrix B in the layer LoRA parameters.

5. The federated fine-tuning method for large language models based on low-rank adaptation LoRA according to claim 1, characterized in that, In step S3, the client The LoRA parameters of each layer are respectively compared with the client with the lowest similarity. Before cross-aggregating LoRA parameters at the same level, a dynamic calculation client is also included. The local update The weights of the dimension-reduced projection matrix A and the dimension-upgraded reconstruction matrix B in the layer LoRA parameters are: ； in, For the client The local update The weights of the dimension-reduced projection matrix A or the dimension-upgraded reconstruction matrix B in the layer LoRA parameters. To find the maximum value, This is the lower bound of the aggregation weight. and For aggregate weight pairs and training progress factor Sensitivity For the client and client Similarity factor of LoRA parameters in the same layer This represents the number of training rounds.

6. The federated fine-tuning method for large language models based on low-rank adaptation LoRA according to claim 5, characterized in that, The client and client The functional expression for the similarity factor of the LoRA parameter in the same layer is: ； in, For the client and client The Similarity factor of layer LoRA parameters This is the similarity calculation function.

7. A federated fine-tuning system for large language models based on low-rank adaptation (LoRA), comprising interconnected microprocessors and memory, characterized in that, The microprocessor is programmed or configured to execute the large language model federated fine-tuning method based on low-rank adaptation LoRA as described in any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the large language model federated fine-tuning method based on low-rank adaptation LoRA as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the large language model federated fine-tuning method based on low-rank adaptation LoRA as described in any one of claims 1 to 6.

Images