Federal fine-tuned large model evaluation report generation system, method, and storage medium

Abstract

The application discloses a federal fine-tuning large model evaluation report generation system and method and a storage medium, wherein the system comprises a cross-domain regulation event collection and semantic graph evolution unit, a federal multi-modal evaluation-generation unit, an adaptive model and knowledge synchronization unit, a parameter efficient fine-tuning unit and an explainability and robustness verification unit. The explainable federal fine-tuning large model evaluation report generation system provided by the application realizes cross-modal efficient evaluation in the field of intellectual property rights, ensures real-time performance, privacy security and explainability, and effectively improves the accuracy and reliability of the evaluation result.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a federated fine-tuning large model evaluation report generation system, method, and storage medium. Background Technology

[0002] With the rapid development of the knowledge economy, the quantity and complexity of intellectual property rights are growing exponentially. This has made traditional manual review and evaluation methods increasingly inadequate to meet practical needs, driving the rapid development of intelligent evaluation technologies. In recent years, large language models (LLMs) based on artificial intelligence technology have been increasingly widely used in the patent and intellectual property fields. They possess powerful semantic understanding and generation capabilities, and can, to some extent, replace manual tasks such as intellectual property document retrieval, text classification, and intelligent evaluation. This technology typically relies on deep learning and pre-trained models, trained in a self-supervised manner using large-scale text corpora, thus possessing strong generalization and understanding capabilities.

[0003] However, while existing large-scale language models excel in semantic understanding and generation tasks, they still have significant shortcomings in intellectual property assessment applications. First, the intellectual property field involves complex and diverse data sources, including legal provisions, technical cases, judicial precedents, full-text patents, images, and tables—multimodal information. Existing models primarily focus on text-based single-modal processing, making it difficult to efficiently integrate and analyze multi-source heterogeneous data, resulting in a one-sided and insufficient assessment process. Second, the laws, regulations, and market environment in the intellectual property field change frequently. Existing semantic understanding models often rely on static knowledge graphs as background support, lacking dynamic real-time update capabilities and failing to quickly respond to regulatory changes and market dynamics. This lag reduces the timeliness and reliability of assessment conclusions and may lead to the risk of misjudgment.

[0004] To address these issues, some studies have attempted to achieve cross-institutional data collaboration through federated learning techniques and protect data privacy with secure aggregation protocols, achieving initial attempts at joint training of multi-party models. However, they have not yet effectively integrated cross-modal data, and most lack sufficient attention to model interpretability and evaluation credibility. Furthermore, regarding model fine-tuning, current conventional methods for fine-tuning large language models involve a large number of parameters and high computational costs, making them difficult for resource-constrained enterprises to deploy effectively and to meet the need for rapid updates at the minute level. Simultaneously, the traceability and immutability of model updates during fine-tuning lack effective guarantees, failing to meet the practical application requirements of high credibility and regulatory oversight in intellectual property assessment.

[0005] Recent research has shown that Generative Adversarial Networks (GANs) and gradient attribution techniques are gaining increasing attention for improving the credibility and interpretability of model outputs. These methods analyze the relationship between model outputs and inputs through gradient attribution analysis, or construct adversarial examples to test model stability. However, current technologies typically apply these methods independently, lacking organic integration, and have not yet formed a systematic solution for intellectual property assessment scenarios. Furthermore, existing solutions lack timely and effective automatic correction mechanisms to address insufficient model evaluation credibility, making it difficult to ensure the long-term stability and reliability of the evaluation model's output.

[0006] Therefore, how to provide a system, method, and storage medium for generating evaluation reports for federally tuned large models is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the aforementioned technical problems, the purpose of this invention is to overcome the shortcomings of existing technologies, which typically apply these methods independently, lacking organic integration and failing to form a systematic solution for intellectual property assessment scenarios. Furthermore, existing solutions lack timely and effective automatic correction mechanisms to address insufficient model assessment reliability, making it difficult to ensure the long-term stability and reliability of the assessment model's output results. To achieve the above objectives, the present invention provides a federated fine-tuning large model evaluation report generation system, the system comprising: The cross-domain regulatory event acquisition and semantic graph evolution unit is used to acquire regulatory events in global patent office announcements, judicial rulings and technology news in real time. Event nodes are incorporated into the intellectual property semantic graph through incremental graph convolution-difference algorithm, and causal reasoning and confidence labeling of the five-tuple of "article-term-case-precedent-market indicator" are completed based on multi-hop graph attention. The Federal Multimodal Assessment-Generation Unit is used to encode text, images, tables, and full-text patents locally at each data holder. It employs a gated expert hybrid model and a secure aggregation protocol to complete privacy-preserving joint training, and generates a structured assessment report containing verifiable evidence chains in parallel at the central side based on a cross-modal alignment matrix. An adaptive model and knowledge synchronization unit is used to map four-dimensional feedback signals of law, technology, market and competitive situation into a multi-agent reinforcement learning strategy, and dynamically adjust the gating expert weights, dual adapter parameters and semantic graph confidence intervals to achieve minute-level online updates; The parameter-efficient fine-tuning unit is used to quickly fine-tune different domain models by using parallel stacked LoRA-IA3 adapters and meta-learning initialization strategies, without exceeding one percent of the total number of trainable parameters of the original large language model. The hash values ​​of key fine-tuning parameters are written to the blockchain through zero-knowledge proof to ensure the traceability and immutability of model updates. The interpretability and robustness verification unit is used to generate cross-modal significance heatmaps and causal chain descriptions based on dual gradient attribution for each evaluation conclusion, and to perform adversarial sample robustness tests using generative adversarial networks. When the confidence of the conclusion is lower than a preset threshold, it triggers an automatic rollback to the previous trusted version.

[0008] Preferably, the cross-domain regulatory event collection and semantic graph evolution unit performs the following steps: Establish a multi-source crawler task queue Polling to retrieve global patent office announcements data sources Data source of judicial judgments Technology News Data Source Forming an event text stream ; For text streams Word segmentation, named entity recognition, and relation extraction are performed to obtain a set of event nodes. With edge set ; The increment vector is calculated using the following formula; in For activation function, For the normalized adjacency matrix, The matrix represents the nodes at the previous time step. This is the weight matrix; Will Injection Semantic Graph get A multi-hop graph attention mechanism is used to calculate the causal score for quintuple nodes. And generate confidence labels. Write .

[0009] Preferably, the federal multimodal evaluation-generation unit performs the following steps: At each data holder Establish a multimodal dataset locally And embedded via a shared encoder-decoder ; Expert weights are calculated using a gated expert mixture model. And generate expert output ; Using secure aggregation protocol Homomorphic encryption summation yields the global gradient Update the model after decryption; Based on the cross-modal alignment matrix Parallel generation of evidence chains The structured assessment report.

[0010] Preferably, the adaptive model and knowledge synchronization unit are used to perform the following steps: Constructing a four-dimensional feedback vector Input Multi-Agent Policy Network Generate gated route weights With adapter gain and according to reports Calculate policy gradient according to Update strategy parameters, press Adjust the expert routing matrix and adapter weights; Correcting the confidence interval of the semantic graph The update log will then be written to the blockchain traceability channel.

[0011] Preferably, the parameter fine-tuning unit is used to perform the following steps: Domain dataset Freeze the original weights Insert the LoRA-IA3 adapter to form an incremental connection. And constrain the proportion of trainable parameters Initialize parameters with meta-learning Obtain domain-specific parameters as the starting point Calculate the hash of the fine-tuning parameters Generating zero-knowledge proofs And write it along with the timestamp into the blockchain ledger. .

[0012] Preferably, the interpretability and robustness verification unit is used to perform the following steps: For input Calculation and prediction And generate a dual gradient heatmap based on Form a significant area , will support samples With hash Forming a causal chain ; By generating adversarial networks Constructing adversarial examples Calculate robustness score When confidence level or At that time, roll back to the previous trusted version. And rollback hash Upload to the blockchain.

[0013] The present invention also provides a method for generating an evaluation report, the method comprising: Collect cross-domain legal events and update the intellectual property semantic graph, and complete causal reasoning and confidence labeling; Multimodal encoding of text, images, tables, and full-text patents is performed locally at each data holder, and federated training is completed through secure aggregation. A structured evaluation report and a verifiable chain of evidence are generated based on the cross-modal alignment matrix; Utilizing four-dimensional feedback to drive multi-agent reinforcement learning, expert routes, adapter parameters, and semantic graph confidence intervals are synchronized within minutes. LoRA-IA3 fine-tuning is performed with parameters accounting for no more than one percent of the original model, and hashes and zero-knowledge proofs are written into the blockchain. Perform dual gradient interpretability analysis and adversarial robustness verification on the evaluation results, and automatically roll back the model version if necessary.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the evaluation report generation method of claim 7.

[0015] According to the above technical solution, the beneficial effects of the federated fine-tuning large model evaluation report generation system provided by the present invention in use are as follows: (1) This invention collects cross-domain regulatory events in real time and uses incremental graph convolution-difference algorithm and multi-hop graph attention mechanism to dynamically update semantic graph and causal reasoning, thereby realizing the real-time update and dynamic reasoning function of intellectual property evaluation system, effectively improving the timeliness of evaluation information update and the accuracy of semantic reasoning, and enhancing the real-time performance and decision reliability of intellectual property evaluation.

[0016] (2) This invention uses a federated multimodal evaluation-generation unit, a gated expert hybrid (MoE) model and a secure aggregation protocol to achieve joint training with cross-agency privacy protection, and generates a structured evaluation report with verifiable evidence chains on the central side based on the cross-modal dynamic alignment matrix. This significantly improves the fusion analysis capability of multi-source heterogeneous data and the interpretability of evaluation results, and demonstrates better data security and model credibility in the application scenario of multi-agency collaborative evaluation of intellectual property.

[0017] (3) In terms of adaptive model update and parameter fine-tuning, this invention effectively solves the shortcomings of existing technology models with large number of parameters and difficulty in fast update by using a multi-agent reinforcement learning strategy driven by four-dimensional feedback signals and a high-efficiency fine-tuning mechanism of parallel stacking of LoRA-IA3 adapters. It breaks through the limitations of existing technology models in terms of dynamic update response speed and number of fine-tuned parameters, and realizes minute-level online adaptive update and extremely low parameter fine-tuning of the model, thereby effectively improving the flexibility, adaptability and deployment economy of the model in the field of intellectual property intelligent evaluation.

[0018] (4) This invention achieves high credibility and stability of model output results by using dual gradient attribution analysis and adversarial robustness testing mechanism of generative adversarial network. When the credibility of the evaluated model output is lower than the threshold, it automatically rolls back to the previous credible version, effectively reducing the risk of model misjudgment, breaking through the technical bottleneck of insufficient credibility of existing intellectual property intelligent evaluation system, and thus significantly improving the reliability and credibility of intellectual property evaluation and report generation in practical applications.

[0019] Other features and advantages of the present invention will be described in detail in the following detailed description section; and all parts not covered in the present invention are the same as or can be implemented using the prior art. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural block diagram of a federated fine-tuning large model evaluation report generation system provided in a preferred embodiment of the present invention; Figure 2 A flowchart of a method for generating evaluation reports for a federated fine-tuning large model provided in a preferred embodiment of the present invention; Figure 3 A schematic diagram of the structure of an electronic device provided in a preferred embodiment of the present invention.

[0021] Explanation of reference numerals in the attached figures 1. Cross-domain regulatory event collection and semantic graph evolution unit 2. Federated Multimodal Evaluation - Generation Unit 3. Adaptive Model and Knowledge Synchronization Unit 4. High-efficiency parameter fine-tuning unit 5. Interpretability and Robustness Verification Unit Detailed Implementation The cross-domain regulatory event collection and semantic graph evolution unit of this invention employs a distributed crawler program to continuously poll and crawl data from global patent office announcements, judicial rulings, and technology news sources at a 2-minute interval. A text cleaning module removes irrelevant characters and formatting information, resulting in structured event nodes and edge relationship data. This data is then input into an incremental graph convolution-differential update algorithm, where the normalized adjacency matrix of the incremental graph convolution... Defined as ,in It is an adjacency matrix. It is the identity matrix. Let be the degree matrix, and let be the activation function. Specifically, the ReLU function is used; the multi-hop graph attention mechanism uses three attention heads, as shown by the formula... Calculate the causal scores between nodes, with the embedding vector of each node set to 128 dimensions, and add confidence scores. The value range is [0.85, 1.0]; the federated multimodal evaluation-generation unit adopts a gated expert hybrid (MoE) architecture, consisting of 4 sub-expert networks, with each sub-network having a uniform output dimension of 256 dimensions; the secure aggregation protocol specifically adopts the federated averaging algorithm FedAvg, which homomorphically encrypts the expert outputs from multiple parties, sums them, and decrypts them at the center to obtain the aggregation model parameters; the cross-modal dynamic alignment matrix uses cosine similarity calculation; the multi-agent reinforcement learning of the adaptive model and knowledge synchronization unit adopts the PPO algorithm, with a learning rate set. Route update step size Adapter update step size Semantic graph confidence interval adjustment coefficient The reinforcement learning feedback signal is updated every minute; the parameter-efficient fine-tuning unit uses the LoRA-IA3 adapter with an adapter rank r set to 8, and the total number of training parameters is limited to 1% of the original large model parameters. Meta-learning initialization uses the MAML algorithm, and zero-knowledge proof uses the Groth16 protocol to write the SHA256 hash values ​​of key parameters into the blockchain; the interpretability and robustness verification unit uses the dual gradient attribution analysis baseline input. Assuming all-zero vectors, the Generative Adversarial Network (GAN) uses the FGSM algorithm to generate adversarial examples, and the robustness score is... The threshold is set to 0.95. When the value falls below this threshold, a rollback mechanism is triggered, and the parameter hash value of the previous trusted version is recorded again in the blockchain ledger.

[0022] This invention enhances the timeliness of semantic graphs through real-time cross-domain regulatory data acquisition and dynamic graph update algorithms; it achieves cross-institutional multimodal data collaboration by utilizing federated learning and secure aggregation protocols, effectively ensuring data privacy; it reduces model update latency and significantly decreases parameter update volume through multi-agent reinforcement learning and efficient fine-tuning techniques; and it integrates dual gradient attribution analysis and adversarial robustness verification, making model output more credible and reliable, and assessment results traceable and highly credible. Compared with existing technologies, this invention significantly improves the response efficiency, privacy protection capabilities, and decision reliability of intellectual property assessment systems.

[0023] The cross-domain regulatory event collection and semantic graph evolution unit of this invention executes the following process sequentially: First, a distributed crawler task queue Q is established, and a multi-threaded polling method is used to collect global patent office announcement data every 2 minutes. Data source of judicial judgments Technology News Data Source The text stream T is generated by crawling the data separately. Then, the text stream T is segmented using the Jieba word segmentation tool, and named entity recognition and entity relationship extraction are achieved through the BERT-BiLSTM-CRF joint model to generate a clear set of event nodes. Sum of edges Then, an incremental graph convolution-difference algorithm is used, where the update vector... Through formula Calculation, where Using the ReLU activation function, This represents the normalized form of the adjacency matrix A. It is the previous time step graph node representation matrix, with a dimension of 128, and W is the trainable parameter matrix, with a dimension of 128×128; then , and update vector Write the intellectual property semantic graph using the graph update function. The updated semantic graph is obtained. Next, a multi-hop graph attention mechanism is adopted, where the number of attention heads is explicitly set to 3, and the query node represents... and target node representation All are 128-dimensional embedding vectors, obtained through the formula... Calculate causal scores Finally, based on the scores Generate confidence labels The confidence level labeling threshold is specifically set between 0.85 and 1.0, and the labeling results are... Write to the updated semantic graph middle.

[0024] This invention, by specifying concrete data acquisition methods, incremental update mechanisms, and implementation details of graph attention, enables the legal semantic graph to have higher real-time performance and dynamic update capabilities, improves the accuracy and timeliness of intellectual property assessment in processing complex legal data, and effectively overcomes the technical bottleneck of information lag in existing static graph schemes.

[0025] The specific implementation process of the federated multimodal evaluation-generation unit of this invention is as follows: First, for the data holder... For locally held text data Image data Table data and full-text patent data BERT, ResNet-50, and TabNet were used as encoders for feature encoding, respectively, to uniformly generate embedding vector representations with a dimension of 256. Secondly, a gated expert hybrid model (MoE) is used, with four expert subnetworks, each with a fixed output dimension of 256, and the output dimension is determined by the formula. Calculate the expert weight vector, where the weight matrix is... The dimension is 256×256, using weights The expert output vector is obtained from the output of the scheduling expert network. Then, using the secure aggregation protocol in federated learning, specifically the FedAvg algorithm, for joint model training, the expert output vectors of each participant are processed. Perform Paillier homomorphic encryption and then homomorphic decryption at the central side to obtain the global aggregated gradient vector. and gradient After decryption, update the global model parameters; then, use the updated embedding vectors to calculate the cross-modal dynamic alignment matrix. The matrix is ​​specifically determined using the cosine similarity calculation method, and the embedding vectors for different modalities are denoted as follows: , The vector dimension is fixed at 256; finally, based on the matrix... Parallel generation of structured evaluation reports, with each evaluation conclusion in the report corresponding to a chain of evidence. Supported by specific samples and its corresponding SHA256 hash value The composition is designed to ensure that conclusions are verifiable and reports are credible.

[0026] This invention achieves efficient cross-modal information fusion under multi-agency data collaboration by defining specific federated learning models, secure aggregation protocols, and encoding and alignment mechanisms for cross-modal data. This significantly enhances the comprehensiveness, data security, and objectivity of intellectual property assessments, thereby effectively overcoming the single-modal processing and data silo problems of existing technologies.

[0027] The specific implementation process of the adaptive model and knowledge synchronization unit of this invention is as follows: First, the legal dimension score is calculated. Technical dimension score Market dimension score Score based on competitive situation Clearly combine them into a feedback vector with a dimension of 4. The vector element values ​​are set to a range of [0,1]; secondly, the feedback vector is... Input Multi-Agent Reinforcement Learning Policy Network The network is specifically implemented using the PPO algorithm, with a learning rate of... The value is explicitly set to 0.0003, updated every minute, and the expert routing weight vector is output through the policy network. and dual adapter gain coefficient vector The vector dimension is set to 256, and a strategy action is selected. Then, based on the real-time assessed return value. (Values ​​range [-1, 1]) Calculate the policy gradient The calculation formula is: And based on this gradient with the learning rate Update the strategy parameters, i.e. Secondly, utilize expert routing weight vectors. Update step size by route Update the global routing matrix of the gated expert hybrid model The specific update method is as follows: At the same time, using the gain coefficient vector Update step size with adapter Adjust the parameters of the parallel-stacked LoRA-IA3 adapter using the following formula: Next, the confidence intervals of the intellectual property semantic graph are determined. Perform synchronous correction, and the correction formula is specified as follows: The confidence adjustment coefficient Explicitly set it to 0.1; finally, include the policy parameters in the updated log data. Routing weight Adapter parameter increment and confidence interval correction value The SHA256 hash value is used as log data and is explicitly recorded to the blockchain ledger once every minute to ensure that the update process is traceable and tamper-proof.

[0028] This invention clarifies the implementation details of specific multi-agent reinforcement learning algorithms, feedback vector dimensions, update step sizes, and blockchain traceability technology, thereby achieving automation, efficiency, and high reliability in synchronizing parameters and knowledge in intellectual property evaluation systems. It effectively solves the problems of model update lag and frequent manual intervention in existing technologies, and significantly improves the online response speed, real-time adaptability, and reliability of the model.

[0029] The specific implementation process of the parameter high-efficiency fine-tuning unit of this invention is as follows: First, clearly select the domain dataset set. The data volume of each domain dataset was set between 1000 and 5000 samples; secondly, for each domain dataset... In freezing the original large language model parameter matrix Under the premise of explicitly setting the rank, insert LoRA adapter matrix With IA3 scalar vector , where the vector dimension Set to 512 to form the adapter incremental parameter matrix. Then, statistical constraints are explicitly applied to the adapter parameters for all domains to ensure that the proportion of trainable parameters is [percentage missing]. Strictly not exceeding 1% of the original model's parameter count, i.e., satisfying the formula Next, the meta-learning algorithm is used, specifically the MAML algorithm for initialization. The meta-learning initialization parameters are... The determination was made after performing three steps of gradient descent on multiple domain datasets; subsequently, the learning rate was used. Domain-based loss function Domain-specific parameters The update process specifically satisfies the formula. Next, the domain-specific parameters and adapter increments are aggregated to obtain the final parameter set after fine-tuning. Calculate its SHA256 hash value Finally, using the Groth16 zero-knowledge proof protocol, a zero-knowledge proof is generated. This zero-knowledge proof and parameter hash value The corresponding timestamp is recorded every minute and explicitly written into the blockchain ledger to ensure strict traceability and immutability of the parameter update process.

[0030] This invention, by explicitly defining adapter parameter constraints, meta-learning initialization methods, and specific protocols for zero-knowledge proofs, effectively achieves extremely low computational resource consumption, efficient domain adaptation, and secure traceability during the fine-tuning of large model parameters. It significantly reduces fine-tuning costs and complexity, solves the problems of excessive fine-tuning overhead and difficulty in guaranteeing the reliability of parameter updates in existing technologies, and improves the practicality and reliability of model fine-tuning schemes in the intellectual property field. The specific implementation process of the interpretability and robustness verification unit of this invention is as follows: First, for each evaluation conclusion, explicitly call the pre-trained prediction function. Generate initial prediction output Input data vector The dimension is set to 512; secondly, based on the dual gradient attribution analysis method, the baseline input vector is explicitly selected. Calculate the significance heatmap for a vector consisting entirely of zeros. Its calculation method is as follows and based on Extracting the set of salient regions A significance threshold of 0.7 was selected; then, based on the set of significant regions... Clearly extract supporting samples And calculate the SHA256 hash value of the sample. The two clearly form a causal chain. Next, the Fast Gradient Sign Method (FGSM) from Generative Adversarial Networks (GANs) is explicitly adopted to perturb the amplitude parameter. Generate adversarial examples Then, adversarial examples were used. Calculate robustness score The specific calculation formula is as follows: The number of samples The number of results is explicitly set to 1000; finally, the confidence level corresponding to the causal score is labeled. Below 0.85, or robustness score When the value drops below 0.95, the automatic rollback mechanism is explicitly triggered, restoring the model to the parameters of the previous trusted version. And calculate the SHA256 hash value of the rollback event. The data, along with a timestamp, is explicitly written into the blockchain ledger every minute to ensure the stability, traceability, and immutability of the model evaluation results.

[0031] This invention, through a clear dual gradient attribution method and a specific implementation scheme of generative adversarial networks, ensures the high interpretability and robustness against disturbances of the model prediction results, significantly reduces the risk of model misjudgment, effectively solves the problems of insufficient credibility of existing models and difficulty in backtracking and correcting abnormal results, and improves the reliability and credibility of the intellectual property intelligent evaluation model in real complex application environments.

[0032] like Figure 2 As shown, the present invention also provides a method for generating an evaluation report, the method comprising: Step S1: Collect cross-domain legal events and update the intellectual property semantic graph to complete causal reasoning and confidence labeling; Step S2: Multimodal encoding of text, images, tables and full-text patents is performed locally on each data holder's site, and federated training is completed through secure aggregation. Step S3: Generate a structured evaluation report and a verifiable chain of evidence based on the cross-modal alignment matrix; Step S4: Utilize four-dimensional feedback to drive multi-agent reinforcement learning, and synchronize expert routes, adapter parameters, and semantic graph confidence intervals within minutes. Step S5: Perform LoRA-IA3 fine-tuning with parameters not exceeding one percent of the original model, and write the hash and zero-knowledge proof into the blockchain. Step S6: Perform dual gradient interpretability analysis and adversarial robustness verification on the evaluation results, and automatically roll back the model version if necessary.

[0033] In the above scheme, the specific implementation process of the intellectual property assessment and credible report generation method of the present invention is as follows: First, through a distributed crawler task queue Q, global patent office announcement data sources, judicial judgment data sources, and technology news data sources are continuously polled and crawled at a cycle of 2 minutes to form structured event node and edge relationship data. Then, the semantic graph is updated through an incremental graph convolution-differential update algorithm. The graph node embedding dimension is 128 dimensions. A multi-hop graph attention mechanism with 3 attention heads is used to calculate the causal score and label the confidence level. Second, BERT, ResNet-50, and TabNet are used to encode text, image, and tabular data respectively, and a feature embedding vector with a dimension of 256 is uniformly generated. Multimodal data federated learning is completed through a gated expert hybrid (MoE) model containing 4 expert sub-networks and the FedAvg secure aggregation protocol. Cosine similarity is used to calculate the cross-modal dynamic alignment matrix. Based on this matrix, a structured assessment report containing a verifiable evidence chain with supporting sample SHA256 hashes is generated in parallel. Third, a four-dimensional assessment report is constructed based on legal, technical, market, and competitive landscape data. The feedback vector, composed of multiple dimensions, utilizes the PPO algorithm for multi-agent reinforcement learning with a learning rate set to 0.0003. Policy updates are performed every minute, with route update step sizes set to 0.01 and adapter update step sizes set to 0.005. Expert route weights and adapter parameters are adjusted in real-time, and the semantic graph confidence interval is dynamically adjusted. Update logs are written to the blockchain ledger every minute. Then, efficient fine-tuning is performed using a LoRA-IA3 adapter with an explicitly defined rank of 8, with a limit not exceeding 1% of the original model's total parameters. Meta-learning initialization parameters are obtained through... The MAML algorithm is used to obtain the SHA256 hash value of the fine-tuned parameters, which is then processed by the Groth16 zero-knowledge proof protocol and recorded to the blockchain every minute. Finally, for each evaluation conclusion, a significance heatmap is generated using the dual gradient attribution method, with a significance threshold set to 0.7. The generative adversarial network uses the FGSM method with a perturbation amplitude of 0.01 to generate adversarial examples and calculates the robustness score. When the confidence level is lower than 0.85 or the robustness is lower than 0.95, it automatically rolls back to the previous trusted version, and the rollback parameter hash is recorded to the blockchain every minute. This invention also discloses a computer-readable storage medium, specifically a 128GB solid-state drive, in which a computer program implementing the above method is stored. The program is explicitly written in Python and can be executed in a Linux system environment equipped with an Intel Xeon processor and 64GB of memory.

[0034] like Figure 3 As shown, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an evaluation report generation method.

[0035] The electronic device includes a processor 901, a memory 902, a communication interface 903, and a bus 904; wherein the processor 901, the memory 902, and the communication interface 903 communicate with each other through the bus 904; the communication interface 903 is used for information transmission between the electronic device and the communication device of the terminal; and the processor 901 is used to call program instructions in the memory 902.

[0036] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0037] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0038] Example 1 To verify the feasibility of this invention, the interpretable federated fine-tuning large-scale model intellectual property assessment and credible report generation system of this invention was applied to an undergraduate university in Wuhu City, Anhui Province (hereinafter referred to as University C). As a key local university, University C has achieved significant research results in recent years in fields such as intelligent manufacturing, new energy vehicles, and electronic information, with approximately 1200 patent applications annually. However, with the increase in the number of patent applications, the university's intellectual property management department faces prominent problems such as poor assessment timeliness, scattered and difficult data integration, slow model parameter adjustment, and a lack of credibility assurance in assessment results. Traditional manual assessment methods can no longer meet the rapidly growing patent assessment needs of universities. To effectively solve the above problems, University C decided to deploy the system of this invention to improve the efficiency of intellectual property assessment and management and enhance the quality of technology transfer.

[0039] The system of this invention was deployed in the Information Center of University C, located in Jinghu District, Wuhu City, Anhui Province. This center is equipped with a high-performance server cluster, including computing nodes with Intel Xeon processors, 256GB of memory, and Tesla V100 GPUs, meeting the system's requirements for real-time data processing and complex computational tasks. The system first employs a distributed crawler task queue, polling every 2 minutes to retrieve announcements from the State Intellectual Property Office, the Supreme People's Court's case law database, and technical news from CNKI, VIP, and major global technology media websites. It collects patent law changes and technological developments in real time, and after cleaning and processing, forms structured node data. Each data update takes only about 85 seconds to inject the node data into the intellectual property semantic graph. The graph update uses the incremental graph convolution-difference algorithm of this invention, with a node embedding dimension of 128. The graph attention mechanism uses three attention heads to achieve efficient causal reasoning and dynamic confidence labeling of "article-terminology-case-judgment-market indicators".

[0040] University C comprises multiple colleges and research teams, each possessing abundant but scattered research data, including research papers, patent texts, engineering drawings, design images, and experimental data tables. This data is stored in a dispersed manner and exhibits significant cross-modal characteristics, making traditional aggregation and analysis methods inefficient and posing high data security risks. This invention addresses this issue by establishing federated learning nodes in the School of Mechanical Engineering, the School of Electrical Engineering, and the School of Information Science and Technology through a federated multimodal evaluation-generation unit. Employing a gated expert hybrid (MoE) model, text data is uniformly encoded into 256-dimensional feature vectors using a BERT model, image data using a ResNet-50 model, and tabular data using a TabNet model. A federated averaging algorithm (FedAvg) is then used for secure aggregation via homomorphic encryption, and the data is jointly trained at the central node. This enables secure analysis and fusion of cross-departmental data without leaving the campus. During a three-month trial run, the system has securely processed approximately 3,600 cross-modal data items and generated over 1,200 intellectual property assessment reports.

[0041] Furthermore, to address the issues of slow parameter adjustment and inability to quickly adapt to dynamic changes in scientific research using traditional evaluation models, this invention deploys an adaptive model and knowledge synchronization unit. It employs the PPO algorithm based on multi-agent reinforcement learning, using feedback signals from legal, technological, market, and competitive landscape aspects as input to dynamically adjust expert routing weights, adapter parameters, and the confidence interval of the semantic graph. The model parameters are automatically updated and adjusted every minute. Implementation results show that this invention reduces the parameter adjustment delay from several hours using manual methods to an average of less than 40 seconds, effectively improving the model's real-time response capability.

[0042] Meanwhile, to achieve rapid and efficient domain-specific model fine-tuning, the parameter-efficient fine-tuning unit of this invention, after being deployed in the information center of University C, uses a LoRA-IA3 adapter with a explicitly set rank of 8 for efficient parameter fine-tuning. The parameter update amount is strictly controlled within 1% of the total number of parameters in the original model, and the meta-learning initialization uses the MAML algorithm. On-site data shows that the average single fine-tuning process has been reduced from 120 minutes to 12 minutes, and resource consumption has been reduced by approximately 90% compared to traditional fine-tuning methods. In addition, by recording parameter hash values ​​using zero-knowledge proofs and a blockchain ledger, once per minute, strict supervision and high trust in the model fine-tuning process are achieved.

[0043] Furthermore, to enhance the reliability and credibility of the model evaluation conclusions, this invention deploys an interpretability and robustness verification unit. Based on the dual gradient attribution method, the threshold for generating saliency heatmaps and identifying salient regions for each evaluation conclusion is set to 0.7. Adversarial samples with a perturbation amplitude of 0.01 are generated using the Fast Gradient Sign Method (FGSM) in Generative Adversarial Networks (GANs) to evaluate the model's robustness. Once an evaluation conclusion's confidence level is detected to be below 0.85 or its robustness below 0.95, it is automatically rolled back to the previous credible version, and the rollback hash is recorded in the blockchain ledger every minute. During implementation, a total of 16 rollback operations were triggered, ensuring the high credibility and stability of the evaluation conclusions.

[0044] Table 3 shows the significant improvement in the intellectual property evaluation system of University C three months after implementing the system of this invention: Table 1. Comparison of Intellectual Property Assessment Results Before and After Implementation at University C As shown in Table 1 above, after deploying the system of this invention, the intellectual property assessment process of University C has been significantly improved: the assessment time for a single patent has been reduced from 5 hours to 18 minutes, and the number of patent reports generated per month has increased dramatically from 100 to 400; the response time to regulatory changes has been shortened from 48 hours to 85 seconds; the model fine-tuning time has been reduced from approximately 120 minutes to 12 minutes, saving approximately 90% of computing resources; the accuracy of multimodal data analysis has increased from 80% to 95%; a total of 16 timely rollback operations were triggered during the deployment period, ensuring that the model results were highly reliable and that no data security incidents occurred. Furthermore, the response delay for model parameter adjustments has been significantly shortened from several hours to several days to an average of 40 seconds, greatly improving the overall assessment efficiency and reliability assurance capabilities of the intellectual property management department of University C.

[0045] As seen in the above examples, the implementation of the system of the present invention at University C has not only effectively solved the problems of low evaluation efficiency, difficulty in multimodal data fusion, lag in parameter adjustment, and lack of credibility of the original manual evaluation method, but also significantly improved the timeliness, accuracy, data security and credibility of intellectual property evaluation. It has won unanimous recognition from the technical management department of the university, which proves the outstanding advantages and technological advancement of the present invention in practical applications.

[0046] Example 2 To verify the feasibility of this invention, the interpretable federated fine-tuning large-scale model intellectual property assessment and credible report generation system of this invention was applied to a medium-sized technology company (hereinafter referred to as Company D) located in Hefei City, Anhui Province. Company D is a technology enterprise specializing in artificial intelligence chips, Internet of Things devices, and new energy technologies, with an average annual patent application volume of approximately 800. Patent assessment has become a crucial link in the company's transformation of innovative achievements and protection of intellectual property rights. However, with the continuous expansion of the company's technological fields and the rapid growth of patent applications, the traditional manual assessment methods have become increasingly problematic due to their time-consuming nature, low efficiency, slow response to regulatory changes, and difficulty in ensuring data security. These issues severely restrict the company's rapid response to technological innovation and the accuracy of decision-making. To effectively solve the above problems, Company D decided to deploy the assessment system of this invention, aiming to improve the efficiency, reliability, and security of intellectual property management.

[0047] This invention's system is deployed in the R&D data center of Company D in Hefei High-tech Zone. The data center is equipped with a high-performance server cluster, including Intel Xeon Gold processors, 128GB of memory, and Tesla A100 GPU computing units, to meet the system's complex computing needs and large-scale data processing tasks. During system operation, a distributed crawler task queue is first used to crawl regulatory events, technological developments, and patent announcements from the official websites of global patent institutions such as the China National Intellectual Property Administration, the USPTO, and the European Patent Office (EPO), as well as judicial precedent databases and international science and technology media websites, every 2 minutes. This forms a real-time updated structured data of event nodes and edge relationships. After each data collection, the semantic graph is updated using an incremental graph convolution-differential update algorithm. The node embedding dimension is set to 128 dimensions, and a multi-hop graph attention mechanism with three attention heads is used for dynamic causal reasoning and confidence labeling of the "article-term-case-judgment-market indicator" five-tuple. The average time for each graph update is approximately 80 seconds, significantly improving the real-time update capability of the graph and ensuring Company D's timely response to global regulatory and technological changes.

[0048] Due to the numerous R&D departments within Company D, the multimodal data, including patent texts, drawings, images, and R&D data tables, are stored in a scattered manner, leading to information silos and data security risks in cross-departmental data fusion. To address this, this invention establishes federated learning nodes in the company's R&D, engineering, and intellectual property departments through a federated multimodal assessment-generation unit. It employs a gated expert hybrid (MoE) model, using BERT to encode text data, ResNet-50 to encode image data, and TabNet to encode tabular data, uniformly generating a 256-dimensional embedding vector. Through federated averaging (FedAvg) and homomorphic encryption, multimodal data can be securely fused without leaving the departments. Approximately 500 data points are analyzed daily, generating an average of over 40 assessment reports.

[0049] Furthermore, the adaptive model and knowledge synchronization unit of this invention adjust model parameters in real time based on feedback signals from laws, technology, markets, and competitive landscape, once per minute. Employing the PPO reinforcement learning algorithm, it dynamically updates expert routing weights, adapter parameters, and semantic graph confidence intervals. During actual operation, the average adaptive adjustment response time of the model parameters is only about 30 seconds, significantly better than the parameter update cycle of traditional models, which can take several hours or even longer, greatly improving the model's ability to respond to changes in real time.

[0050] Meanwhile, the high-efficiency parameter fine-tuning unit of this invention achieves rapid domain fine-tuning by adjusting parameters at a ratio not exceeding 1% of the original large language model, through a LoRA-IA3 adapter with an explicitly set rank of 8. The fine-tuning process is initialized using the MAML meta-learning algorithm, reducing the time consumed by each fine-tuning task from approximately 90 minutes to about 9 minutes, and reducing computational resource usage by about 90%. Through the Groth16 zero-knowledge proof protocol, the hash value of the fine-tuned parameters is recorded to the blockchain ledger every minute, ensuring the transparency and trustworthiness of parameter updates.

[0051] Furthermore, to ensure the credibility and robustness of the model's prediction results, the interpretability and robustness verification unit of this invention utilizes the dual gradient attribution method to generate a significance heatmap. With a significance threshold of 0.7, it clearly identifies key factors affecting the prediction results and generates adversarial examples using the Fast Gradient Sign Method (FGSM) in Generative Adversarial Networks (GANs) with a perturbation amplitude of 0.01 to test model robustness. When the model's output confidence score falls below 0.85 or its robustness score falls below 0.95, an automatic rollback to the previous trusted version is triggered, and the rollback parameter hash value is automatically recorded in the blockchain ledger. During the three-month deployment and operation period, a total of nine rollback events were triggered, promptly correcting abnormal model outputs and ensuring the reliability of the evaluation results.

[0052] Three months after deploying the system of this invention, Company D's intellectual property evaluation system showed significant improvement, as detailed in Table 4: Table 2 Comparison of Data on the Implementation Effects of Intellectual Property Assessment by Company D As clearly shown in Table 2 above, after implementing this invention, Company D's patent evaluation efficiency and quality have significantly improved. The average evaluation time for a single patent has decreased from 6 hours to 15 minutes, and the monthly number of patent evaluation reports generated has jumped from 80 to approximately 1200. The dynamic response capability to regulations has improved from approximately 24 hours to an average of 80 seconds. The time for a single model fine-tuning task has been shortened from 90 minutes to 9 minutes, saving approximately 90% of fine-tuning resources. The accuracy of multimodal data analysis has increased from 78% to 95.5%. Nine automatic rollbacks were triggered during implementation, effectively ensuring the high reliability of the evaluation results. No data security incidents occurred during implementation, significantly improving the company's intellectual property data security, and the real-time response capability for model parameter adjustments has also been significantly improved.

[0053] As can be seen from the above practical examples, the implementation of the evaluation system of the present invention effectively solves the practical problems of low efficiency, high data security risks, difficulty in adjusting model parameters, and insufficient credibility of evaluation results in the intellectual property evaluation process of Company D. It significantly improves the company's intellectual property evaluation capabilities, data processing efficiency, and decision-making reliability, and has won unanimous recognition and high praise from within the company, proving the outstanding practical value and technical advantages of the present invention in practical applications.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0055] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

[0056] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable way without contradiction. In order to avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0057] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. A system for generating evaluation reports for a federally tuned large model, characterized in that, The system includes: The cross-domain regulatory event acquisition and semantic graph evolution unit is used to acquire regulatory events in global patent office announcements, judicial rulings and technology news in real time. Event nodes are incorporated into the intellectual property semantic graph through incremental graph convolution-difference algorithm, and causal reasoning and confidence labeling of the "article-term-case-precedent-market indicator" quintuple are completed based on multi-hop graph attention. The Federal Multimodal Assessment-Generation Unit is used to encode text, images, tables, and full-text patents locally at each data holder. It employs a gated expert hybrid model and a secure aggregation protocol to complete privacy-preserving joint training, and generates a structured assessment report containing verifiable evidence chains in parallel at the central side based on a cross-modal alignment matrix. An adaptive model and knowledge synchronization unit is used to map four-dimensional feedback signals of law, technology, market and competitive situation into a multi-agent reinforcement learning strategy, and dynamically adjust the gating expert weights, dual adapter parameters and semantic graph confidence intervals to achieve minute-level online updates; The parameter-efficient fine-tuning unit is used to quickly fine-tune different domain models by using parallel stacked LoRA-IA3 adapters and meta-learning initialization strategies, without exceeding one percent of the total number of trainable parameters of the original large language model. The hash values ​​of key fine-tuning parameters are written to the blockchain through zero-knowledge proof to ensure the traceability and immutability of model updates. The interpretability and robustness verification unit is used to generate cross-modal significance heatmaps and causal chain descriptions based on dual gradient attribution for each evaluation conclusion, and to perform adversarial sample robustness tests using generative adversarial networks. When the confidence of the conclusion is lower than a preset threshold, it triggers an automatic rollback to the previous trusted version.

2. The federal fine-tuning large model evaluation report generation system according to claim 1, characterized in that, The cross-domain regulatory event collection and semantic graph evolution unit performs the following steps: Establish a multi-source crawler task queue Polling to retrieve global patent office announcements data sources Data source of judicial judgments Technology News Data Source Forming an event text stream ; For text streams Word segmentation, named entity recognition, and relation extraction are performed to obtain a set of event nodes. With edge set ; The increment vector is calculated using the following formula; in For activation function, For the normalized adjacency matrix, The matrix represents the nodes at the previous time step. This is the weight matrix; Will Injection Semantic Graph get A multi-hop graph attention mechanism is used to calculate the causal score for quintuple nodes. And generate confidence labels. Write .

3. The federal fine-tuning large model evaluation report generation system according to claim 1, characterized in that, The federal multimodal evaluation-generation unit performs the following steps: At each data holder Establish a multimodal dataset locally And embedded via a shared encoder-decoder ; Expert weights are calculated using a gated expert mixture model. And generate expert output ; Using secure aggregation protocol Homomorphic encryption summation yields the global gradient Update the model after decryption; Based on the cross-modal alignment matrix Parallel generation of evidence chains The structured assessment report.

4. The federal fine-tuning large model evaluation report generation system according to claim 1, characterized in that, The adaptive model and knowledge synchronization unit are used to perform the following steps: Constructing a four-dimensional feedback vector Input Multi-Agent Policy Network Generate gated route weights With adapter gain and according to reports Calculate the policy gradient according to Update strategy parameters, press Adjust the expert routing matrix and adapter weights; Correcting the confidence interval of the semantic graph The update log will then be written to the blockchain traceability channel.

5. The federal fine-tuning large model evaluation report generation system according to claim 1, characterized in that, The parameter fine-tuning unit is used to perform the following steps: Domain dataset Freeze the original weights Insert the LoRA-IA3 adapter to form an incremental connection. And constrain the proportion of trainable parameters Initialize parameters with meta-learning Obtain domain-specific parameters as the starting point Calculate the hash of the fine-tuning parameters Generating zero-knowledge proofs And write it along with the timestamp into the blockchain ledger. .

6. The federal fine-tuning large model evaluation report generation system according to claim 1, characterized in that, The interpretability and robustness verification unit is used to perform the following steps: For input Calculation and prediction And generate a dual gradient heatmap based on Form a significant area , will support samples With hash Forming a causal chain ; By generating adversarial networks Constructing adversarial examples Calculate robustness score When confidence level or At that time, roll back to the previous trusted version. And rollback hash Upload to the blockchain.

7. A method for generating an evaluation report for a system according to any one of claims 1-6, characterized in that, The method includes: Collect cross-domain legal events and update the intellectual property semantic graph, and complete causal reasoning and confidence labeling; Multimodal encoding of text, images, tables, and full-text patents is performed locally at each data holder, and federated training is completed through secure aggregation. A structured evaluation report and a verifiable chain of evidence are generated based on the cross-modal alignment matrix; Utilizing four-dimensional feedback to drive multi-agent reinforcement learning, expert routes, adapter parameters, and semantic graph confidence intervals are synchronized within minutes. LoRA-IA3 fine-tuning is performed with parameters accounting for no more than one percent of the original model, and hashes and zero-knowledge proofs are written into the blockchain. Perform dual gradient interpretability analysis and adversarial robustness verification on the evaluation results, and automatically roll back the model version if necessary.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the evaluation report generation method of claim 7.

Images