Business planning intelligent decomposition method, device and medium based on large model fine tuning

Abstract

The application discloses a business planning intelligent decomposition method and device based on large model fine tuning and a medium, relates to the technical field of artificial intelligence, and comprises the following steps: acquiring business comprehensive data, and performing scene-based general logic analysis on the business comprehensive data to determine a pre-fine tuning baseline model; based on the pre-fine tuning baseline model, real-time fine tuning streaming data is obtained through directional fine tuning of an element mask; model dynamic hierarchical adjustment analysis is performed on the real-time fine tuning streaming data to determine a business dynamic planning knowledge graph; according to the business dynamic planning knowledge graph, an initial decomposition scheme is obtained through context retrieval enhanced business planning decomposition; and task compliance verification is performed on the initial decomposition scheme to determine a business planning decomposition scheme. The application solves the technical problems of insufficient business planning decomposition adaptation and non-standard verification in the prior art through the above method.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a business planning intelligent decomposition method, device and medium based on large model fine-tuning. Background Technology

[0002] With the rapid development of artificial intelligence technology, Large Language Models (LLMs) are increasingly widely used in fields such as natural language processing and content generation. In the field of enterprise business management, using artificial intelligence to assist in the generation, decomposition, and tracking of business plans has become an important trend to improve management efficiency and scientific rigor. Traditional automation methods typically rely on pre-defined fixed rule templates and decision trees to generate structured task lists through simple logical judgments.

[0003] In existing technologies, business planning for enterprise departments is a key node in multi-agent analysis, and its decomposition has become a technical solution urgently requiring intelligent processing. However, existing large-scale models frequently exhibit a disconnect between semantic understanding and business logic, and the compliance and data privacy of business decomposition rules cannot meet the actual needs of intelligent business decomposition. Summary of the Invention

[0004] This application provides a business planning intelligent decomposition method, device, and medium based on large model fine-tuning, which solves the technical problems of insufficient adaptability and non-standard verification in the prior art.

[0005] In a first aspect, embodiments of this application provide a business planning intelligent decomposition method based on large model fine-tuning. The method includes: acquiring comprehensive business data and performing scenario-based general logic analysis on the comprehensive business data to determine a pre-fine-tuning baseline model; obtaining real-time fine-tuning streaming data based on the pre-fine-tuning baseline model through targeted fine-tuning using feature masks; performing dynamic hierarchical adjustment analysis on the real-time fine-tuning streaming data to determine a business dynamic planning knowledge graph; obtaining an initial decomposition scheme based on the business dynamic planning knowledge graph through context-enhanced business planning decomposition using context retrieval; and verifying the task compliance of the initial decomposition scheme to determine the business planning decomposition scheme.

[0006] In one implementation of this application, scenario-based general logic analysis is performed on comprehensive business data to determine a pre-fine-tuned baseline model. Specifically, this includes: analyzing the comprehensive business data using business management terminology to identify business management keywords; constructing a scenario-based dataset based on these keywords, and freezing the basic parameters of the scenario-based dataset through lightweight adjustments to obtain the model training range; determining the pre-fine-tuned model based on the model training range through joint optimization of loss terms; wherein the joint optimization of loss terms includes: cross-entropy loss optimization and rule matching loss optimization; and evaluating the pre-fine-tuned model's baseline to determine the pre-fine-tuned baseline model.

[0007] In one implementation of this application, real-time fine-tuned streaming data is obtained based on a pre-fine-tuned baseline model through targeted fine-tuning of feature masks. Specifically, this includes: acquiring non-public departmental data and performing local masking on the non-public departmental data to determine anonymized departmental data; fine-tuning the anonymized departmental data based on the pre-fine-tuned baseline model to obtain specialized fine-tuning agents; wherein, the specialized fine-tuning agents include: decomposition agents, tracking agents, and summarizing agents; and recording the gradient aggregation process corresponding to the specialized fine-tuning agents to obtain real-time fine-tuned streaming data.

[0008] In one implementation of this application, local masking is performed on non-public departmental data to determine anonymized departmental data. Specifically, this includes: calculating the local model gradient of the non-public departmental data, and performing local masking on the local model gradient to obtain mask gradient data; aggregating the mask gradient data into model update gradient, and updating the global model based on the model update gradient to determine the anonymized departmental data.

[0009] In one implementation of this application, dynamic hierarchical adjustment analysis of real-time fine-tuning streaming data is performed to determine the business dynamic planning knowledge graph. Specifically, this includes: determining the model warning level based on real-time fine-tuning streaming data through multi-indicator model health monitoring; wherein the monitoring indicators for multi-indicator model health monitoring include: output compliance rate, user correction rate, and environmental change rate; performing online model fine-tuning triggered by the model warning level to obtain the fine-tuned model version; verifying the fine-tuning effect of the fine-tuned model version; if the verification result is successful, automatically saving the fine-tuned model version and updating the dynamic scenario memory to determine the business dynamic planning knowledge graph; if the verification result is unsuccessful, automatically reverting to the model version before fine-tuning and re-analyzing the warning reasons.

[0010] In one implementation of this application, an initial decomposition scheme is obtained by decomposing the business plan using contextual retrieval-enhanced methods based on a business dynamic planning knowledge graph. Specifically, this includes: performing departmental effective task retrieval on the business dynamic planning knowledge graph to obtain an initial business plan; adding rule-encoded metadata to the initial business plan to obtain context-embedded business plan data; performing rule-mixed retrieval on the context-embedded business plan data, and evaluating the business rule matching degree of the decomposition rules obtained from the rule-mixed retrieval to obtain a rule matching degree score; wherein, the rule-mixed retrieval includes: BM25 matching and semantic vector retrieval; and setting decomposition rules with rule matching degree scores higher than a preset threshold as the initial decomposition scheme.

[0011] In one implementation of this application, the initial decomposition scheme is subject to task compliance verification to determine the business planning decomposition scheme. Specifically, this includes: performing hierarchical association verification on the initial decomposition scheme to determine the weights of first-level and second-level tasks; performing task integrity checks on the initial decomposition scheme to determine the weights of third-level tasks; performing key element matching on the weights of first-level, second-level, and third-level tasks to determine the performance of the decomposition scheme; and, if the performance of the decomposition scheme reaches the pre-approval state, determining the business planning decomposition scheme through milestone time sequence coherence analysis.

[0012] In one implementation of this application, after verifying the task compliance of the initial decomposition scheme to determine the business planning decomposition scheme, the method further includes: inputting the business planning decomposition scheme into a knowledge graph and setting the input business planning decomposition scheme as the basic data for the department's annual assessment.

[0013] Secondly, embodiments of this application also provide a business planning intelligent decomposition device based on large model fine-tuning, characterized in that the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: acquire comprehensive business data and perform scenario-based general logic analysis on the comprehensive business data to determine a pre-fine-tuning baseline model; based on the pre-fine-tuning baseline model, obtain real-time fine-tuning streaming data through directional fine-tuning of feature masks; perform model dynamic hierarchical adjustment analysis on the real-time fine-tuning streaming data to determine a business dynamic planning knowledge graph; obtain an initial decomposition scheme through context-enhanced business planning decomposition based on the business dynamic planning knowledge graph; and perform task compliance verification on the initial decomposition scheme to determine a business planning decomposition scheme.

[0014] Thirdly, embodiments of this application also provide a non-volatile computer storage medium for intelligent decomposition of business planning based on large model fine-tuning, storing computer-executable instructions. The computer-executable instructions are characterized by: acquiring comprehensive business data and performing scenario-based general logic analysis on the comprehensive business data to determine a pre-fine-tuning baseline model; obtaining real-time fine-tuning streaming data based on the pre-fine-tuning baseline model through directional fine-tuning using feature masks; performing dynamic hierarchical adjustment analysis on the real-time fine-tuning streaming data to determine a business dynamic planning knowledge graph; obtaining an initial decomposition scheme based on the business dynamic planning knowledge graph through context-enhanced business planning decomposition using context retrieval; and performing task compliance verification on the initial decomposition scheme to determine a business planning decomposition scheme.

[0015] This application provides a business planning intelligent decomposition method, device, and medium based on large model fine-tuning. Through pre-fine-tuning, special adjustment, online adjustment, and hierarchical triggering of model health self-optimization, it solves the technical problems of insufficient adaptability and non-standard verification in the prior art of business planning decomposition. It realizes adaptive execution of some business decomposition operations and improves the adaptability and verification standardization of some business planning decompositions. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1 A flowchart of a business planning intelligent decomposition method based on large model fine-tuning provided in this application embodiment;

[0018] Figure 2 This is a schematic diagram of the internal structure of a business planning intelligent decomposition device based on large model fine-tuning, provided as an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] This application provides a business planning intelligent decomposition method, device, and medium based on large model fine-tuning. Through pre-fine-tuning, special adjustment, online adjustment, and hierarchical triggering of model health self-optimization, it solves the technical problems of insufficient adaptability and non-standard verification in the prior art of business planning decomposition. It realizes adaptive execution of some business decomposition operations and improves the adaptability and verification standardization of some business planning decompositions.

[0021] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0022] Figure 1 A flowchart illustrating a business planning intelligent decomposition method based on large model fine-tuning, provided as an embodiment of this application. Figure 1 As shown in the figure, the intelligent decomposition method for business planning based on large model fine-tuning provided in this application embodiment specifically includes the following steps:

[0023] Step 101: Obtain comprehensive business data and perform scenario-based general logic analysis on the comprehensive business data to determine the pre-fine-tuned baseline model.

[0024] For example, fine-tuning of large models is a prerequisite for multi-agent systems to meet corresponding functions. In the prior art, fine-tuning of large models is usually a single adjustment or multiple unrelated adjustments, which cannot meet the data correlation of business data scenarios. This application determines the pre-fine-tuning baseline model by performing scenario-based general logic analysis on comprehensive business data, and realizes the construction of the pre-fine-tuning baseline model, providing a fine-tuning baseline for fine-tuning large models.

[0025] Specifically, scenario-based general logic analysis is performed on comprehensive business data to determine the pre-fine-tuned baseline model. This includes: analyzing the comprehensive business data for business management terminology to identify business management keywords; constructing a scenario-based dataset based on these keywords, and freezing the basic parameters of the scenario-based dataset through lightweight adjustments to obtain the model training range; determining the pre-fine-tuned model through joint optimization of loss terms based on the model training range; wherein, joint optimization of loss terms includes: cross-entropy loss optimization and rule matching loss optimization; and evaluating the pre-fine-tuned model's baseline to determine the pre-fine-tuned baseline model.

[0026] In one embodiment, the pre-tuning is not aimed at department-specific issues, but rather at enabling general basic models (such as GPT-4, LLaMA 3, etc.) to quickly grasp the three core objectives of industrial software business management scenarios: unified semantic understanding, mastery of basic rule logic, and construction of basic business relationships.

[0027] First, the core of constructing a scenario-based general dataset is to select data that is "companion-wide applicable, free of departmental sensitive information, and supports basic logic." This includes three categories: First, a general business terminology library (derived from a knowledge graph), containing basic terms and definitions for industrial software business management; second, a cross-departmental basic rule library (derived from a clause rule library), which filters out general rules that all departments must follow and eliminates department-specific rules; and third, a historical general case library (derived from historical summary data), which extracts decomposed cases that are recognized across departments, eliminates cases containing departmental sensitive data, and labels each case with "goal-task level-weight allocation-compliance explanation."

[0028] It is important to note that the dataset size should be controlled between 2,000 and 3,000 records to avoid data overload leading to poor generalization and to ensure the data's universality and accuracy.

[0029] Then, LoRA (Low-Rank Adaptation) lightweight fine-tuning is used to freeze more than 95% of the parameters of the base model and train only the general logic adaptation layer (the newly added 2-3 Transformer layers) to avoid overtraining and causing the model to forget its general language capabilities.

[0030] The loss function combines cross-entropy loss with rule matching loss, and monitors the accuracy of term understanding and the compliance rate of basic rules in real time. Training stops when both metrics are stable at ≥90%.

[0031] Finally, the pre-fine-tuning baseline assessment evaluates three core evaluation indicators to ensure that the pre-fine-tuning effect meets the standards. These core evaluation indicators include:

[0032] Terminology alignment rate: Input industrial software business terms, and the model output definitions should match the company's standard terminology library by ≥95%;

[0033] Basic rule compliance rate: For a given general scenario, the proportion of task weights, milestone times, etc. generated by the model that comply with the basic rules is ≥90%.

[0034] General link consistency refers to the "business-task" basic link that is decomposed multiple times in the model with the same input business goal.

[0035] Step 102: Based on the pre-fine-tuned baseline model, obtain real-time fine-tuned streaming data through directional fine-tuning of feature masks.

[0036] For example, this application addresses the common cross-departmental data privacy issues in existing technologies by adopting a federated fine-tuning framework. Through targeted fine-tuning of feature masks, it enables each department to complete data preprocessing and model fine-tuning locally, thereby improving the accuracy of business-oriented adjustments and data security.

[0037] Specifically, based on the pre-fine-tuned baseline model, real-time fine-tuned streaming data is obtained through targeted fine-tuning of feature masks. This includes: acquiring non-public data from departments and performing local masking on the non-public data to determine the anonymized departmental data; fine-tuning the anonymized departmental data based on the pre-fine-tuned baseline model to obtain specialized fine-tuning agents; wherein, the specialized fine-tuning agents include: decomposition agents, tracking agents, and summarizing agents; and recording the gradient aggregation process corresponding to the specialized fine-tuning agents to obtain real-time fine-tuned streaming data.

[0038] Furthermore, local masking is performed on the non-public data of departments to identify the data to be anonymized. Specifically, this includes: calculating the local model gradient of the non-public data of departments, and performing local masking on the local model gradient to obtain mask gradient data; aggregating the mask gradient data into model update gradient, and updating the global model based on the model update gradient to identify the data to be anonymized.

[0039] In one embodiment, based on the pre-tuned model, a federated learning framework is used to perform specific fine-tuning on three agents (decomposition agent, tracking agent, and summarizing agent).

[0040] The decomposition agent is used to label the correlation scores of the decomposition chains provided locally by the R&D department; the tracking agent is used for historical progress reporting and verification records provided by departments such as testing and project management; and the summarizing agent is used to summarize the annual / quarterly reports of some business operations.

[0041] First, data from each department is stored on local servers. The central server coordinates the federated learning process, employing a feature masking mechanism to protect sensitive information from each department. Key parameters for each agent are fine-tuned as follows: the decomposition agent uses LoRA with a learning rate of 5e-5, focusing on the attention layer related to task decomposition logic; the tracking agent uses LoRA with a learning rate of 3e-5, additionally incorporating resource load-lag risk prediction capabilities based on gradient boosting trees; and the summarizing agent uses full-parameter fine-tuning with a learning rate of 2e-5 to better learn complex cross-dimensional inductive reasoning.

[0042] Step 103: Perform dynamic hierarchical adjustment analysis on the real-time fine-tuning streaming data to determine the business dynamic planning knowledge graph.

[0043] For example, this application solves the problems of lag and resource waste in fixed-cycle online model adjustment in the prior art by performing dynamic hierarchical adjustment analysis on real-time fine-tuning streaming data to determine the business dynamic planning knowledge graph, thereby improving the analysis efficiency of business planning decomposition.

[0044] Specifically, the system performs dynamic hierarchical adjustment analysis on real-time fine-tuning streaming data to determine the business dynamic planning knowledge graph. This includes: determining the model warning level based on real-time fine-tuning streaming data through multi-indicator model health monitoring; the monitoring indicators for multi-indicator model health monitoring include: output compliance rate, user correction rate, and environmental change rate; online model fine-tuning triggered by the model warning level to obtain the fine-tuned model version; verifying the fine-tuning effect of the fine-tuned model version; if the verification result is successful, automatically saving the fine-tuned model version and updating the dynamic scenario memory to determine the business dynamic planning knowledge graph; if the verification result is unsuccessful, automatically reverting to the model version before fine-tuning and re-analyzing the reasons for the warning.

[0045] In one embodiment, model health monitors are deployed for three agents to collect three types of core indicators in real time and compare them with preset thresholds (set based on the needs of industrial software strategic management scenarios) to automatically determine whether to trigger fine-tuning.

[0046] If any indicator triggers the threshold, the monitor will automatically generate an early warning report and indicate the cause of the anomaly; if multiple indicators overlap, the early warning level will be automatically upgraded to shorten the response time.

[0047] First, online fine-tuning is triggered in a tiered manner. Based on the alert level, different levels of fine-tuning schemes are automatically invoked. For mild alerts, incremental data updates are used, outputting only the relevant fine-tuning parameters for slightly lower compliance rates and no major strategic or rule changes.

[0048] Furthermore, for moderate alerts, fine-tuning of the adaptation layer parameters is required. Only the department-specific adaptation layer of the model is trained, while the general logic layer in the pre-fine-tuning phase and the core parameters of the department in the specific tuning phase are frozen to prevent the model from forgetting. Through RLHF (Human Feedback Reinforcement Learning), the corrective opinions of business experts are transformed into reward signals to guide the model to optimize its decisions.

[0049] Furthermore, the severe warning adopts full dynamic fine-tuning, integrates historical effective data and newly added scenario data, constructs a temporary fine-tuning dataset, performs full fine-tuning on the model, updates the relevant parameters of the causal inference part, and simultaneously optimizes the decision tracing logic.

[0050] Finally, fine-tuning and verification were completed in the test environment first, and then the online model was gradually replaced. After the optimization was executed, the optimization effect was verified to ensure that the fine-tuning was effective and had no side effects.

[0051] If the verification passes, the fine-tuned model version is automatically saved, and the dynamic memory is updated to determine the business dynamic planning knowledge graph. Otherwise, the system automatically reverts to the model version before fine-tuning and re-analyzes the reasons for the warning.

[0052] Step 104: Based on the business dynamic planning knowledge graph, obtain the initial decomposition scheme by enhancing the business planning decomposition through contextual retrieval.

[0053] Specifically, based on the business dynamic planning knowledge graph, an initial decomposition scheme is obtained through context-enhanced business planning decomposition, including: performing effective task retrieval on the business dynamic planning knowledge graph to obtain an initial business scheme; adding rule-encoded metadata to the initial business scheme to obtain context-embedded business scheme data; performing rule-mixed retrieval on the context-embedded business scheme data, and evaluating the business rule matching degree of the decomposition rules obtained from the rule-mixed retrieval to obtain a rule matching degree score; wherein, rule-mixed retrieval includes: BM25 matching and semantic vector retrieval; and setting decomposition rules with rule matching degree scores higher than a preset threshold as the initial decomposition scheme.

[0054] In one embodiment, after receiving information, the intelligent agent calls the knowledge graph API to retrieve relevant business elements and generate and display relevant documents.

[0055] Then, historical tasks and rule bases in the knowledge graph are retrieved using context-enhanced RAG (Retrieved Algorithm), and preliminary solutions are generated through reasoning. Automatic verification of hierarchical associations, element matching, and weight summation are performed on tasks at different levels.

[0056] Furthermore, for tasks derived from the decomposition of entirely new business operations, the agent searches for decomposition logic for similar tasks, extrapolates it, and prompts for manual calibration.

[0057] Step 105: Verify the task compliance of the initial decomposition scheme to determine the business planning decomposition scheme.

[0058] For example, in the prior art, after business decomposition, the corresponding task objectives have weak semantic association and the task weight allocation has problems such as the weight being out of sync with business priority and time coverage conflict. This application determines the business planning decomposition scheme by verifying the task compliance of the initial decomposition scheme, thereby achieving accurate analysis of the business planning decomposition scheme and improving the compliance of the business planning decomposition scheme.

[0059] Specifically, the initial decomposition plan undergoes task compliance verification to determine the business planning decomposition plan. This includes: performing hierarchical correlation verification on the initial decomposition plan to determine the weights of first-level and second-level tasks; performing task integrity checks on the initial decomposition plan to determine the weights of third-level tasks; matching key elements of the first-level, second-level, and third-level task weights to determine the performance of the decomposition plan; and, if the performance of the decomposition plan reaches the pre-approval stage, determining the business planning decomposition plan through milestone time sequence coherence analysis.

[0060] Furthermore, after verifying the task compliance of the initial decomposition scheme to determine the business planning decomposition scheme, the method also includes: inputting the business planning decomposition scheme into a knowledge graph and setting the input business planning decomposition scheme as the basic data for the department's annual assessment.

[0061] In one embodiment, the agent can use Retrieval Enhanced Generation (RAG) technology to establish semantic connections between strategic goals and task implementation, and traverse the domain knowledge graph to extract business objectives and technical requirements.

[0062] By using text similarity algorithms derived from natural language processing, the semantic fit between the new task objectives and the overall business plan is compared.

[0063] There are two optimization methods: semantic enhancement and element completion. To address the above issues, semantic enhancement is used to add strategic keywords to the task description, and element completion is used to add element tags to the task. The association parameters of the knowledge graph are updated synchronously, and the element matching is verified twice to confirm compliance.

[0064] Next, compliance verification and optimization of weight allocation are performed. The agent extracts priority tags associated with tasks from the knowledge graph and calls the mapping rules for priority tags in the rule base. The agent retrieves historical valid cases through RAG and automatically overlays element weighting coefficients onto the base interval to generate a calibrated interval.

[0065] Subsequently, preliminary optimization was performed, automatically overlaying element weighting coefficients onto the base interval to determine the weights of the three-level tasks.

[0066] Finally, when the timing consistency of milestone nodes needs optimization, the milestone plan is made reasonable and feasible based on timing and resource constraints. After multiple rounds of revision and verification, the department head confirms the final draft, and the intelligent agent enters the solution into the knowledge graph and pushes it to the business tracking intelligent agent as a benchmark for subsequent monthly scoring and annual evaluation.

[0067] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide a business planning intelligent decomposition device based on large model fine-tuning, the structure of which is as follows: Figure 2 As shown.

[0068] Figure 2 This is a schematic diagram of the internal structure of a business planning intelligent decomposition device based on large model fine-tuning, provided as an embodiment of this application. Figure 2 As shown, the device includes:

[0069] At least one processor 201;

[0070] And a memory 202 that is communicatively connected to at least one processor;

[0071] The memory 202 stores instructions executable by at least one processor, which are executed by at least one processor 201 to enable at least one processor 201 to:

[0072] Acquire comprehensive business data and perform scenario-based general logical analysis on the comprehensive business data to determine the pre-fine-tuning baseline model; based on the pre-fine-tuning baseline model, obtain real-time fine-tuning streaming data through targeted fine-tuning of feature masks; perform dynamic hierarchical adjustment analysis on the real-time fine-tuning streaming data to determine the business dynamic planning knowledge graph; based on the business dynamic planning knowledge graph, obtain an initial decomposition scheme through contextual retrieval-enhanced business planning decomposition; and perform task compliance verification on the initial decomposition scheme to determine the business planning decomposition scheme.

[0073] Some embodiments of this application provide corresponding to Figure 1 A non-volatile computer storage medium for intelligent decomposition of business planning based on large model fine-tuning, storing computer-executable instructions, wherein the computer-executable instructions are configured as follows:

[0074] Acquire comprehensive business data and perform scenario-based general logical analysis on the comprehensive business data to determine the pre-fine-tuning baseline model; based on the pre-fine-tuning baseline model, obtain real-time fine-tuning streaming data through targeted fine-tuning of feature masks; perform dynamic hierarchical adjustment analysis on the real-time fine-tuning streaming data to determine the business dynamic planning knowledge graph; based on the business dynamic planning knowledge graph, obtain an initial decomposition scheme through contextual retrieval-enhanced business planning decomposition; and perform task compliance verification on the initial decomposition scheme to determine the business planning decomposition scheme.

[0075] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0076] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

Claims

1. A business planning intelligent decomposition method based on large model fine-tuning, characterized in that, The method includes: Acquire comprehensive business data and perform scenario-based general logic analysis on the comprehensive business data to determine the pre-fine-tuned baseline model; Based on the pre-fine-tuned baseline model, real-time fine-tuned streaming data is obtained through directional fine-tuning of the feature mask; The real-time fine-tuning streaming data is subjected to dynamic hierarchical adjustment analysis of the model to determine the business dynamic planning knowledge graph; Based on the business dynamic planning knowledge graph, an initial decomposition scheme is obtained through contextual retrieval-enhanced business planning decomposition. The initial decomposition scheme is subjected to task compliance verification in order to determine the business planning decomposition scheme; Based on the pre-fine-tuned baseline model, real-time fine-tuned streaming data is obtained through directional fine-tuning of feature masks, specifically including: Obtain non-public data from departments and perform local masking on the non-public data to identify the anonymized department data; Based on the pre-fine-tuned baseline model, the desensitized departmental data is fine-tuned and oriented to obtain a specialized fine-tuning agent; wherein, the specialized fine-tuning agent includes: a decomposition agent, a tracking agent, and a summarizing agent; Record the gradient aggregation process corresponding to the specialized fine-tuning agent to obtain the real-time fine-tuning streaming data; The real-time fine-tuning streaming data is subjected to dynamic hierarchical adjustment analysis of the model to determine the business dynamic planning knowledge graph, specifically including: Based on the real-time fine-tuning streaming data, the model's early warning level is determined through multi-indicator health monitoring; wherein, the monitoring indicators for multi-indicator health monitoring include: output compliance rate, user correction rate, and environmental change rate. The model is fine-tuned online by triggering graded warning levels to obtain a fine-tuned model version; The fine-tuned model version is verified for fine-tuning effect. If the verification result of the fine-tuning effect verification is passed, the fine-tuned model version is automatically saved and the dynamic scene memory is updated to determine the business dynamic planning knowledge graph. If the verification result of the fine-tuning effect verification is unsuccessful, the system will automatically revert to the model version before the fine-tuning and re-analyze the reasons for the warning.

2. The intelligent decomposition method for business planning based on large model fine-tuning according to claim 1, characterized in that, The comprehensive business data is subjected to scenario-based general logical analysis to determine the pre-fine-tuning baseline model, specifically including: The comprehensive business data is analyzed using business management terminology to identify key business management terms. Based on the aforementioned business management keywords, a scenario-based dataset is constructed, and the basic parameters of the scenario-based dataset are frozen after lightweight adjustments to obtain the model training range. Based on the training range of the model, a pre-fine-tuned model is determined through joint optimization of loss terms; wherein, the joint optimization of loss terms includes: cross-entropy loss optimization and rule matching loss optimization; A model baseline evaluation is performed on the pre-fine-tuned model to determine the pre-fine-tuned baseline model.

3. The intelligent decomposition method for business planning based on large model fine-tuning according to claim 1, characterized in that, Local masking is performed on the non-public data of the aforementioned departments to identify the data to be anonymized. Specifically, this includes: Calculate the local model gradient of the department's non-public data, and perform local masking processing on the local model gradient using department-specific data to obtain masked gradient data; The mask gradient data is aggregated into model update gradients, and the global model is updated based on the model update gradients to determine the data of the de-identified department.

4. The intelligent decomposition method for business planning based on large model fine-tuning according to claim 1, characterized in that, Based on the aforementioned business dynamic planning knowledge graph, an initial decomposition scheme is obtained through context-enhanced business planning decomposition, specifically including: The business dynamic planning knowledge graph is used to retrieve effective tasks from departments to obtain an initial business plan; Add rule-encoded metadata to the initial business plan to obtain context-embedded business plan data; The context-embedded business solution data is subjected to rule-based hybrid retrieval, and the decomposed rules obtained from the rule-based hybrid retrieval are evaluated for business rule matching degree to obtain a rule matching degree score; wherein, the rule-based hybrid retrieval includes: BM25 matching and semantic vector retrieval; The decomposition rules whose matching scores are higher than a preset threshold are set as the initial decomposition scheme.

5. The intelligent decomposition method for business planning based on large model fine-tuning according to claim 1, characterized in that, The initial decomposition scheme is subjected to task compliance verification to determine the business planning decomposition scheme, specifically including: The initial decomposition scheme is subjected to hierarchical correlation verification to determine the weights of the first-level tasks and the second-level tasks; The initial decomposition scheme is subjected to task integrity detection to determine the weights of the three levels of tasks; Key elements are matched to determine the performance of the decomposition scheme by assessing the weights of the first-level tasks, second-level tasks, and third-level tasks. Once the performance of the decomposition scheme reaches the preliminary review stage, the business planning decomposition scheme is determined through milestone time sequence coherence analysis.

6. The intelligent decomposition method for business planning based on large model fine-tuning according to claim 1, characterized in that, After verifying the task compliance of the initial decomposition scheme to determine the business planning decomposition scheme, the method further includes: The business planning decomposition scheme is entered into a knowledge graph, and the entered business planning decomposition scheme is set as the basic data for the department's annual evaluation.

7. A business planning intelligent decomposition device based on large model fine-tuning, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: Acquire comprehensive business data and perform scenario-based general logic analysis on the comprehensive business data to determine the pre-fine-tuned baseline model; Based on the pre-fine-tuned baseline model, real-time fine-tuned streaming data is obtained through directional fine-tuning of the feature mask; The real-time fine-tuning streaming data is subjected to dynamic hierarchical adjustment analysis of the model to determine the business dynamic planning knowledge graph; Based on the business dynamic planning knowledge graph, an initial decomposition scheme is obtained through contextual retrieval-enhanced business planning decomposition. The initial decomposition scheme is subjected to task compliance verification in order to determine the business planning decomposition scheme; Based on the pre-fine-tuned baseline model, real-time fine-tuned streaming data is obtained through directional fine-tuning of feature masks, specifically including: Obtain non-public data from departments and perform local masking on the non-public data to identify the anonymized department data; Based on the pre-fine-tuned baseline model, the desensitized departmental data is fine-tuned and oriented to obtain a specialized fine-tuning agent; wherein, the specialized fine-tuning agent includes: a decomposition agent, a tracking agent, and a summarizing agent; Record the gradient aggregation process corresponding to the specialized fine-tuning agent to obtain the real-time fine-tuning streaming data; The real-time fine-tuning streaming data is subjected to dynamic hierarchical adjustment analysis of the model to determine the business dynamic planning knowledge graph, specifically including: Based on the real-time fine-tuning streaming data, the model's early warning level is determined through multi-indicator health monitoring; wherein, the monitoring indicators for multi-indicator health monitoring include: output compliance rate, user correction rate, and environmental change rate. The model is fine-tuned online by triggering graded warning levels to obtain a fine-tuned model version; The fine-tuned model version is verified for fine-tuning effect. If the verification result of the fine-tuning effect verification is passed, the fine-tuned model version is automatically saved and the dynamic scene memory is updated to determine the business dynamic planning knowledge graph. If the verification result of the fine-tuning effect verification is unsuccessful, the system will automatically revert to the model version before the fine-tuning and re-analyze the reasons for the warning.

8. A non-volatile computer storage medium for intelligent decomposition of business planning based on large model fine-tuning, storing computer-executable instructions, characterized in that, The computer-executable instructions are set as follows: Acquire comprehensive business data and perform scenario-based general logic analysis on the comprehensive business data to determine the pre-fine-tuned baseline model; Based on the pre-fine-tuned baseline model, real-time fine-tuned streaming data is obtained through directional fine-tuning of the feature mask; The real-time fine-tuning streaming data is subjected to dynamic hierarchical adjustment analysis of the model to determine the business dynamic planning knowledge graph; Based on the business dynamic planning knowledge graph, an initial decomposition scheme is obtained through contextual retrieval-enhanced business planning decomposition. The initial decomposition scheme is subjected to task compliance verification in order to determine the business planning decomposition scheme; Based on the pre-fine-tuned baseline model, real-time fine-tuned streaming data is obtained through directional fine-tuning of feature masks, specifically including: Obtain non-public data from departments and perform local masking on the non-public data to identify the anonymized department data; Based on the pre-fine-tuned baseline model, the desensitized departmental data is fine-tuned and oriented to obtain a specialized fine-tuning agent; wherein, the specialized fine-tuning agent includes: a decomposition agent, a tracking agent, and a summarizing agent; Record the gradient aggregation process corresponding to the specialized fine-tuning agent to obtain the real-time fine-tuning streaming data; The real-time fine-tuning streaming data is subjected to dynamic hierarchical adjustment analysis of the model to determine the business dynamic planning knowledge graph, specifically including: Based on the real-time fine-tuning streaming data, the model's early warning level is determined through multi-indicator health monitoring; wherein, the monitoring indicators for multi-indicator health monitoring include: output compliance rate, user correction rate, and environmental change rate. The model is fine-tuned online by triggering graded warning levels to obtain a fine-tuned model version; The fine-tuned model version is verified for fine-tuning effect. If the verification result of the fine-tuning effect verification is passed, the fine-tuned model version is automatically saved and the dynamic scene memory is updated to determine the business dynamic planning knowledge graph. If the verification result of the fine-tuning effect verification is unsuccessful, the system will automatically revert to the model version before the fine-tuning and re-analyze the reasons for the warning.

Images