Large model-oriented knowledge elimination learning parameter adjustment method and system

Abstract

The invention discloses a large model-oriented knowledge elimination learning parameter adjustment method and system. The large model-oriented knowledge elimination learning parameter adjustment method comprises the following steps of: obtaining an initial parameter and an elimination request of a large language model; calculating the influence of target data corresponding to the elimination request on the model parameters by using the influence function to obtain a parameter variation; and updating the parameters of the adapter module according to the parameter variation. According to the method, the influence function is introduced, and the influence of data disturbance on the model parameters is accurately estimated, so that the high calculation overhead caused by completely retraining the model in a traditional method is avoided, and the calculation efficiency is remarkably improved. Through an efficient parameter adjustment mechanism, under the condition that the model does not need to be trained again, accurate processing of various instance-level forgetting tasks is achieved. The robustness and accuracy of the model are improved, and the practical application potential of the model in a large-scale industrial scene is shown.

Description

Technical Field

[0001] The present application relates to the technical field of knowledge elimination learning, and specifically to a parameter adjustment method and system for knowledge elimination learning for large models. Background Art

[0002] With the rapid development of large-scale language models (LLMs), they have shown unprecedented potential in natural language understanding and generation tasks. Whether in traditional tasks such as text classification and machine translation, or in complex tasks such as dialogue generation and information extraction, LLMs have demonstrated powerful knowledge expression and reasoning capabilities. However, the powerful performance of the model also brings new problems, especially the risk that the model may remember sensitive, erroneous or unnecessary information during training. Therefore, how to efficiently remove this unnecessary information has become a key challenge in the current research field.

[0003] In order to deal with this problem, the prior art has proposed a variety of methods, which can be mainly summarized into the following two categories: Exact knowledge elimination method: Eliminate the influence of target data by subsetting the training data and retraining the model. This method can ensure that the model completely eliminates the knowledge of the target data, thus achieving the optimal effect in theory. However, its computational cost is extremely high, because each time a knowledge elimination request is processed, the model needs to be retrained, which involves a lot of time and resource investment. In addition, retraining may destroy the original performance or structural integrity of the model, especially when facing large-scale data sets or complex tasks, the efficiency problem is particularly prominent.

[0004] Approximate knowledge elimination methods: These methods achieve approximate knowledge elimination for specific data by adjusting model parameters, avoiding the need for retraining in exact methods. Common implementation methods include fine-tuning model parameters or constraining output differences (for example, using KL divergence). Although they are more efficient than exact methods, approximate knowledge elimination methods are usually only applicable to simple tasks (such as deleting training instances) and have difficulty handling more complex scenarios (such as modifying query content or correcting output responses). In addition, when dealing with changes in data distribution, the model performance of these methods may not be stable enough and may even introduce unexpected deviations.

[0005] Although existing methods have solved the knowledge elimination problem to some extent, they still have the following key limitations: Insufficient task support: Most current methods focus on deleting specific instances. However, in practical applications, more fine-grained operations are often required (such as modifying data content or correcting model outputs). The single nature of existing methods cannot meet these requirements.

[0006] High computational cost: The exact knowledge elimination method has a huge computational cost due to the frequent need to retrain the model, making it difficult to apply to application scenarios with high real-time requirements. Although the approximate knowledge elimination method has improved in efficiency, it still has performance bottlenecks when processing complex tasks or large-scale data sets.

[0007] Poor flexibility and scalability: Many methods rely on specific model architectures or fixed data partitioning strategies. This dependence makes them lack sufficient adaptability when faced with new tasks, new data, or dynamically changing requirements.

[0008] In general, existing methods are insufficient in adaptability, efficiency, and versatility in knowledge elimination tasks. There is an urgent need for a knowledge elimination technology that can efficiently handle multiple tasks and multiple scenarios while taking into account both computing performance and model effects. This technology can not only improve the accuracy of knowledge elimination, but also have a wide range of applicability to meet the diverse needs of practical applications. Summary of the invention

[0009] In this embodiment, a parameter adjustment method, system, electronic device and storage medium for knowledge elimination learning for large models are provided to address the shortcomings of knowledge elimination technology in terms of computational efficiency, task adaptability and model scalability. Traditional methods usually face the problems of high computational cost, single task support and difficulty in adapting to complex application scenarios.

[0010] In a first aspect, an embodiment of the present invention provides a parameter adjustment method for knowledge elimination learning of a large model, and the parameter adjustment method for knowledge elimination learning of a large model comprises: Get the initial parameters of the large language model and eliminate the request; Use the influence function to calculate and eliminate the influence of the target data corresponding to the request on the model parameters, and obtain the parameter change; The parameters of the adapter module are updated according to the parameter change amount.

[0011] In an optional embodiment, obtaining initial parameters of a large language model and eliminating a request includes: Obtain the initial parameters of the large language model through transfer learning or pre-training; A knowledge elimination request is received from a user or a system, the request specifying target data to be removed or corrected.

[0012] In an optional embodiment, the influence of the target data corresponding to the elimination request on the model parameters is calculated using the influence function to obtain the parameter change, including: Parsing knowledge elimination requests to determine specific task types; The gradient information related to the target data is extracted, and combined with the Hessian matrix of the model or its approximation, the influence function is used to estimate the exact impact of the target data on the model parameters to obtain the parameter change.

[0013] In an optional embodiment, based on the derivation of the influence function, the model parameter update is defined as follows: ; in, are model parameters, is the inverse of the Hessian matrix, representing the second-order derivative matrix of the model parameters, which is used to capture the complex interrelationships between the parameters; For the target instance The gradient information of represents the contribution of data to the model loss function. is the model input, is the model output.

[0014] In an optional embodiment, an optimization algorithm is used to transform the inverse Hessian vector product problem into a scalable convex quadratic optimization problem: ; in, is the cumulative contribution vector of the target gradient, q is the optimization target, that is, the parameter change, and H is the Hessian matrix.

[0015] In an optional embodiment, the knowledge elimination request includes a task type of instance removal, query modification, or response correction.

[0016] Compared with the prior art, the parameter adjustment method for knowledge elimination learning for large models of the present invention has the following beneficial effects: The parameter adjustment method for knowledge elimination learning for large models of the present invention shows significant advantages and positive effects in the instance-level forgetting task of large language models (LLMs). First, by introducing an influence function, this method accurately estimates the impact of data perturbations on model parameters, thereby avoiding the high computational overhead of completely retraining the model in traditional methods and significantly improving computational efficiency. Through an efficient parameter adjustment mechanism, accurate processing of a variety of instance-level forgetting tasks is achieved without retraining the model. This not only improves the robustness and accuracy of the model, but also demonstrates its practical application potential in large-scale industrial scenarios.

[0017] In a second aspect, an embodiment of the present invention provides a parameter adjustment system for knowledge elimination learning of a large model, including: Parameter acquisition module, used to obtain the initial parameters of the large language model and eliminate requests; An influence function calculation module is used to calculate and eliminate the influence of the target data corresponding to the request on the model parameters by using the influence function to obtain the parameter change amount; The parameter updating module is used to update the parameters of the adapter module according to the parameter change amount.

[0018] In an optional embodiment, the influence function calculation module includes a gradient extraction subunit and a parameter estimation subunit; The gradient extraction subunit is used to extract gradient information related to the target data; The parameter estimation subunit is used to estimate the variation of the parameters in combination with the Hessian matrix of the model or its approximate value.

[0019] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a processor, a communication interface, a memory and a bus, wherein the processor, the communication interface and the memory communicate with each other through the bus, and the processor can call logic instructions in the memory to execute the steps of the method provided in the first aspect.

[0020] In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the parameter adjustment method for knowledge elimination learning for large models as described in the first aspect.

[0021] Compared with the prior art, the beneficial effects of the parameter adjustment system, electronic device and storage medium for knowledge elimination learning for large models of the present invention are the same as the parameter adjustment method for knowledge elimination learning for large models described in the first aspect, so they will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 It is a flow chart of a parameter adjustment method for knowledge elimination learning for a large model in an embodiment of the present invention; Figure 2 It is a structural block diagram of a parameter adjustment system for knowledge elimination learning for a large model in an embodiment of the present invention; Figure 3 1 is a structural block diagram of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0025] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with general skills in the technical field to which this application belongs. The words "one", "a", "the", "these" and the like in this application do not indicate a quantitative limitation, and they may be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" may mean: A exists alone, A and B exist at the same time, and B exists alone. Generally, the character " / " indicates that the objects associated with each other are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0026] The present invention aims to propose a unified parameter efficient knowledge elimination method based on influence function to solve the shortcomings of existing knowledge elimination technology in terms of computational efficiency, task adaptability and model scalability. Traditional methods usually face the problems of high computational cost, single task support, and difficulty in adapting to complex application scenarios. The present invention avoids the tedious retraining process by directly adjusting the model parameters, which not only significantly improves the computational efficiency, but also expands the applicability of the method in various tasks. This method provides a new idea for efficiently handling the diverse knowledge elimination needs in large-scale language models (LLMs), taking into account the accuracy of knowledge elimination tasks and the maintenance of model performance.

[0027] Specific goals include the following aspects: Support for multiple knowledge elimination tasks: The present invention designs a unified knowledge elimination framework that can flexibly handle multiple task types including instance removal, input modification, and output correction. In practical applications, models often need to face different forms of data modification requirements. For example, deleting sensitive or invalid training instances, correcting noise information in erroneous inputs, or correcting erroneous labels in output results. The present invention provides a highly versatile and adaptable solution by abstracting these different tasks into parameter adjustment problems.

[0028] Improve computational efficiency: Traditional precise knowledge elimination methods require frequent retraining of the model, which has extremely high computational costs. Although approximate methods are more efficient, they also face efficiency issues when data complexity increases. The present invention converts the adjustment process of model parameters into an optimization problem by designing an efficient inverse Hessian vector product calculation method, and uses efficient algorithms such as stochastic gradient descent (SGD) to solve it. This design greatly reduces the time and resource overhead in the parameter update process, allowing the present invention to maintain excellent computing performance when processing large-scale data. In addition, the method supports incremental updates and can quickly adapt to the needs of new data or new tasks, providing the possibility for efficient deployment in practical applications.

[0029] Maintaining model performance: While achieving the knowledge elimination task, the present invention focuses on retaining the original performance of the model, ensuring that the model still has high predictive ability and stability after completing the knowledge elimination operation. By accurately estimating the impact of data perturbations on model parameters through influence functions, the present invention can minimize unnecessary performance losses. For example, when performing data removal tasks, only the affected parameter parts are adjusted without interfering with other unmodified data contributions; when correcting model outputs, the parameter update process can effectively avoid information loss or model confusion, ensuring that the model after knowledge elimination still maintains high-quality reasoning and generation capabilities.

[0030] Through the above improvements, the present invention not only achieves significant improvements in multi-task support, computing efficiency and model performance retention, but also has strong expansion capabilities and can flexibly adapt to various practical application scenarios. For example, user privacy data needs to be removed in the recommendation system, erroneous knowledge fragments need to be deleted in the generative task, or erroneous labels need to be corrected in the classification task. The present invention provides a unified and efficient solution, provides technical support for a variety of knowledge elimination tasks, and has broad application prospects and industrial value.

[0031] Specifically, an embodiment of the present invention provides a parameter adjustment method for large model knowledge elimination learning (LLMIF). Figure 1 is a flow chart of the parameter adjustment method for knowledge elimination learning of a large model of the present invention, such as Figure 1As shown, the process includes the following steps: S100, obtaining initial parameters of a large language model and eliminating a request; Specifically, obtaining the initial parameters of the large language model and the elimination request includes: Obtain the initial parameters of the large language model through transfer learning or pre-training; A knowledge elimination request is received from a user or a system, the request specifying target data to be removed or corrected.

[0032] It should be noted that in the framework initialization stage, the large-scale language model (LLM) usually obtains initial parameters through transfer learning or pre-training.

[0033] The present invention uses Parameter-Efficient Fine-Tuning (PEFT) technology, such as LoRA (Low-Rank Adaptation) or Adapter modules. This technology only adjusts a small number of model parameters and freezes the rest, thereby significantly reducing computing and storage requirements and providing a lightweight foundation for subsequent parameter updates.

[0034] S200, using the influence function to calculate the influence of the target data corresponding to the elimination request on the model parameters, and obtaining the parameter change; First of all, it should be noted that the knowledge elimination request includes the task types of instance removal, query modification or response correction. The user or the system submits a knowledge elimination request through a specific interface, specifying the target data to be removed or corrected. For example, the user may request to delete sensitive training instances, modify the query content, or correct erroneous labels in the generated output. The content of the knowledge elimination request will be parsed by the system into specific task categories (such as instance removal, query modification or response correction), and provide target data input for subsequent parameter adjustments.

[0035] The influence function is used to calculate and eliminate the influence of the target data corresponding to the request on the model parameters, and the parameter changes are obtained, including: Parsing knowledge elimination requests to determine specific task types; The gradient information related to the target data is extracted, and combined with the Hessian matrix of the model or its approximation, the influence function is used to estimate the exact impact of the target data on the model parameters to obtain the parameter change.

[0036] Furthermore, based on the derivation of the influence function, the model parameter update is defined as follows: ; in, are model parameters, is the inverse of the Hessian matrix, representing the second-order derivative matrix of the model parameters, which is used to capture the complex interrelationships between the parameters; For the target instance The gradient information of represents the contribution of data to the model loss function. is the model input, is the model output.

[0037] However, directly computing the inverse of the Hessian matrix This is not practical in large-scale language models.

[0038] Furthermore, an optimization algorithm is used to transform the inverse Hessian vector product problem into a scalable convex quadratic optimization problem: ; in, is the cumulative contribution vector of the target gradient, q is the optimization target, that is, the parameter change, and H is the Hessian matrix.

[0039] The present invention efficiently approximates the inverse Hessian vector product through an optimization algorithm, converts the computational problem into a convex quadratic optimization problem about q, and by solving the optimization problem, LLMIF can obtain accurate parameter adjustment direction and amplitude without significantly increasing the computational cost.

[0040] In order to cope with the complexity of actual calculations, the present invention adopts stochastic gradient descent and other batch optimization algorithms to achieve a good balance between computational efficiency and model performance. This method significantly reduces the computational complexity, making the parameter update process more advantageous in terms of time and space overhead, while ensuring support for a variety of knowledge elimination tasks.

[0041] Exemplarily, after the knowledge elimination request is parsed, the LLMIF framework efficiently calculates the impact of the target data on the model parameters through the influence function. Specifically, the framework first extracts the gradient information related to the target data, and estimates the change in the parameters in combination with the Hessian matrix (or its approximation) of the model. In order to cope with the computational complexity of large-scale models, the present invention designs an optimization algorithm based on stochastic gradient descent (SGD) to transform the Hessian-vector product problem into a scalable optimization process, thereby significantly reducing the computational cost. This step is the core link of the framework, ensuring the accuracy and efficiency of parameter updates.

[0042] S300: Update the parameters of the adapter module according to the parameter change amount.

[0043] According to the parameter changes that affect the function calculation, the present invention directly updates the parameters of the adapter module without retraining the entire model. The adjustment of the adapter parameters only acts on the local structure of the model, avoiding interference with the frozen parameters, thereby effectively retaining the original capabilities of the model. This method not only speeds up the update process, but also reduces possible performance degradation problems, allowing the model to quickly adapt to the needs of knowledge elimination requests.

[0044] After the parameter update is completed, the model is fully evaluated to check whether its performance meets the requirements. The evaluation includes the model's performance on the knowledge elimination task (such as whether the target data is effectively removed or modified) and the prediction accuracy of other tasks. Based on the evaluation results, the parameter adjustment strategy or hyperparameter configuration may be further optimized to ensure that the model maintains a high level of performance and stability after knowledge elimination.

[0045] The updated model will be deployed to actual applications to handle real-time tasks. For example, in a recommendation system, the model can immediately apply the latest parameter adjustment results to provide users with personalized recommendations; in a generative model, the revised model can output more accurate and reliable generated content. Through this step, LLMIF achieves rapid response to dynamic data needs and meets high real-time application scenarios in industry and academia.

[0046] For example, in a recommendation system, user behavior data (such as browsing history, click logs, and shopping history) are widely used to train recommendation models in personalized recommendation systems. However, these data may contain outdated or sensitive information. For example, when a user requests to delete his or her history, LLMIF can quickly remove the impact of the relevant data on the model parameters without retraining the entire recommendation model. At the same time, user behavior in the recommendation system changes dynamically, and LLMIF can adjust the model in real time to adapt to the latest user preferences, ensuring the accuracy and relevance of the recommendation results. Through this method, the recommendation system can not only better meet user privacy needs, but also maintain continuously optimized recommendation performance.

[0047] Privacy protection: In many applications involving personal data (such as medical diagnosis, financial analysis, and social media), protecting user privacy is a top priority. LLMIF is able to remove sensitive data related to user privacy in the model, thereby avoiding the risk of data leakage. For example, in a medical diagnosis model, a user may request to delete their personal medical record data. LLMIF can efficiently achieve this requirement through precise calculation and parameter updates, while ensuring that the performance of the model is not significantly affected after knowledge elimination. Compared with traditional methods of retraining models, LLMIF is faster and saves more computing resources, providing an efficient technical means for privacy protection.

[0048] Generative models: In text generation, image generation, and other generative tasks, models may generate results that contain incorrect information or inaccurate knowledge. For example, a dialogue generation model may cite outdated facts or generate biased responses. With LLMIF, generative models can correct these inaccurate generated content in a timely manner. For example, the model can remove the memory of a certain erroneous knowledge or correct the incorrect logic in the generated output, thereby improving the generation quality and user experience. This method is not only applicable to text generation tasks, but can also be extended to multimodal generation applications (such as image annotation and video summarization) to achieve cross-domain flexibility.

[0049] Legal Compliance and Data Management: In many industries, compliance with laws and regulations on data deletion or the “right to be intellectually erased” is an important challenge for companies. When users request to delete their personal data, LLMIF can respond quickly by directly updating model parameters to eliminate the impact of the target data on the model. Compared with traditional methods, this framework avoids the high cost of retraining the model while ensuring the dual needs of compliance and efficiency. In addition, LLMIF can also help companies optimize data management processes and reduce system downtime or waste of computing resources caused by data updates.

[0050] Real-time decision-making system: In real-time decision-making scenarios such as financial transactions and advertising, the system needs to quickly adjust the model based on the latest data. For example, the advertising platform needs to remove invalid click data, or dynamically optimize the delivery strategy based on newly arrived interaction records. LLMIF can quickly adapt to changes in real-time data through its efficient parameter update mechanism, avoiding decision errors caused by data lag, thereby improving the system's response speed and effectiveness.

[0051] From the above application examples, we can see that the LLMIF framework combines efficient computing power and flexible task adaptability in the knowledge elimination task of large-scale language models, significantly reducing computing costs while meeting diverse practical needs. Whether in the fields of personalized recommendation, privacy protection or generative tasks, LLMIF has demonstrated wide applicability and practical value, providing an advanced solution for processing dynamic changes and large-scale data. The advantages of this technology give it important promotion potential in both industrial and academic scenarios.

[0052] In summary, parameter adjustment is achieved through efficient computational methods to meet various knowledge elimination requirements. The main methods of the framework include the following parts: Complete knowledge elimination of samples. The training of large language models relies on large-scale data, which contains knowledge from different fields and scenarios. However, in practical applications, it is sometimes necessary to remove the knowledge of specific samples from the model to meet privacy protection, data copyright or personalization requirements. This process is called "complete knowledge elimination of samples". Unlike traditional model training updates, its purpose is to completely remove the influence of certain samples on the model behavior without affecting the overall performance. The technical challenges of achieving complete knowledge elimination of samples mainly include two aspects: 1) The complexity of knowledge representation of the model: knowledge in language models is often not simply stored in a certain parameter, but is collaboratively expressed through multiple parameters. Therefore, how to efficiently locate the parameters related to specific samples is one of the difficulties; 2) The balance of model performance: when eliminating specific knowledge, it is necessary to ensure that other parts of the model remain intact to avoid unnecessary damage to its general capabilities. At present, methods for this problem include knowledge elimination based on influence functions, gradient inverse optimization, and parameter pruning techniques. For example, the influence function can be used to analyze the role of specific samples in model parameters, and then the relevant parameters can be reversed or directly adjusted to achieve knowledge elimination. In addition, some studies have proposed to generate pseudo gradients for specific samples and reversely optimize model parameters to delete the knowledge related to the sample. Although these methods are effective, they still need to optimize efficiency when dealing with large-scale models, and further study their impact on model robustness and generalization ability. As regulations (such as GDPR and CCPA) continue to increase their requirements for users' right to delete data, "complete knowledge elimination of samples" will become a key issue that cannot be ignored in the development and application of language models.

[0053] Input and output optimization adjustment. The output effect of large language models depends heavily on the quality and structure of the input. In some tasks, the input data may contain noise or erroneous information, such as spelling errors, inaccurate descriptions, or irrelevant content in user queries. These problems may have a negative impact on the model's understanding and prediction capabilities. To address this challenge, LLMIF provides an efficient input data correction mechanism that can intelligently modify noise or errors in the input, such as removing irrelevant noise information, correcting spelling errors, or replacing erroneous content, thereby ensuring that the model can more accurately understand user intent. Specifically, LLMIF combines input data preprocessing with dynamic response adjustment to achieve consistency in output responses while modifying input data. The corrected input can not only optimize the model's contextual understanding, but also further dynamically adjust the model's parameters by influencing function calculations, so that the model can better adapt to the modified input data. This process ensures that the model's prediction results can reflect the true intent of the modified data without being disturbed by the original erroneous information. In addition, the LLMIF framework not only improves the model's response accuracy, but also enhances its robustness to complex or noisy environments. Whether there are spelling errors in user queries or inaccurate description information, LLMIF can quickly adapt to the changes of these imperfect data through precise input optimization and parameter adjustment mechanisms, providing a solid guarantee for the accuracy and reliability of task results. This capability makes LLMIF unique in dealing with scenarios where user input is diverse and contains erroneous or uncertain information, especially in natural language processing tasks. Although the text generated by large language models is usually highly fluent and consistent, its output results sometimes face problems such as being lengthy, inaccurate, or not meeting user needs. When the response labels of training samples are found to be incorrectly labeled, the LLMIF framework can quickly adjust the model parameters to make its output consistent with the correct labels. This process not only reduces the negative impact of incorrect labels on the model learning process, but also avoids the long-term damage of error propagation to predictive ability. Specifically, LLMIF introduces a dynamic parameter optimization mechanism. After discovering label errors, it can quickly update the model's gradient calculation and parameter adjustment methods to minimize the interference of erroneous samples. In addition, the framework can re-evaluate the model's performance on the task based on the corrected labels, thereby ensuring that the model can more accurately capture the true patterns in the data. Thanks to this efficient response correction mechanism, LLMIF significantly improves the performance of the model on related tasks, not only improving the accuracy of predictions, but also enhancing the robustness of the model to uncertain data. This rapid response and correction to incorrect labels enables LLMIF to use data more effectively during training, especially in scenarios with noisy labels or low-quality data annotations.

[0054] In order to efficiently calculate the changes in model parameters, this paper designs a new optimization algorithm that transforms the inverse Hessian vector product problem in the influence function into a convex quadratic optimization problem. Through this transformation, the framework can efficiently solve using modern optimization algorithms such as stochastic gradient descent, significantly reducing the computational complexity and improving the scalability of the computation.

[0055] The LLMIF method of the present invention shows significant advantages and positive effects in the instance-level forgetting task of large language models (LLMs). First, LLMIF introduces an influence function to accurately estimate the impact of data perturbations on model parameters, thereby avoiding the high computational overhead of completely retraining the model in traditional methods and significantly improving computational efficiency. In addition, compared with existing methods, LLMIF performs well in both model performance preservation and efficiency optimization.

[0056] Experimental results demonstrate that LLMIF outperforms existing state-of-the-art methods on multiple benchmark tasks and datasets.

[0057] Table 1: AUC prediction results of different methods on the book dataset

[0058] In the query modification and response correction tasks, LLMIF further demonstrated its robustness and resilience to noisy data. In the query modification task of the MovieLens dataset, by correcting the randomly injected noise, LLMIF achieved an average performance improvement of 5.1% over the contaminated model and was closest to the performance of the retraining method. In the response correction task, LLMIF achieved a 14% performance improvement over the contaminated model in the experiment of correcting 40% of the wrong labels, and the average performance gap with the retraining method was only 3%.

[0059] Table 2: Experimental results of different methods on the MovieLens dataset

[0060] In addition, in terms of efficiency evaluation, LLMIF demonstrated excellent computational efficiency. In the query modification task, LLMIF only took 1.37×10³ seconds, which is about 31 times faster than the retraining method, significantly reducing the computing resource requirements.

[0061] In summary, LLMIF achieves accurate processing of multiple instance-level forget tasks through an efficient parameter adjustment mechanism without retraining the model. This not only improves the robustness and accuracy of the model, but also demonstrates its practical application potential in large-scale industrial scenarios.

[0062] The embodiment of the present invention also provides a parameter adjustment system for knowledge elimination learning for large models, which is used to implement the above method embodiments, and will not be repeated hereafter. The terms "module", "unit", "subunit", etc. used below can implement a combination of software and / or hardware for predetermined functions. Although the systems described in the following embodiments are preferably implemented in software, the implementation of hardware or a combination of software and hardware is also possible and conceivable.

[0063] like Figure 2 As shown, Figure 2 : is a structural block diagram of a parameter adjustment system for knowledge elimination learning for a large model in the present invention, the system comprising: A parameter acquisition module 101 is used to acquire initial parameters of a large language model and eliminate requests; An influence function calculation module 102 is used to calculate the influence of the target data corresponding to the elimination request on the model parameters by using the influence function to obtain the parameter change amount; The parameter updating module 103 is used to update the parameters of the adapter module according to the parameter variation.

[0064] The influence function calculation module 102 includes a gradient extraction subunit and a parameter estimation subunit; The gradient extraction subunit is used to extract gradient information related to the target data; The parameter estimation subunit is used to estimate the variation of the parameters in combination with the Hessian matrix of the model or its approximate value.

[0065] Figure 3 A structural block diagram of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the following method: Get the initial parameters of the large language model and eliminate the request; Use the influence function to calculate and eliminate the influence of the target data corresponding to the request on the model parameters, and obtain the parameter change; The parameters of the adapter module are updated according to the parameter change amount.

[0066] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0067] An embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method provided in each of the above embodiments is implemented, for example, including: Get the initial parameters of the large language model and eliminate the request; Use the influence function to calculate and eliminate the influence of the target data corresponding to the request on the model parameters, and obtain the parameter change; The parameters of the adapter module are updated according to the parameter change amount.

[0068] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A parameter adjustment method for knowledge elimination learning of large models, characterized in that: The parameter adjustment method comprises: Get the initial parameters of the large language model and eliminate the request; Use the influence function to calculate and eliminate the influence of the target data corresponding to the request on the model parameters, and obtain the parameter change; The parameters of the adapter module are updated according to the parameter change amount.

2. The parameter adjustment method for knowledge elimination learning for large models according to claim 1 is characterized in that: Get the initial parameters of the large language model and eliminate the request, including: Obtain the initial parameters of the large language model through transfer learning or pre-training; A knowledge elimination request is received from a user or a system, the request specifying target data to be removed or corrected.

3. The parameter adjustment method for knowledge elimination learning for large models according to claim 1, characterized in that: The influence function is used to calculate and eliminate the influence of the target data corresponding to the request on the model parameters, and the parameter changes are obtained, including: Parsing knowledge elimination requests to determine specific task types; The gradient information related to the target data is extracted, and combined with the Hessian matrix of the model or its approximation, the influence function is used to estimate the exact impact of the target data on the model parameters to obtain the parameter change.

4. The parameter adjustment method for knowledge elimination learning for large models according to claim 3 is characterized in that: Based on the derivation of the influence function, the model parameter update is defined as follows: ； in, are model parameters, is the inverse of the Hessian matrix, representing the second-order derivative matrix of the model parameters, which is used to capture the complex interrelationships between the parameters; For the target instance The gradient information of represents the contribution of data to the model loss function. is the model input, is the model output.

5. The parameter adjustment method for knowledge elimination learning for large models according to claim 4 is characterized in that: An optimization algorithm is used to transform the inverse Hessian vector product problem into a scalable convex quadratic optimization problem: ； in, is the cumulative contribution vector of the target gradient, q is the optimization target, that is, the parameter change, and H is the Hessian matrix.

6. The parameter adjustment method for knowledge elimination learning for large models according to claim 2, characterized in that: The knowledge elimination request includes a task type of instance removal, query modification, or response revision.

7. A parameter adjustment system for knowledge elimination learning of large models, characterized in that: include: Parameter acquisition module, used to obtain the initial parameters of the large language model and eliminate requests; An influence function calculation module is used to calculate and eliminate the influence of the target data corresponding to the request on the model parameters by using the influence function to obtain the parameter change amount; The parameter updating module is used to update the parameters of the adapter module according to the parameter change amount.

8. The parameter adjustment system for knowledge elimination learning for large models according to claim 7, characterized in that: The influence function calculation module includes a gradient extraction subunit and a parameter estimation subunit; The gradient extraction subunit is used to extract gradient information related to the target data; The parameter estimation subunit is used to estimate the variation of the parameters in combination with the Hessian matrix of the model or its approximate value.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the parameter adjustment method for knowledge elimination learning for large models as described in any one of claims 1 to 6.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the parameter adjustment method for knowledge elimination learning for large models as described in any one of claims 1 to 6 are implemented.

Images