Distributed aigc model dynamic fine-tuning method and system

Abstract

The application discloses a distributed AIGC model dynamic fine-tuning method and system. The method relates to the technical field of model dynamic fine-tuning and comprises the following steps: multi-modal business data processing and training sample generation, basic model field knowledge internalization fine-tuning, and model role adaptation and compliance constraint fine-tuning. The application generates training samples by collecting multi-modal business data and performing standardized and knowledge graph association processing; the first level fine-tuning is performed on the basic model by using supervised fine-tuning, so that the basic model masters professional terms and business processes; on the basis, the second level role adaptation fine-tuning is performed, and the compliance checking mechanism based on the law enforcement rule constraint library is introduced during decoding, so that the safety of generated content is ensured, and the problem of low safety of the distributed AIGC model dynamic fine-tuning due to the field distribution deviation of training data and the lack of parameterization and non-parameterization knowledge fusion mechanism in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of model dynamic fine-tuning technology, and in particular to a method and system for dynamic fine-tuning of distributed AIGC models. Background Technology

[0002] First, edge computing nodes are deployed in each prison area to collect non-privacy ubiquitous data (such as personnel movement and environmental conditions) through cameras and IoT sensors. The data is then processed by a local federated learning client for differential privacy noise addition and feature extraction. Only the de-identified feature vectors are uploaded to the central server of the supervisory private cloud.

[0003] Next, the central server employs a federated aggregation algorithm based on secure multi-party computation to weight and fuse the feature vectors uploaded by each edge node, updating the global AIGC (Artificial Intelligence Generated Content) model. The fine-tuning trigger mechanism falls into two categories: timed incremental learning and event-driven mechanisms (such as automatic triggering upon detection of abnormal clustering behavior). Parameter-efficient methods such as LoRA (Low-Rank Adaptation) are used, adjusting only some attention layer parameters to quickly adapt to regulatory scenarios while maintaining the general capabilities of the base model.

[0004] Then, it is deployed on the intranet of the regulatory site through API (Application Programming Interface) service to support two types of applications: one is generative tasks, such as automatically generating patrol report summaries based on natural language instructions, or training images to simulate violation scenarios; the other is analytical tasks, such as using time series models fine-tuned with video time series data to predict the risk of group behavior.

[0005] Finally, a lightweight version is compressed using model distillation technology and executed locally for real-time inference (such as detecting abnormal actions in real-time surveillance videos). Edge nodes report cases with low model confidence or uncertain decisions, along with encrypted tags, to the central server. If model drift or an increase in the false positive rate for a specific group is detected, a rollback mechanism is automatically initiated to the previous stable version, triggering a manual review process.

[0006] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems: In existing technologies, the first issue is the bias in the domain distribution of training data. The pre-training corpus for general-purpose models primarily comes from broad texts such as publicly available web pages, books, and news articles, with very little emphasis on specialized content related to regulatory enforcement, detention approval records, and surveillance procedures. This data bias means that even with fine-tuning using some legal provisions, the model cannot build a deep domain knowledge graph. It can only learn the superficial co-occurrence relationships between terms, failing to internalize the rigorous legal logic, operational procedures, and power boundaries behind them. For example, the model might confuse "detention" with the concept of ordinary isolation, failing to connect it to the complete enforcement procedure requiring deliberation, approval, time limits, and notification of rights.

[0007] Secondly, the existing architecture fails to effectively integrate parametric and non-parametric knowledge. When generating content, the model primarily relies on pre-trained parametric memory (i.e., internal weights), while externally updated regulatory databases, case libraries, and operating procedures serve only as supplementary references. It lacks a reliable mechanism to force the model to perform a "retrieval-verification-citation" process before generating key conclusions. This leads to the possibility that when processing queries involving newly revised regulations, the model may still call outdated text from the training data, resulting in factual errors. The model may incorrectly cite replaced clause numbers or procedural requirements, and it lacks adaptability to different perspectives, failing to dynamically adjust the legal depth, information security level, and stance of the expression based on the interacting parties. Furthermore, the distributed AIGC model suffers from low security due to domain distribution bias in the training data and the lack of a parametric and non-parametric knowledge integration mechanism. Summary of the Invention

[0008] To address the low security of dynamic fine-tuning of distributed AIGC models in existing technologies due to domain distribution bias in training data and the lack of a mechanism for fusing parametric and non-parametric knowledge, this invention provides a method and system for dynamic fine-tuning of distributed AIGC models. The technical solution is as follows: On the one hand, a method for dynamic fine-tuning of a distributed AIGC model is provided. This method includes: S01, collecting multimodal business data from target regulatory sites, including at least textual data, time-series behavioral data, and structured record data; performing natural language processing on the textual data to construct behavior-result-evidence triples; standardizing and mapping the multimodal business data; and associating and labeling the multimodal business data according to a pre-set law enforcement knowledge graph to generate a training sample set for model fine-tuning; S02, using a pre-trained large-scale AI-generated content model as the base model. Based on the training sample set generated in step S01, the basic model is fine-tuned at the first level using a supervised fine-tuning method to enable the model to master the professional terminology, legal document formats, and business process norms in the field of law enforcement in regulatory venues. In step S03, based on the model after the first level of fine-tuning, a second level of fine-tuning is performed to enable the model to adjust the generated content according to the identity of the interaction object. A regulatory venue law enforcement rule constraint library is also constructed. In the model decoding stage, a rule constraint mechanism is introduced to perform real-time compliance verification on the candidate content generated by the model and to filter or correct the generated content that violates the rules in the regulatory venue law enforcement rule constraint library.

[0009] On the other hand, a distributed AIGC model dynamic fine-tuning system is provided. This system includes: a multimodal business data processing and training sample generation module, a basic model domain knowledge internalization fine-tuning module, and a model role adaptation and compliance constraint fine-tuning module. The multimodal business data processing and training sample generation module collects multimodal business data from target regulatory sites. This multimodal business data includes at least textual data, time-series behavioral data, and structured record data. Natural language processing is performed on the textual data to construct behavior-result-evidence triples. The multimodal business data is standardized and mapped, and it is associated and labeled according to a pre-defined law enforcement knowledge graph to generate a training sample set for model fine-tuning. The basic model domain knowledge... The internalization fine-tuning module uses a pre-trained AI content generation model as the base model. Based on the training sample set generated in step S01, it performs first-level fine-tuning of the base model using supervised fine-tuning to enable the model to master professional terminology, legal document formats, and business process norms in the field of law enforcement in regulatory venues. The model role adaptation and compliance constraint fine-tuning module performs second-level fine-tuning on the model after the first-level fine-tuning, enabling the model to adjust the generated content according to the identity of the interaction object. It also builds a regulatory venue law enforcement rule constraint library. During the model decoding stage, a rule constraint mechanism is introduced to perform real-time compliance verification on the candidate content generated by the model and to filter or correct generated content that violates the rules in the regulatory venue law enforcement rule constraint library.

[0010] On the other hand, a computing device is provided, characterized in that it includes: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

[0011] On the other hand, a computer-readable storage medium is provided, characterized in that the computer-readable storage medium stores a computer program or instructions, which, when executed, cause the method according to any one of claims 1 to 7 to be performed.

[0012] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: First, through step S01, multimodal business data undergoes deep semantic deconstruction and knowledge graph association, transforming fragmented and unstructured law enforcement records into structured knowledge with a logical chain of "behavior-result-basis" that machines can deeply understand. This allows the model to accurately learn the terminology, document norms, and business processes specific to law enforcement in regulatory settings during the S02 stage of monitoring and fine-tuning. This ensures that the final law enforcement analysis reports, early warnings, or legal opinions generated by the intelligent agent not only meet the requirements of professional documents in form but also possess accurate legal basis and practical logic tailored to regulatory scenarios. Thus, the large language model is transformed from a "generalist" into a "domain expert" capable of providing high-quality professional assistance.

[0013] Secondly, this method systematically ensures the high compliance and operational security of AI-generated content in the high-risk scenario of regulation. In stage S03, the innovative introduction of a "regulatory site enforcement rule constraint library" and a real-time decoding and verification mechanism adds an insurmountable "rule firewall" to the model. This mechanism can automatically conduct compliance reviews of every output, every legal citation, and every disposal suggestion during the real-time reasoning process of the model-generated content, and hard-filter or automatically correct any statements that deviate from regulatory provisions and internal discipline. This effectively prevents the model from generating illegal, non-compliant, sensitive information-leaking, or potentially conflict-inducing content due to "illusions" or knowledge gaps, ensuring that the AI's generation behavior is always constrained within the rigid framework of laws, regulations, and regulatory policies. This greatly reduces the institutional risks of technology application and provides crucial security guarantees for deploying AI in highly sensitive law enforcement environments.

[0014] Finally, this method significantly enhances the human-computer interaction intelligence and scene adaptability of the agent. Through the second-level fine-tuning in S03 for different interaction objects, the model can dynamically understand the context, purpose and role relationship of the dialogue. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart of the distributed AIGC model dynamic fine-tuning method provided in this application embodiment; Figure 2 A flowchart of the adaptive training scheduling mechanism based on knowledge internalization feedback for the distributed AIGC model dynamic fine-tuning method provided in this application embodiment; Figure 3 Flowchart of the query rewrite threshold adaptive adjustment mechanism based on reference quality comparison for the distributed AIGC model dynamic fine-tuning method provided in this application embodiment; Figure 4 This is a schematic diagram of the structure of the distributed AIGC model dynamic fine-tuning system provided in an embodiment of this application. Detailed Implementation

[0017] The technical solution provided in this application will now be described with reference to the accompanying drawings.

[0018] To facilitate understanding of the embodiments of this application, the following points will be explained first: First, in this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates an "or" relationship between the preceding and following related objects, but it does not exclude the possibility of indicating an "and" relationship; the specific meaning can be understood in context. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c; a and b; a and c; b and c; or a and b and c. Here, a, b, and c can be single or multiple.

[0019] Second, the use of prefixes such as "first" and "second" in this application is solely for the purpose of distinguishing and describing different things belonging to the same category, and does not constrain the order, size, or quantity of things. For example, "first message" and "second message" are simply different messages, and there is no chronological, size, or priority relationship between them.

[0020] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0021] like Figure 1 The diagram shown is a flowchart of a distributed AIGC model dynamic fine-tuning method provided in an embodiment of this application. The method includes the following steps: S01. Collect multimodal business data from the target regulatory site. The multimodal business data includes at least text data, time-series behavioral data, and structured record data. Perform natural language processing on the text data to construct behavior-result-evidence triples. Standardize and map the multimodal business data. Associate and label the multimodal business data according to the preset law enforcement knowledge graph to generate a training sample set for model fine-tuning.

[0022] It needs to be explained that the preset law enforcement knowledge graph is specifically a pre-constructed knowledge graph of the regulatory enforcement field. This knowledge graph includes an ontology layer, a data layer, and a rule layer. The ontology layer defines law enforcement entity classes, relationship classes, and attribute classes. Law enforcement entity classes include behavior entity classes, result entity classes, basis entity classes, and role entity classes. Relationship classes include the association relationships of "behavior-need-procedure", "behavior-generation-result", and "behavior-basis-clause". The data layer contains a set of triples in the regulatory enforcement field. The set of triples includes at least the "behavior-result-basis" triple and the "entity-attribute-value" triple. The rule layer contains law enforcement compliance constraint rules, including at least the rules for closed approval procedures, the rules for setting up surveillance and intelligence, and the rules for time limits for notification of rights.

[0023] It should be understood that the specific steps for generating the training sample set for model fine-tuning are as follows: Based on the entity relationships in the pre-defined law enforcement knowledge graph, textual data, time-series behavioral data, and structured record data are cross-modal linked to form a complete law enforcement business scenario data chain. The linked multimodal business data is then labeled according to the pre-defined law enforcement knowledge graph. Labels include knowledge tags, compliance tags, and procedure tags. Knowledge tags are used to label the law enforcement domain knowledge nodes corresponding to the data, associating them with entities and triples in the knowledge graph. Compliance tags are used to label the compliance results of law enforcement actions, including compliant and non-compliant results. Procedure tags are used to label the completeness of law enforcement processes. The labeled multimodal business data is then sampled and formatted to generate a training sample set for model fine-tuning. The training sample set is formatted as an input text-tag-knowledge basis triple structure.

[0024] In this embodiment, by leveraging the entity relationships within a pre-defined law enforcement knowledge graph, textual data (such as interrogation records), temporal behavioral data (such as personnel activity trajectories), and structured record data (such as penalty decisions) from regulatory sites are cross-modal linked and aligned. This forms a complete law enforcement business scenario data chain spanning "behavior triggering - process recording - document generation," effectively solving the scenario fragmentation problem caused by isolated multi-source data. Subsequently, the linked data is labeled with multi-dimensional tags based on this knowledge graph: knowledge tags are used to label specific legal clauses, procedural nodes, and other knowledge graph entities corresponding to the data, establishing a direct connection between the data and the domain knowledge system; compliance tags are used to clearly label the compliance judgment results (compliant / non-compliant) of law enforcement actions, providing clear positive and negative learning signals for the model; and procedural tags are used to label the integrity level of the law enforcement process, strengthening the model's understanding of procedural norms. Finally, the labeled data undergoes sample partitioning and formatting to generate a triplet training sample set with a unified structure of "input text - tag - knowledge basis." This process not only transforms scattered data into structured, high-quality training resources rich in domain logic and compliance semantics, but also, through the triple reinforcement of "data chain - knowledge chain - tag chain", enables the model to learn the specific context, rigid rules and procedural requirements of regulatory business in parallel during the fine-tuning process, significantly improving the professionalism, compliance accuracy and procedural rationality of the generated content.

[0025] S02, using a pre-trained AI content generation model as the base model, and based on the training sample set generated in step S01, the base model is fine-tuned in the first level using supervised fine-tuning to enable the model to master the professional terminology, legal document formats and business process norms in the field of law enforcement in regulatory venues.

[0026] Furthermore, the specific steps for performing the first-level fine-tuning are as follows: the first-level fine-tuning includes an adaptive training scheduling mechanism based on knowledge internalization feedback and a two-factor adaptive regulation mechanism based on attention head efficiency evaluation.

[0027] It needs to be explained that, such as Figure 2The diagram shows the flowchart of the adaptive training scheduling mechanism based on knowledge internalization feedback for the distributed AIGC model dynamic fine-tuning method provided in this application embodiment. The specific process is as follows: First, initialize relevant parameters, predefine the initial graph information injection frequency and the maximum number of domain training rounds, and simultaneously preset the law enforcement knowledge graph injection frequency standard, the critical lower and upper limits of the number of domain training rounds, and the baseline lower and upper limits of the efficiency factor; then calculate the core loss and convergence rate, first obtain the deviation between the current graph information injection frequency and the preset standard (i.e., graph alignment loss), then calculate the decrease rate of the exponential moving average of this loss within the preset time window (i.e., graph loss smoothing convergence rate), and obtain the graph information internalization feedback based on the two. The efficiency factor is then used to adjust the number of training rounds. If the efficiency factor is not lower than the preset upper limit, it indicates high internalization efficiency and smooth knowledge fusion, and the number of training rounds is reduced to the critical lower limit. If the efficiency factor is within the preset benchmark range, it indicates normal internalization efficiency, and the current number of training rounds is maintained. If the efficiency factor is not higher than the preset lower limit, it indicates that the current injection frequency is insufficient to effectively absorb graph knowledge, and the number of training rounds is expanded to the critical upper limit. Finally, the training round parameters are updated and domain training is continuously performed. If the predefined maximum number of domain training rounds is not reached, the core loss and convergence rate are recalculated to enter the next round of closed-loop adjustment. If the target is reached, the adaptive training scheduling ends.

[0028] It should be further explained that the specific steps of the adaptive training scheduling mechanism based on knowledge internalization feedback are as follows: An initial graph information injection frequency and a predefined maximum number of domain training rounds are defined. The deviation between the current graph information injection frequency and the injection frequency standard in the preset law enforcement knowledge graph is recorded as the graph alignment loss. The rate of decrease of the exponential moving average of the graph alignment loss within a preset time window is recorded as the graph loss smoothing convergence rate. Based on the graph alignment loss and the graph loss smoothing convergence rate, the graph information internalization efficiency factor is obtained. The following closed-loop adjustment is performed based on the graph information internalization efficiency factor: if the graph information internalization efficiency factor is greater than or equal to the preset efficiency factor benchmark upper limit, it indicates that the current internalization efficiency of graph information is high. If knowledge integration is smooth, the number of domain training rounds will be reduced to the predefined critical lower limit of the number of domain training rounds. If the graph information internalization efficiency factor is within the preset efficiency factor benchmark range, it indicates that the internalization efficiency is within the normal range, and the current number of domain training rounds will remain unchanged. The preset efficiency factor benchmark range represents the open interval formed by the preset efficiency factor benchmark lower limit and the preset efficiency factor benchmark upper limit. If the graph information internalization efficiency factor is less than or equal to the preset efficiency factor benchmark lower limit, it indicates that it is difficult to effectively absorb graph knowledge at the current graph information injection frequency, and the number of domain training rounds will be expanded to the predefined critical upper limit of the number of domain training rounds.

[0029] It needs to be explained that, based on the graph alignment loss and the graph loss smoothing convergence rate, the specific constraint expression for the graph information internalization efficiency factor is as follows: ; ; In the formula, This represents the map alignment loss, normalized to the [0,1] interval; This represents the efficiency factor for internalizing spectral information. Indicates the smoothing convergence rate of the spectral loss; and This represents the preset weighting coefficients used to balance the contribution of absolute loss level and convergence trend. By quantifying and evaluating the effectiveness of the "graph information injection" intervention in real time and using this as feedback signal to dynamically adjust the frequency of intervention and the overall training duration, the following are achieved: Optimized training efficiency: Avoiding redundant injection and training when the model has quickly mastered knowledge, or insufficient training when the model encounters bottlenecks, thus saving computational resources. Guaranteed knowledge fusion quality: By dynamically increasing the number of training rounds and injection frequency, ensuring that the model receives sufficient correction and guidance when facing complex logical gaps, thereby improving the depth and reliability of domain knowledge internalization. Adaptive and automated: The entire process requires no repeated manual parameter tuning; the system can automatically find the optimal training rhythm and cycle based on the model's actual learning state.

[0030] In this embodiment, firstly, an initial graph information injection frequency and a maximum number of training rounds are predefined as training benchmarks. During training, the real-time graph alignment loss is obtained by calculating the deviation between the current graph information injection frequency and the standard frequency in the preset law enforcement knowledge graph. Furthermore, the exponential moving average decline rate of this loss within a preset time window is calculated as the graph loss smoothing convergence rate. Based on this, the alignment loss and convergence rate are fused to calculate the graph information internalization efficiency factor, which quantifies the model's efficiency in absorbing structured knowledge. Subsequently, closed-loop adjustments are implemented based on the efficiency factor: when the efficiency factor is higher than the benchmark upper limit, it indicates smooth knowledge fusion, and the number of training rounds is automatically reduced to the critical lower limit to avoid overtraining and save computational resources; when the efficiency factor is within the normal benchmark range, the original training plan is maintained; when the efficiency factor is lower than the benchmark lower limit, it indicates difficulty in knowledge absorption, and the system expands the number of training rounds to the critical upper limit, giving the model sufficient learning time to overcome the bottleneck. This method achieves an optimal balance between training efficiency and knowledge fusion depth by monitoring the knowledge internalization status in real time and dynamically adjusting the training intensity. While ensuring that the model fully grasps the domain knowledge, it significantly reduces unnecessary computational consumption and improves the intelligence and economy of the fine-tuning process.

[0031] It should be further explained that the specific steps of the two-factor adaptive regulation mechanism based on attention head efficacy assessment are as follows: Assign identifiable identifiers to specific attention heads used for processing knowledge graph injection information. Predefine initial values ​​for attention weight scaling factors and temporary gradient scaling factors. The initial values ​​for attention weight scaling factors are used to amplify the gradients generated by the loss during the training of knowledge graph injection information. Obtain the graph attention head weights and the average attention probability values ​​assigned to graph entity tokens in the attention weight matrix within a preset injection period. In each training step of knowledge graph injection information, calculate the comprehensive performance index of each graph attention head. The comprehensive performance index is obtained by the weighted geometric mean of the graph attention head weights and the average attention probability values.

[0032] It should be noted that the two-factor adaptive regulation mechanism based on attention head efficacy assessment also includes: Based on the overall performance index of each graph attention head, the attention weight scaling factor and the temporary gradient scaling factor are adjusted as follows: If the overall performance index is greater than the performance threshold, it indicates that the attention head is working efficiently. The initial value of the attention weight scaling factor is increased to the upper limit of the attention weight scaling factor, while the initial value of the temporary gradient scaling factor remains unchanged. If the overall performance index is within the performance threshold range, it indicates that the head is working normally. The predefined initial values ​​of the attention weight scaling factor and the temporary gradient scaling factor are maintained. The performance threshold range represents the closed interval formed by the lower and upper limits of the performance threshold. If the overall performance index is less than the lower limit of the performance threshold, it indicates that the attention head has not been effectively activated. In the next forward propagation, the initial value of the attention weight scaling factor is increased to the upper limit of the attention weight scaling factor, while the initial value of the temporary gradient scaling factor is reduced to the lower limit of the gradient scaling factor.

[0033] In this embodiment, firstly, a recognizable identifier is assigned to a specific attention head that processes knowledge graph injection information, and initial values ​​for its attention weight scaling factor and temporary gradient scaling factor are predefined. During training, the weight parameters of each attention head and the average probability value of the tokens assigned to the graph entities in its attention matrix are collected in real time within each preset injection cycle. Then, a comprehensive performance index for each head is calculated using a weighted geometric average. This index quantifies the activation intensity and contribution of the head in the knowledge fusion task. Subsequently, precise control is implemented based on the performance index: when the index is higher than the critical upper limit, it indicates that the head is in a highly efficient working state, and its attention weight scaling factor is increased to the baseline upper limit to enhance its influence in forward propagation, while keeping the gradient scaling factor unchanged to avoid over-optimization; when the index is within the normal critical range, the initial values ​​of the two factors are maintained; when the index is lower than the critical lower limit, it indicates that the head has not been effectively activated, and in the next training step, its attention weight factor is simultaneously increased and its gradient scaling factor is decreased. By strengthening attention guidance and weakening gradient updates, the head is synergistically incentivized to refocus on the knowledge injection information. This mechanism achieves intelligent allocation of computing resources on the attention head through real-time evaluation and dynamic adjustment, effectively improving the efficiency and depth of knowledge graph information and model parameter fusion, while enhancing the stability and convergence quality of the training process.

[0034] S03. Based on the model after the first level of fine-tuning, a second level of fine-tuning is performed to enable the model to adjust the generated content according to the identity of the interactive object, and to build a regulatory site law enforcement rule constraint library. In the model decoding stage, a rule constraint mechanism is introduced to perform real-time compliance verification on the candidate content generated by the model, and to filter or correct the generated content that violates the rules in the regulatory site law enforcement rule constraint library.

[0035] Furthermore, the second level of fine-tuning includes implementing an adaptive adjustment mechanism for query rewrite thresholds based on reference quality comparison and an adaptive control mechanism for traffic recovery ratio based on feedback from circuit breaker severity and quiet period.

[0036] It should be noted that, as Figure 3The diagram shows the flowchart of the adaptive adjustment mechanism for the query rewriting threshold based on reference quality comparison in the distributed AIGC model dynamic fine-tuning method provided in this application embodiment. The specific process is as follows: At the end of each preset statistical period, the effective forced reference rate (i.e., "reference rate rewriting value" and "reference rate non-triggered value") of the query subsets that triggered rewriting and those that did not are calculated respectively. The effect of the current rewriting strategy is evaluated by comparing the two values: if the rewriting value is higher than or equal to the non-triggered value, the rewriting is deemed effective, and the system lowers the query rewriting confidence threshold to the lower limit of the baseline to allow more queries to optimize retrieval through rewriting; conversely, if the rewriting value is lower than the non-triggered value, the rewriting may introduce noise, and the system raises the threshold to the upper limit of the baseline to tighten the rewriting triggering conditions. The adjusted new threshold will be applied in the next statistical period, thereby forming a continuous closed-loop optimization, aiming to dynamically improve the reference quality and system reliability of the overall query processing.

[0037] It should be further explained that the specific steps for implementing the query rewrite threshold adaptive adjustment mechanism based on reference quality comparison are as follows: At the end of each preset statistical period, the effective forced citation rate is obtained for the subset of queries that triggered rewriting and the subset of queries that did not trigger rewriting, denoted as the citation rate rewriting value and the citation rate not triggered value. Based on the citation rate rewriting value and the citation rate not triggered value, the query rewriting confidence threshold is adjusted as follows: If the citation rate rewriting value is greater than or equal to the citation rate not triggered value, it indicates that the current query rewriting has effectively improved the citation quality and the citation frequency is normal. The query rewriting confidence threshold is reduced to the lower limit of the query rewriting confidence benchmark, allowing more queries to optimize their retrieval through rewriting, thereby improving the overall citation quality. If the citation rate rewriting value is less than the citation rate not triggered value, it indicates that the current rewriting has introduced noise and the citation is inappropriate. The query rewriting confidence threshold is increased to the upper limit of the query rewriting confidence benchmark to tighten the rewriting triggering conditions and reduce the decline in citation quality caused by inappropriate rewriting.

[0038] In this embodiment, at the end of each preset statistical period, the proportion of generated content deemed appropriately and necessary for citation is calculated in the subset of requests that triggered query rewriting and the subset of requests that did not trigger rewriting, respectively; these are known as the "citation rate rewriting value" and the "citation rate not triggered value." By comparing these two key indicators, the system can objectively assess the actual impact of query rewriting operations on the final quality of knowledge citation: if the citation rate rewriting value is higher than or equal to the non-triggered value, it proves that the current rewriting strategy is effective, and the query rewriting confidence threshold will be automatically lowered to the lower limit of the baseline, thereby relaxing the rewriting triggering conditions and allowing more original queries to improve their citation quality through the optimized retrieval process; conversely, if the citation rate rewriting value is low, it indicates that the rewriting may introduce noise or distort the original meaning, and the system will raise the threshold to the upper limit of the baseline, tightening the triggering conditions to avoid low-quality rewriting. This mechanism, through closed-loop feedback and dynamic optimization, achieves real-time alignment between the query preprocessing strategy and the final output quality, significantly improving the overall reliability and knowledge citation accuracy of the retrieval enhancement generation system, while reducing the cost of manual intervention.

[0039] It should be further explained that the specific steps of the adaptive control mechanism for the flow recovery ratio based on feedback from the severity of the circuit breaker and the quiet period are as follows: When a circuit breaker is triggered, a severity coefficient is assigned based on whether the warning is low-level, high-level, or fatal. The actual silent period within the time window before the circuit breaker is triggered is taken, and a recovery caution coefficient is calculated based on the actual silent period duration, adjusting the recovery ratio accordingly. After the silent period ends, the recovery caution coefficient is calculated again based on the actual silent period duration. The actual silent period duration and the recovery caution coefficient are negatively correlated. The initial recovery ratio sequence is dynamically adjusted using the recovery caution coefficient to generate the final execution sequence. For each target recovery ratio in the initial sequence, a dynamic value of the target recovery ratio is calculated. The dynamic value of the target recovery ratio is obtained by proportionally processing each target recovery ratio to the recovery caution coefficient. A lower limit and an upper limit of the dynamic baseline for the recovery ratio are applied to protect the dynamic value of the target recovery ratio, ensuring that the dynamic value of the target recovery ratio is within the dynamic baseline range of the recovery ratio, forming a more conservative recovery schedule. The dynamic baseline range of the recovery ratio represents the range of the formed closed interval.

[0040] It should be explained that the specific constraint expression for calculating the recovery caution coefficient based on the actual duration of the quiet period is as follows: ; In the formula, This indicates a return to a cautious level; This represents a predefined conservative baseline value for recovery; This represents a predefined attenuation factor; Indicates the actual duration of the silent period; This represents a predefined baseline value for the quiet period duration. By transforming the quiet period duration from a fixed value into a dynamic variable reflecting the severity of the fault and the system state, and using this as the core to derive matching, more cautious or more aggressive traffic recovery strategies, the following are achieved: Intelligent trade-off between recovery security and efficiency: severe faults automatically receive a longer cooling-off period and a slower recovery, while slower faults can accelerate recovery, avoiding the risks or inefficiencies of a "one-size-fits-all" strategy. A flexible and adaptive recovery process: the recovery process itself can be fine-tuned in real time based on feedback during execution, and even restart a more stringent recovery cycle, improving the system's resilience and self-healing capabilities in the face of complex faults. Data-driven strategy optimization: the parameters of the entire recovery process can be self-optimized based on historical performance, forming continuously improving operational capabilities.

[0041] In this embodiment, when a circuit breaker is triggered, a severity coefficient is first assigned according to the circuit breaker level (low-level warning, high-level warning, or fatal error). The fluctuations of key indicators within a preset time window before the circuit breaker is recorded, and the actual quiet period duration is dynamically determined accordingly. After the quiet period ends, a recovery caution coefficient is calculated based on this actual duration. This coefficient is negatively correlated with the quiet period duration; that is, the more severe the problem and the longer the quiet period, the more cautious the recovery process must be. Subsequently, this coefficient is used to dynamically scale a preset initial traffic recovery ratio sequence: each target recovery ratio in the sequence is proportional to the recovery caution coefficient to obtain a corresponding dynamic recovery value. By applying lower and upper limits for protection, it is ensured that the traffic ratio at each stage remains within a safe and controllable baseline range, thereby generating a more conservative, progressive recovery schedule that matches the severity of the fault. This method dynamically couples circuit breaker severity, system history, and recovery strategy to achieve intelligent closed-loop management from fault identification to safe recovery. It avoids overly conservative recovery under minor faults and ensures robust self-healing of the system after severe faults, significantly improving the availability of distributed AIGC services in high-risk regulatory environments.

[0042] like Figure 4The diagram shown illustrates the structure of the distributed AIGC model dynamic fine-tuning system provided in this embodiment. The distributed AIGC model dynamic fine-tuning system includes: a multimodal business data processing and training sample generation module, a basic model domain knowledge internalization fine-tuning module, and a model role adaptation and compliance constraint fine-tuning module. The multimodal business data processing and training sample generation module collects multimodal business data from target regulatory locations. This multimodal business data includes at least textual data, time-series behavioral data, and structured record data. Natural language processing is performed on the textual data to construct behavior-result-evidence triples. The multimodal business data is standardized and mapped, and it is associated and labeled according to a preset law enforcement knowledge graph to generate data for model fine-tuning. The system includes: a training sample set for adjustment; a basic model domain knowledge internalization fine-tuning module, which uses a pre-trained AI content generation model as the basic model, and performs first-level fine-tuning on the basic model based on the training sample set generated in step S01 using supervised fine-tuning, to enable the model to master professional terminology, legal document formats, and business process norms in the field of law enforcement in regulatory venues; and a model role adaptation and compliance constraint fine-tuning module, which performs second-level fine-tuning on the model after the first-level fine-tuning, enabling the model to adjust the generated content according to the identity of the interaction object, and constructing a regulatory venue law enforcement rule constraint library. In the model decoding stage, a rule constraint mechanism is introduced to perform real-time compliance verification on the candidate content generated by the model, and to filter or correct the generated content that violates the rules in the regulatory venue law enforcement rule constraint library.

[0043] In this embodiment, the multimodal business data processing and training sample generation module is responsible for collecting and integrating multimodal data such as text records, time-series behavioral data, and structured forms from regulatory sites. It constructs "behavior-result-basis" triples by performing natural language processing on the text data, and performs standardized mapping and association annotation based on a pre-set law enforcement knowledge graph on all data. This generates a training sample set rich in domain logic and compliance labels, solving the problems of scattered and heterogeneous regulatory data and obscure professional logic, providing a high-quality, structured knowledge source for model learning. The basic model domain knowledge internalization fine-tuning module, based on the pre-trained AIGC large model, uses the aforementioned generated training sample set for first-level supervised fine-tuning. This enables the model to deeply internalize the professional terminology system, legal document norms, and business process logic of the regulatory site law enforcement domain, significantly improving the domain professionalism and format accuracy of the model-generated content. The model role adaptation and compliance constraint fine-tuning module, based on the internalization of domain knowledge, implements a second level of fine-tuning to endow the model role with perceptual capabilities. This allows it to dynamically adjust the expression style, information depth, and legal stance of the generated content according to the identity of the interaction object. Simultaneously, by constructing a regulatory site enforcement rule constraint library and introducing a real-time compliance verification mechanism during the decoding stage, the generated candidate content is automatically filtered or corrected, ensuring that the output content strictly complies with regulatory rules and fundamentally preventing the generation of illegal, sensitive, or inappropriate content. These three modules are interconnected and progressively enhance each other, jointly achieving a complete transformation from raw data to an intelligent generation model with deep domain knowledge, precise role adaptability, and rigid compliance guarantees. This provides regulatory sites with a safe, professional, and reliable AI-assisted tool.

[0044] This application also provides a computing device, characterized in that it includes: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

[0045] One embodiment of this application also provides a computer-readable storage medium for storing a program that, when executed by a processor, implements a method for dynamically fine-tuning a distributed AIGC model.

[0046] Based on the above embodiments, this disclosure also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a computer, it causes the computer to perform any of the methods provided in the above embodiments. The storage medium can be any available medium accessible to a computer. By way of example, but not limited to, a computer-readable medium may include RAM, ROM, electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having an instruction or data structure form and accessible by a computer.

[0047] The various features and processes described above can be used independently of each other or can be combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of this disclosure. Furthermore, certain method or process blocks may be omitted in some embodiments. The methods and processes described herein are not limited to any particular order, and the blocks or states associated with them may be performed in other suitable orders. For example, the described blocks or states may be performed in an order different from the order specifically disclosed, or multiple blocks or states may be combined in a single block or state. Example blocks or states may be performed serially, in parallel, or in some other manner. Blocks or states may be added to or removed from the disclosed example embodiments. The exemplary systems and components described herein may be configured differently from those described. For example, elements may be added to, removed from, or rearranged compared to the disclosed example embodiments.

[0048] The various operations of the example methods described herein can be performed at least in part by an algorithm. This algorithm can be contained in program code or instructions stored in memory (e.g., the aforementioned non-transitory computer-readable storage medium). Such an algorithm may include a machine learning algorithm. In some embodiments, the machine learning algorithm may not be explicitly programmed into the computer to perform the function, but can learn from training data to create a predictive model that performs the function.

[0049] The various operations of the example methods described herein can be performed, at least in part, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors can constitute the engine of a processor implementation that operates to perform one or more of the operations or functions described herein.

[0050] Similarly, the methods described herein can be implemented at least in part by a processor, where one or more specific processors are examples of hardware. For example, at least some operations of a method can be performed by one or more processors or an engine implemented by a processor. Furthermore, one or more processors can also be operated to support the performance of related operations in a “cloud computing” environment or as “Software as a Service” (SaaS). For example, at least some operations can be performed by a set of computers (as an example of a machine including processors), where these operations are accessible via a network (e.g., the Internet) and via one or more suitable interfaces (e.g., application programming interfaces (APIs)).

[0051] The performance of certain operations can be distributed across processors, residing not only within a single machine but also deployed across multiple machines. In some example embodiments, the processor or processor-implemented engine may reside in a single geographic location (e.g., within a home environment, office environment, or server cluster). In other example embodiments, the processor or processor-implemented engine may be distributed across multiple geographic locations.

[0052] In this specification, multiple instances may implement components, operations, or structures described as single instances. Although individual operations of one or more methods are shown and described as separate operations, one or more of the separate operations may be performed simultaneously and do not need to be performed in the order shown. Structures and functions presented as separate components in the example configuration may be implemented as composite structures or components. Similarly, structures and functions presented as single components may be implemented as separate components. These and other variations, modifications, additions, and improvements fall within the scope of this document.

[0053] While an overview of the subject matter has been described with reference to specific example embodiments, various modifications and changes can be made to these embodiments without departing from the broader scope of embodiments of this disclosure. Such embodiments of the subject matter are referred to herein, individually or collectively, by the term "invention," and are used for convenience only and are not intended to limit the scope of this application to any single disclosure or concept, should more than one disclosure or concept be disclosed in fact.

[0054] The embodiments described herein have been described in sufficient detail to enable those skilled in the art to practice the disclosed teachings. Other embodiments may be used and derived therefrom, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. Therefore, the detailed description should not be construed as limiting, and the scope of the various embodiments is defined only by the appended claims and the full scope of their equivalents.

Claims

1. A method for dynamic fine-tuning of a distributed AIGC model, characterized in that, Includes the following steps: S01, collect multimodal business data from the target regulatory site. The multimodal business data includes at least text data, time-series behavioral data, and structured record data. Perform natural language processing on the text data to construct behavior-result-evidence triples. Standardize and map the multimodal business data. Associate and label the multimodal business data according to a preset law enforcement knowledge graph to generate a training sample set for model fine-tuning. S02, using a pre-trained AI content generation model as the base model, and based on the training sample set generated in step S01, the base model is fine-tuned in the first level using a supervised fine-tuning method to enable the model to master the professional terminology, legal document formats and business process norms in the field of law enforcement in regulatory venues. S03. Based on the model after the first level of fine-tuning, a second level of fine-tuning is performed to enable the model to adjust the generated content according to the identity of the interactive object, and to build a regulatory site law enforcement rule constraint library. In the model decoding stage, a rule constraint mechanism is introduced to perform real-time compliance verification on the candidate content generated by the model, and to filter or correct the generated content that violates the rules in the regulatory site law enforcement rule constraint library.

2. The distributed AIGC model dynamic fine-tuning method as described in claim 1, characterized in that, The specific steps for generating the training sample set for model fine-tuning are as follows: Based on the entity relationships in the pre-set law enforcement knowledge graph, textual data, time-series behavioral data and structured record data are cross-modal associated to form a complete law enforcement business scenario data chain; The associated multimodal business data is labeled according to a pre-defined law enforcement knowledge graph, and the labels include knowledge labels, compliance labels, and procedure labels. The knowledge tags are used to label the law enforcement knowledge nodes corresponding to the data and associate them with entities and triples in the knowledge graph. The compliance label is used to indicate the compliance outcome of law enforcement actions, and the compliance outcome includes compliance and non-compliance; The program tags are used to indicate the completeness of the law enforcement process; The labeled multimodal business data is divided and formatted to generate a training sample set for model fine-tuning. The training sample set is in the format of an input text-label-knowledge basis triplet structure.

3. The distributed AIGC model dynamic fine-tuning method as described in claim 1, characterized in that, The specific steps for performing the first-level fine-tuning are as follows: The first level of fine-tuning includes an adaptive training scheduling mechanism based on knowledge internalization feedback and a two-factor adaptive regulation mechanism based on attention head efficiency evaluation. The specific steps of the adaptive training scheduling mechanism based on knowledge internalization feedback are as follows: The initial graph information injection frequency and the maximum number of training rounds in the domain are predefined. The deviation between the current graph information injection frequency and the injection frequency standard in the preset law enforcement knowledge graph is recorded as the graph alignment loss. The rate of decrease of the exponential moving average of the graph alignment loss within the preset time window is recorded as the graph loss smoothing convergence rate. Based on the graph alignment loss and graph loss smoothing convergence rate, the graph information internalization efficiency factor is obtained. The following closed-loop adjustment is performed based on the graph information internalization efficiency factor: if the graph information internalization efficiency factor is greater than or equal to the preset efficiency factor benchmark upper limit, it indicates that the current internalization efficiency of graph information is high and the knowledge fusion is smooth. The number of domain training rounds is reduced to the predefined critical lower limit of the number of domain training rounds. If the internalization efficiency factor of the graph information is within the preset efficiency factor benchmark range, it indicates that the internalization efficiency is within the normal range, and the current number of domain training rounds remains unchanged. The preset efficiency factor benchmark range represents the open interval formed by the preset lower limit of the efficiency factor benchmark and the preset upper limit of the efficiency factor benchmark. If the graph information internalization efficiency factor is less than or equal to the preset efficiency factor baseline lower limit, it indicates that it is difficult to effectively absorb graph knowledge under the current graph information injection frequency, and the domain training rounds should be extended to the predefined critical upper limit of the domain training rounds.

4. The distributed AIGC model dynamic fine-tuning method as described in claim 3, characterized in that, The specific steps of the two-factor adaptive regulation mechanism based on attention head efficacy assessment are as follows: Assign identifiable identifiers to specific attention heads used for processing knowledge graph injection information, predefine initial values ​​for attention weight scaling factors and temporary gradient scaling factors, wherein the initial values ​​for attention weight scaling factors are used to amplify the gradient generated by the loss during training of knowledge graph injection; Within a preset injection period, the graph attention head weights and the average attention probability values ​​assigned to graph entity tokens in the attention weight matrix are obtained. In each training step of graph information injection, the comprehensive performance index of each graph attention head is calculated. The comprehensive performance index is obtained by the weighted geometric average of the graph attention head weights and the average attention probability values.

5. The distributed AIGC model dynamic fine-tuning method as described in claim 4, characterized in that, The two-factor adaptive regulation mechanism based on attention head efficacy assessment also includes: Based on the overall performance index of each graph attention head, the attention weight scaling factor and the temporary gradient scaling factor are adjusted as follows: If the overall efficiency index is greater than the efficiency threshold, it indicates that the attention head of the graph is in a highly efficient working state. The initial value of the attention weight scaling factor is increased to the upper limit of the attention weight scaling factor benchmark, while the initial value of the temporary gradient scaling factor remains unchanged. If the overall performance index is within the performance critical interval, it indicates that the head is working normally and the initial values ​​of the predefined attention weight scaling factor and the temporary gradient scaling factor are maintained. The performance critical interval represents the closed interval formed by the lower limit of the performance critical limit and the upper limit of the performance critical limit. If the overall performance index is less than the lower limit of the performance threshold, it indicates that the attention head of the graph has not been effectively activated. In the next forward propagation, the initial value of the attention weight scaling factor will be increased to the upper limit of the attention weight scaling factor benchmark, and the initial value of the temporary gradient scaling factor will be reduced to the lower limit of the gradient scaling factor benchmark.

6. The distributed AIGC model dynamic fine-tuning method as described in claim 1, characterized in that, The second-level fine-tuning includes implementing an adaptive adjustment mechanism for query rewrite thresholds based on reference quality comparison and an adaptive control mechanism for traffic recovery ratio based on feedback from circuit breaker severity and quiet period. The specific steps for implementing the query rewrite threshold adaptive adjustment mechanism based on reference quality comparison are as follows: At the end of each preset statistical period, the effective forced reference rate is obtained in the query subset that triggered rewriting and the query subset that did not trigger rewriting, and is recorded as the reference rate rewriting value and the reference rate non-triggered value. Based on the reference rate rewrite value and the reference rate not triggered value, the query rewrite confidence threshold is adjusted as follows: If the citation rate rewrite value is greater than or equal to the citation rate not triggered value, it indicates that the current query rewrite has effectively improved the citation quality and the citation frequency is normal. The query rewrite confidence threshold is reduced to the lower limit of the query rewrite confidence benchmark, so that more queries can be optimized through rewriting, thereby improving the overall citation quality. If the reference rate rewrite value is less than the reference rate not triggered value, it indicates that the current rewrite has introduced noise and the reference is improper. The query rewrite confidence threshold will be increased to the upper limit of the query rewrite confidence benchmark to tighten the rewrite triggering conditions and reduce the decline in reference quality caused by improper rewrite.

7. The distributed AIGC model dynamic fine-tuning method as described in claim 6, characterized in that, The specific steps of the adaptive control mechanism for flow recovery ratio based on feedback of circuit breaker severity and silent period are as follows: When a circuit breaker is triggered, it is triggered based on whether it is a low-level warning, a high-level warning, or a fatal error. A severity coefficient is assigned to each level. The actual silence period within a certain time window before the circuit breaker is triggered is taken. The recovery caution coefficient is calculated based on the actual silence period duration, and the recovery ratio is adjusted accordingly. After the silence period ends, the recovery caution coefficient is calculated based on the actual silence period duration. The actual silence period duration is negatively correlated with the recovery caution coefficient. The initial recovery ratio sequence is dynamically adjusted using the recovery caution coefficient to generate the final execution sequence. For each target recovery ratio in the initial sequence, a dynamic value of the target recovery ratio is calculated. The dynamic value of the target recovery ratio is obtained by proportionally processing each target recovery ratio with a recovery caution coefficient. A lower limit and an upper limit of the dynamic benchmark for the recovery ratio are applied to the dynamic value of the target recovery ratio to ensure that the dynamic value of the target recovery ratio is within the dynamic benchmark interval of the recovery ratio, thus forming a more conservative recovery timetable. The dynamic benchmark interval of the recovery ratio represents the range of the closed interval formed.

8. A system applying the distributed AIGC model dynamic fine-tuning method as described in any one of claims 1-7, characterized in that, include: The module includes a multimodal business data processing and training sample generation module, a basic model domain knowledge internalization fine-tuning module, and a model role adaptation and compliance constraint fine-tuning module. The multimodal business data processing and training sample generation module is used to collect multimodal business data from the target regulatory site. The multimodal business data includes at least text data, time-series behavioral data and structured record data. Natural language processing is performed on the text data to construct behavior-result-evidence triples. The multimodal business data is standardized and mapped, and the multimodal business data is associated and labeled according to a preset law enforcement knowledge graph to generate a training sample set for model fine-tuning. The basic model domain knowledge internalization fine-tuning module is used to use a pre-trained artificial intelligence content generation model as the basic model. Based on the training sample set generated in step S01, the basic model is fine-tuned in a supervised manner to enable the model to master the professional terminology, legal document format and business process norms in the field of law enforcement in regulatory venues. The model role adaptation and compliance constraint fine-tuning module is used to perform a second-level fine-tuning based on the model after the first-level fine-tuning, so that the model can adjust the generated content according to the identity of the interaction object, and build a regulatory site law enforcement rule constraint library. In the model decoding stage, a rule constraint mechanism is introduced to perform real-time compliance verification on the candidate content generated by the model, and to filter or correct the generated content that violates the rules in the regulatory site law enforcement rule constraint library.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program or instructions that, when executed, cause the method according to any one of claims 1 to 7 to be performed.

Images