An agent handling task dataset generation method based on structured prompt words

Abstract

The application discloses an agent disposal task dataset generation method based on a structured prompt word, forms a standardized requirement list by sorting out core elements affecting agent task allocation, constructs a structured micro-scene template containing a fixed framework and global rules, a parameterized variable pool and random instructions, strategy and logic constraint embedding, and a standardized output format forced structure, and analyzes the template by using a large language model deployed locally and generates a standardized question and answer pair data containing a complete reasoning chain. The method improves the civil scene adaptability, data quality controllability and instruction design accuracy of the dataset, and the construction method is standardized and reproducible, and is suitable for the autonomous and intelligent allocation of civil cluster agent disposal tasks.

Description

Technical Field

[0001] The present invention relates to the technical fields of computer application technology and artificial intelligence data construction technology, and specifically relates to a method for generating an intelligent agent disposal task dataset based on structured prompt words. Background Art

[0002] With the continuous iterative upgrade of intelligent agent technology, civilian cluster intelligent agents are increasingly widely used in fields such as public safety, emergency rescue, patrol inspection, and logistics delivery. Among them, the intelligent agent disposal task refers to a full-process operation task such as target detection, disaster situation verification, material transfer, hidden danger disposal, and collaborative return in a complex operation environment. Relying on the advantages of multi-aircraft collaborative networking and distributed linkage operations, civilian cluster intelligent agents have become the key carriers of various core operations. The scientific rationality of its disposal task allocation mechanism directly determines the overall operation efficiency and task execution success rate of the cluster.

[0003] In recent years, the implementation and application of large model technology in the field of intelligent decision-making have provided a new technical direction for intelligent agent task allocation. Relying on large models to achieve autonomous and intelligent allocation of civilian cluster intelligent agent disposal tasks has become the key to improving the automation level of airspace operations. However, the high-quality instruction fine-tuning dataset required in the large model supervised fine-tuning (SFT) link is the core basis for ensuring the fine-tuning accuracy of the model and optimizing the decision-making logic. At present, the construction technology of large model fine-tuning datasets mostly focuses on the field of general natural language processing, and a dedicated construction system adapted to the civilian cluster intelligent agent disposal task allocation scenario has not yet been formed.

[0004] The existing technical solutions have the following obvious shortcomings: First, the scene adaptability is insufficient, and the scene characteristics of intelligent agent cluster collaborative operations and the characteristics of dynamic and complex environments are not deeply integrated. The task process design is limited to a single link, lacking a full-link closed loop; Second, the data quality control is missing, and there are problems such as data redundancy, noise doping, and inconsistent formats, interfering with the model learning logic; Third, the instruction design is weakly targeted, lacking hard constraint conditions and quantitative evaluation criteria in the task allocation link, resulting in the disconnection between the model decision output and the actual combat requirements. Therefore, there is an urgent need to develop a method for constructing an instruction fine-tuning dataset that fits the actual combat scenario, has controllable quality throughout the process, and is highly targeted at the scenario. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for generating an intelligent agent disposal task dataset based on structured prompt words to solve the problems of poor scene adaptability, lack of quality control, and insufficient instruction design targeting in the intelligent agent task allocation dataset in the prior art.

[0006] To achieve the above purpose, the present invention provides the following technical solutions: A method for generating datasets for intelligent agent handling tasks based on structured prompts includes: Identify the core factors affecting task allocation for intelligent agents and create a standardized requirements list. Based on a standardized requirements list, a structured micro-scene template is constructed, which includes a fixed framework and global rules, a parameterized variable pool and random instructions, embedded strategies and logical constraints, and a standardized output format. The structured micro-scenario template is parsed using a locally deployed large language model, and a parameterized variable pool is called to generate the job status. The embedded constraint rules are combined to perform planning and reasoning, and standardized question-answer pairs containing a complete reasoning chain are output.

[0007] Furthermore, the core elements influencing task allocation for intelligent agents include: Environmental parameters, objective parameters, agent parameters, and constraint and policy parameters; The environmental parameters include meteorological parameters and operational interference intensity; The target parameters include target attributes, number of targets, coordinate precision, and task priority; The intelligent agent parameters include machine type, deployment location, performance threshold, and quantity configuration; The constraints and strategy parameters include aircraft type operating constraints, performance evaluation rules, action strategies, and operating methods.

[0008] Furthermore, the structured micro-scene template includes: Establish a fixed framework and global rules to uniformly define the operational airspace, basic performance parameters of the agent, and principles for global operations; Construct a parameterized variable pool, define the value range and generation logic of environment variables, target variables, agent variables and policy variables, and set random instruction requirements; Knowledge from the civilian operations domain is transformed into hard constraints and written into templates to form a strategy and logic constraint embedded module; Set a standardized output format and force all samples to be organized into a three-part structure based on the problem description, analysis process, and answer.

[0009] Furthermore, the environmental variables include weather, temperature, humidity, takeoff wind direction, wind speed, and intensity of operational interference; The target variables include target type, number of target categories, number of targets per category, coordinate accuracy, and target priority; Agent variables include the number of bases, the number of agents per base, the combination of machine types, and the base coordinates; Strategic variables include action strategies, work methods, and work performance thresholds.

[0010] Furthermore, the strategy and logic constraint embedding module includes device model adaptation constraints and strategy execution constraints; Among them, the model-adaptive constraint-limited coordinate delivery intelligent agent is only adapted to precise coordinate targets, the comprehensive rescue intelligent agent is adapted to fuzzy or composite coordinate targets, and the photoelectric inspection intelligent agent is only responsible for reconnaissance and performance evaluation. The strategy execution constraints stipulate that the low-cost strategy prioritizes coordinate delivery agents, while the high-efficiency strategy prioritizes scheduling agents closest to the target, and both strategies execute a closed loop from task execution and evaluation to follow-up attack.

[0011] Furthermore, the standardized output format is set, forcing all samples to be organized into a three-part structure based on the problem description, analysis process, and answer, including: The problem description section includes the task name, action strategy, operation method, target attributes and coordinate accuracy, weather conditions, agent deployment information, and detailed operation objectives; The analysis process segment sequentially completes the total number of targets, priority ranking, aircraft type compatibility analysis, base site selection calculation, and quantity allocation derivation. The answer section clearly outlines the step-by-step action plan, including agent dispatch information, operational steps, performance evaluation process, and cyclical replenishment rules.

[0012] Furthermore, in the automated data generation based on the local large model, the situation simulation generation process is as follows: the local large model parses the structured micro-scenario template instructions, calls the parameterized variable pool to randomly generate a logically consistent complete operational situation, verifies the compliance of the parameters, and outputs standardized situation text according to the problem description format.

[0013] Furthermore, the process involves using a locally deployed large language model to parse the structured micro-scene template, calling a parameterized variable pool to generate a work situation, combining embedded constraint rules for planning and reasoning, and outputting standardized question-answer pairs containing a complete reasoning chain, including: Based on the generated operational situation, the model combines embedded constraint rules and priority logic to carry out resource allocation, path calculation and time sequence planning, deduce executable action plans and externalize the complete reasoning process.

[0014] Furthermore, the method also includes: Batch question-answer pairs are generated by iterating variable values ​​through templates. The model verifies the consistency of parameter logic in real time. After generation, invalid samples with incomplete format, contradictory parameters, and logical conflicts are automatically removed, while compliant and valid data are retained.

[0015] The present invention has the following beneficial effects: 1. Strong adaptability to civilian scenarios: Guided by the needs of civilian cluster intelligent agents for collaborative operation, it fully covers conventional and special scenarios. The instruction templates are derived from practical data, which solves the pain point of poor civilian adaptability of general datasets.

[0016] 2. Controllable data quality: Through full-process quality control and dual verification mechanisms, low-quality noise data is eliminated, improving the stability of decision-making.

[0017] 3. Precise and efficient instruction design: The structured instruction templates are designed to fit the semantic understanding logic of large models, guiding the models to quickly learn allocation rules and improve decision-making accuracy.

[0018] 4. The construction method is standardized and reproducible: the steps are clear and the standards are well-defined, with strong repeatability and scalability, making it easy to implement and promote. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of a method for generating datasets for intelligent agent handling tasks based on structured prompt words, according to the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments use a deep emergency rescue scenario as an example to illustrate the specific implementation of the present invention in detail.

[0021] In this application, "intelligent agent handling task" specifically refers to the full-process operation tasks carried out by civilian swarm intelligent agents in complex operating environments, targeting public safety, emergency rescue, inspection and patrol, logistics delivery and other scenarios, including target detection, disaster verification, material transfer, hidden danger handling and collaborative return.

[0022] In this application, "structured micro-scenario template" refers to a template that, for a specific operational scenario (such as in-depth emergency rescue), structures the core elements affecting task allocation according to a fixed framework, rules, variable pool, and constraint logic, and is used to guide the generation of standardized data by a large model.

[0023] In this application, "local large model" refers to a large language model deployed on a local server or computing device for parsing instructions, generating text, and reasoning logic, ensuring the privacy and controllability of data processing.

[0024] Existing techniques for constructing instruction fine-tuning datasets are mostly focused on general natural language processing, and a dedicated construction system adapted to civilian swarm intelligence agent task allocation scenarios has not yet been formed. Existing solutions suffer from insufficient scenario adaptability, failing to deeply integrate the scenario characteristics of intelligent agent swarm collaborative operations; lack of data quality control, resulting in noise redundancy; and weak instruction design targeting, lacking hard constraints and quantitative evaluation standards. These shortcomings lead to large models being unable to accurately align with the core intent of task allocation in real-world scenarios.

[0025] In this application, the intelligent agent can be a drone.

[0026] Example 1 This embodiment provides a method for generating datasets for intelligent agent handling tasks based on structured prompts. This method aims to construct a standardized instruction fine-tuning dataset that closely matches actual civilian operational scenarios, ensures full control over data quality, and provides precise instruction adaptation. For example... Figure 1 As shown, the method specifically includes the following steps: Step S1: Identify the core elements that affect task allocation for intelligent agents and form a standardized requirements list.

[0027] This provides a standardized data foundation for subsequent template design, ensuring that the source of data generation meets practical needs and solving the problem of insufficient scenario adaptability in existing technologies.

[0028] In one embodiment of this application, the core elements affecting the task allocation of the intelligent agent include: environmental parameters, target parameters, intelligent agent parameters, and constraint and policy parameters.

[0029] The environmental parameters include meteorological parameters and operational interference intensity. Specifically, meteorological parameters cover weather type, ambient temperature, relative humidity, takeoff wind direction, wind speed, etc.; operational interference intensity is dynamically linked to target protection or hazard level. For example, strong interference corresponds to high-risk targets such as hazardous chemical storage areas and important power plants, while weak interference corresponds to low-risk targets such as ordinary inspection points.

[0030] The target parameters include target attributes, number of targets, coordinate precision, and task priority. Specifically, target attributes are divided into fixed targets (such as trapped personnel and disaster sites) and mobile targets (such as emergency supplies and rescue vehicles); target priorities are preset according to actual combat needs, for example, trapped personnel have the highest priority.

[0031] The intelligent agent parameters include aircraft type, deployment location, performance threshold, and quantity configuration. Specifically, aircraft types include comprehensive rescue type, coordinate delivery type, and photoelectric inspection type, etc., and different aircraft types have different performance thresholds (such as flight speed and endurance) and operational capabilities.

[0032] The constraints and strategy parameters include aircraft type operational constraints, performance evaluation rules, action strategies, and operational methods. Specifically, aircraft type operational constraints limit the operational scope of a particular aircraft type; action strategies include low-cost operational strategies, high-efficiency operational strategies, etc.

[0033] By employing the above technical solutions and refining the four dimensions of parameters, we can comprehensively cover the key variables in real-world scenarios, providing detailed element support for building high-quality datasets and significantly improving the scenario adaptability of the datasets.

[0034] Step S2: Based on the standardized requirements list, construct a structured micro-scene template that includes a fixed framework and global rules, a parameterized variable pool and random instructions, embedded strategies and logical constraints, and a standardized output format.

[0035] By executing step S2, domain knowledge can be structured and modeled, providing parsable and logically rigorous input templates for large models, thus solving the problems of weak instruction design and lack of constraint system in existing technologies.

[0036] In one embodiment of this application, the structured micro-scene template includes: setting a fixed framework and global rules to uniformly define the operational spatial scope, basic performance parameters of the agent, and global operational principles; constructing a parameterized variable pool to define the value range and generation logic of environmental variables, target variables, agent variables, and policy variables, and setting random instruction requirements; transforming civilian operational domain knowledge into hard constraints and writing it into the template to form a policy and logic constraint embedding module; setting a standardized output format to force all samples to construct a three-part structure organization based on problem description, analysis process, and answer.

[0037] Thus, through modular template design, comprehensive control over the data generation process is achieved, ensuring the standardization and logic of the generated data.

[0038] In one embodiment of this application, the environmental variables include weather, temperature, humidity, takeoff wind direction, wind speed, and operational interference intensity; the target variables include target type, number of target types, number of single-type targets, coordinate accuracy, and target priority; the agent variables include number of bases, number of agents per base, aircraft type combination, and base coordinates; and the strategy variables include action strategy, operation mode, and operation efficiency threshold.

[0039] Specifically, when constructing the parameterized variable pool, the value range and generation logic of each variable are defined. For example, in environmental variables, the weather is randomly selected between sunny and rainy days; the temperature is randomly generated as an integer within the range of 15-45℃. The randomization instruction requires the local large model to generate variables according to the value range and enforces the verification of variable logic consistency, such as humidity not being lower than 60% on rainy days, and disabling coordinate projection agents for fuzzy coordinates.

[0040] By using a parameterized variable pool, the randomness and diversity of the generated data are ensured, while logical verification ensures the rationality of the data and avoids the generation of invalid data.

[0041] In one embodiment of this application, the strategy and logic constraint embedding module includes device model adaptation constraints and strategy execution constraints.

[0042] Among them, the model-adaptive constraint-limited coordinate delivery intelligent agent is only adapted to precise coordinate targets, the comprehensive rescue intelligent agent is adapted to fuzzy or composite coordinate targets, and the photoelectric inspection intelligent agent is only responsible for reconnaissance and effectiveness evaluation.

[0043] The strategy execution constraints stipulate that the low-cost strategy prioritizes the use of coordinate delivery agents, and the high-efficiency strategy prioritizes scheduling the agent closest to the target, and both execute a closed loop from operation, evaluation to follow-up attack.

[0044] Specifically, knowledge from the civilian operational domain is transformed into hard constraints and written into the template. For example, the model adaptation constraint strictly matches the model according to the coordinate accuracy, and the one-time intelligent agent cannot be repeatedly scheduled; the policy execution constraint enforces the "job-evaluation-replacement" closed loop to ensure the complete execution of the task.

[0045] By embedding hard constraints, the generated data is ensured to fit the logic of real-world applications, preventing the model from learning incorrect decision-making paths and enhancing the practical guidance value of the dataset.

[0046] In one embodiment of this application, the standardized output format is set, forcing all samples to be organized into a three-part structure based on problem description, analysis process, and answer. The problem description section covers the task name, action strategy, operation method, target attributes and coordinate accuracy, weather conditions, agent deployment information, and operation target details. The analysis process section sequentially completes the total number of targets, priority ranking, model adaptation analysis, base site selection calculation, and quantity allocation derivation. The answer section clarifies the step-by-step action plan, including agent dispatch information, operation steps, performance evaluation process, and cyclical replenishment rules.

[0047] Specifically, the problem description fully describes the current operational situation; the analysis process externalizes a complete reasoning chain, such as first statistically analyzing and sorting the targets, and then matching the machine models according to coordinate precision; the answer clearly outlines a step-by-step action plan to ensure that the plan is implementable and reproducible.

[0048] By employing the above technical solution and a three-stage structure, not only are standardized input and output formats provided, but also the externalized reasoning chain guides the large model to learn the logical reasoning process, significantly improving the fine-tuning effect and the interpretability of model decisions.

[0049] Step S3: Use the locally deployed large language model to parse the structured micro-scene template, call the parameterized variable pool to generate the job status, combine the embedded constraint rules to perform planning and reasoning, and output standardized question-answer pairs containing a complete reasoning chain.

[0050] This step enables a deep integration of domain expert knowledge and large model reasoning capabilities, eliminating the need for manual annotation and efficiently producing high-quality training samples. It solves the problems of lack of data quality control and noise redundancy in existing technologies.

[0051] In one embodiment of this application, the process of situation simulation generation in the automated data generation based on the local large model is as follows: the local large model parses the structured micro-scenario template instructions, calls the parameterized variable pool to randomly generate a logically consistent complete operational situation, verifies the compliance of the parameters, and outputs standardized situation text according to the problem description format.

[0052] Specifically, the local large model parses template instructions, calls the variable pool to randomly generate the situation, strictly verifies the compliance of parameters (such as the upper limit of base deployment and meteorological parameter matching), and outputs standardized situation text in a three-part problem description format.

[0053] By employing the above technical solutions, the uniqueness and professionalism of the samples are ensured, the generation of situational awareness is automated, and the efficiency of dataset generation is significantly improved.

[0054] In one embodiment of this application, the process of parsing the structured micro-scene template using a locally deployed large language model, generating a work situation by calling a parameterized variable pool, performing planning and reasoning in conjunction with embedded constraint rules, and outputting standardized question-answer pairs containing a complete reasoning chain includes: the model, based on the generated work situation, combined with embedded constraint rules and priority logic, performing resource allocation, path calculation and time-series planning, deriving executable action plans and externalizing the complete reasoning process.

[0055] Specifically, the model combines embedded constraint rules and priority logic to carry out resource allocation, path calculation, and time-series planning. For example, in a deep emergency rescue scenario, targets are first counted and sorted, then aircraft models are matched according to coordinate accuracy, and location allocation is combined with action strategies. Simultaneously, an efficiency evaluation agent is configured to ultimately form a step-by-step operation plan.

[0056] By employing the above technical solutions, the generated data contains complete decision-making logic, which helps to improve the decision-making accuracy and logical consistency of large models.

[0057] In one embodiment of this application, the method further includes: generating batch question-answer pairs by iterating variable values ​​through templates, verifying the consistency of parameter logic in real time, and automatically removing invalid samples with incomplete format, contradictory parameters, or logical conflicts after generation, while retaining compliant and valid data.

[0058] Specifically, batch generation is achieved through template iteration and variable value taking, and the model verifies the consistency of parameter logic in real time. After generation, invalid samples with incomplete format, contradictory parameters, or logical conflicts are automatically removed, while compliant and valid data are retained.

[0059] By employing the above technical solutions, low-quality noise data is effectively eliminated, data quality is made fully controllable, large model fine-tuning errors are reduced, and decision stability is improved.

[0060] Example 2 like Figure 1 As shown, this invention provides a method for generating an intelligent agent handling task dataset based on structured prompts, comprising the following steps: Step 1: Define actual operational needs and classify scenarios For in-depth emergency rescue missions, the core elements affecting task allocation by intelligent agents were identified, and a standardized requirements list was formed. This mission specifically refers to long-range, cross-regional, and end-to-end emergency response operations conducted by civilian swarm intelligent agents in remote, isolated, and terrain-complex in-depth rescue areas.

[0061] The core elements specifically include: 1.1 Environmental Parameters Meteorological parameters: Weather type is divided into sunny and rainy; ambient temperature is set to 15℃ < x < 45℃; relative humidity is set according to the scene, 60% < x < 90% for rainy days and 20% < x < 70% for sunny days; takeoff wind direction is divided into headwind, tailwind, and crosswind; wind speed is set according to the mode, 0m / s < x < 15m / s in headwind mode and 0m / s < x < 7m / s in tailwind / crosswind mode.

[0062] Operational interference intensity: divided into strong interference (corresponding to high-risk targets such as hazardous chemical storage areas and important power plants) and weak interference (corresponding to low-risk targets such as ordinary inspection points).

[0063] 1.2 Target Parameters Target attributes are divided into fixed (such as trapped personnel, disaster sites) and mobile (such as emergency supplies, rescue vehicles); the number of targets is divided into single / multiple; the coordinate accuracy is divided into precise / fuzzy; the task priority from high to low is as follows: trapped personnel > secondary hazards > emergency supplies, emergency equipment > disaster sites > work teams > rescue vehicles > communication relays > inspection vehicles.

[0064] 1.3 Agent Parameters The aircraft types include: integrated rescue intelligent agents (high cost, single-use, with reconnaissance, assessment, and precision delivery functions), coordinate delivery intelligent agents (low cost, single-use, delivery-only), and electro-optical inspection intelligent agents (recoverable and reusable, no delivery function, used only for reconnaissance and effectiveness assessment). All three types of aircraft have a uniform flight speed of 200 km / h and an endurance of 4 hours. The deployment plan is as follows: five fixed operational bases, randomly distributed on the west side of a 400 km × 400 km operational airspace.

[0065] 1.4 Constraints and Strategy Parameters Model-specific operational constraints: Coordinate delivery type intelligent agents are only compatible with targets with precise coordinates; integrated rescue type intelligent agents can be compatible with targets with fuzzy / composite coordinates; photoelectric inspection type intelligent agents are only responsible for reconnaissance and effectiveness assessment.

[0066] Performance evaluation rules: After each round of operation, a performance evaluation must be carried out by a photoelectric inspection-type intelligent agent. If the target has not been dealt with, supplementary operation will be started, and the cycle will be iterated until all targets are dealt with.

[0067] Action strategies include low-cost operation strategies and high-efficiency operation strategies.

[0068] Work modes include single-objective single-machine work, sequential work, and collaborative saturation work.

[0069] Step 2: Structured Micro-scene Template Design Based on the above list of requirements, a structured micro-scene template is constructed, comprising four core modules: 2.1 Fixed Frame and Global Rules The operational airspace is uniformly defined as 400km×400km; the basic performance parameters of the intelligent agent are standardized (speed 200km / h, endurance 4 hours); and the principles of full-domain operation, the deployment specifications of the intelligent agent, and the basic rules for target generation are clearly defined.

[0070] 2.2 Parameterized Variable Pool and Random Instructions Define the range of values ​​for all-dimensional variables and their generation logic: Environmental variable pool: Weather (sunny / rainy random), temperature (random integer from 15-45℃), humidity (random depending on sunny / rainy scenario), takeoff wind direction (headwind / tailwind / crosswind random), wind speed (random depending on wind direction mode), and operational interference intensity (strong / weak random).

[0071] Target variable pool: target attributes (fixed / moving / composite), number of target types (randomly selected based on coordinate precision), number of single-type targets (1-5 random integers), coordinate precision (precise / fuzzy random), target priority (automatically matched according to preset rules).

[0072] Intelligent agent variable pool: number of bases (5 fixed in the depth scene), number of intelligent agents per base (1-12 random), aircraft type combination (random ratio of three types of aircraft), base coordinates (randomly generated on the west side of the airspace).

[0073] Strategy variable pool: Action strategy (low cost / high efficiency random / specified), operation method (three types of random / specified), operation efficiency threshold (full handling of the target is considered as meeting the standard).

[0074] Random instructions require that local large models generate variables according to their value ranges and that the logical consistency of variables be strictly verified (e.g., humidity ≥ 60% on rainy days, and coordinate delivery agents are disabled for fuzzy coordinates).

[0075] 2.3 Embedding of Strategy and Logic Constraints Transform domain knowledge into hard constraints and write them into the template: Model compatibility constraints: Models must be matched strictly according to coordinate accuracy; one-time intelligent agents cannot be repeatedly scheduled; the number of photoelectric inspection machines must be consistent with the number of target types.

[0076] Strategy execution constraints: Low-cost strategies prioritize coordinate delivery agents, while cooperative positioning requires at least three reconnaissance agents; high-efficiency strategies prioritize scheduling agents closest to the target; both strategies enforce the "operation-evaluation-follow-up attack" closed loop.

[0077] Environmental adaptation constraints: The accuracy of photoelectric guidance decreases in rainy weather, and coordinate delivery type intelligent agents are preferred for targets with precise coordinates; when the wind speed exceeds 10m / s, 1-2 additional reusable photoelectric inspection type intelligent agents are added as backups.

[0078] Task allocation constraints: Sequential tasks are handled from high to low priority; collaborative saturation tasks allocate 2-5 agents to high-value targets and single agents to ordinary targets; single-target single-agent tasks are assigned one agent to one target.

[0079] 2.4 Standardized Output Format Mandation All samples strictly follow a three-part structure: "problem description - analysis process - answer". Problem Description: Fully cover the task name, action strategy, operation method, target attributes and coordinate accuracy, weather conditions, agent deployment information, and detailed operation objectives.

[0080] Analysis process: The process involves sequentially completing the total number of targets, prioritization, aircraft compatibility analysis, base site selection calculation, and quantity allocation derivation, forming a complete reasoning chain.

[0081] Answer: Establish a fixed step-by-step action plan format, clearly defining agent dispatch information, operational steps, performance evaluation process, and cyclical replenishment rules.

[0082] Step 3: Automated data generation based on local large model The core generation engine is a locally deployed large language model. The specific process is as follows: 3.1 Situation Simulation Generation The local large model parses structured template instructions, calls a parameterized variable pool to randomly generate a logically consistent complete operational situation, strictly verifies parameter compliance (such as base deployment limits, meteorological parameter matching, and aircraft coordinate adaptation), and outputs standardized situation text in a three-part problem description format. For example, in the randomly generated situation, the meteorological conditions are selected as rainy, with a temperature of 32 degrees Celsius and humidity of 78%; due to the ambiguity of the operational targets, six types are selected from the moving target set.

[0083] 3.2 Planning Reasoning and Solution Output Based on the generated operational situation, the model combines embedded constraint rules and priority logic to perform resource allocation, path calculation, and time-series planning, deriving executable action plans and externalizing the complete reasoning process. For example, the analysis process needs to clarify that "there are a total of 15 operational targets, with priority order as communication base stations > emergency equipment > ...; due to the ambiguity of target coordinates, it is adapted to integrated rescue and photoelectric inspection intelligent agents...".

[0084] 3.3 Batch Generation and Filtering By iterating through template variables, batch question-answer pairs are generated, and the model verifies the consistency of parameters and logic in real time. After generation, invalid samples with incomplete format, contradictory parameters, or logical conflicts are automatically removed, retaining only compliant and valid data. This process achieves a deep integration of domain expert knowledge and the reasoning capabilities of large-scale models, eliminating the need for manual annotation and significantly improving the efficiency and quality of dataset generation.

[0085] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for generating datasets for intelligent agent handling tasks based on structured prompts, characterized in that, include: Identify the core factors affecting task allocation for intelligent agents and create a standardized requirements list. Based on a standardized requirements list, a structured micro-scene template is constructed, which includes a fixed framework and global rules, a parameterized variable pool and random instructions, embedded strategies and logical constraints, and a standardized output format. The structured micro-scenario template is parsed using a locally deployed large language model, and a parameterized variable pool is called to generate the job status. The embedded constraint rules are combined to perform planning and reasoning, and standardized question-answer pairs containing a complete reasoning chain are output.

2. The method for generating a dataset for an intelligent agent's handling task based on structured prompts as described in claim 1, characterized in that, The core elements influencing task allocation for intelligent agents include: Environmental parameters, objective parameters, agent parameters, and constraint and policy parameters; The environmental parameters include meteorological parameters and operational interference intensity; The target parameters include target attributes, number of targets, coordinate precision, and task priority; The intelligent agent parameters include machine type, deployment location, performance threshold, and quantity configuration; The constraints and strategy parameters include aircraft type operating constraints, performance evaluation rules, action strategies, and operating methods.

3. The method for generating a dataset for intelligent agent handling tasks based on structured prompts as described in claim 1, characterized in that, The structured micro-scene template includes: Establish a fixed framework and global rules to uniformly define the operational airspace, basic performance parameters of the agent, and principles for global operations; Construct a parameterized variable pool, define the value range and generation logic of environment variables, target variables, agent variables and policy variables, and set random instruction requirements; Knowledge from the civilian operations domain is transformed into hard constraints and written into templates to form a strategy and logic constraint embedded module; Set a standardized output format and force all samples to be organized into a three-part structure based on the problem description, analysis process, and answer.

4. The method for generating a dataset for an intelligent agent's handling task based on structured prompts as described in claim 3, characterized in that, The environmental variables include weather, temperature, humidity, takeoff wind direction, wind speed, and intensity of operational interference; The target variables include target type, number of target categories, number of targets per category, coordinate accuracy, and target priority; Agent variables include the number of bases, the number of agents per base, the combination of machine types, and the base coordinates; Strategic variables include action strategies, work methods, and work performance thresholds.

5. The method for generating a dataset for an intelligent agent's handling task based on structured prompts as described in claim 3, characterized in that, The strategy and logic constraint embedding module includes device model adaptation constraints and strategy execution constraints; Among them, the model-adaptive constraint-limited coordinate delivery intelligent agent is only adapted to precise coordinate targets, the comprehensive rescue intelligent agent is adapted to fuzzy or composite coordinate targets, and the photoelectric inspection intelligent agent is only responsible for reconnaissance and performance evaluation. The strategy execution constraints stipulate that the low-cost strategy prioritizes coordinate delivery agents, while the high-efficiency strategy prioritizes scheduling agents closest to the target, and both strategies execute a closed loop from task execution and evaluation to follow-up attack.

6. The method for generating a dataset for an intelligent agent's handling task based on structured prompts as described in claim 3, characterized in that, The established standardized output format forces all samples to be organized into a three-part structure based on the problem description, analysis process, and answer, including: The problem description section includes the task name, action strategy, operation method, target attributes and coordinate accuracy, weather conditions, agent deployment information, and detailed operation objectives; The analysis process segment sequentially completes the total number of targets, priority ranking, aircraft type compatibility analysis, base site selection calculation, and quantity allocation derivation. The answer section clearly outlines the step-by-step action plan, including agent dispatch information, operational steps, performance evaluation process, and cyclical replenishment rules.

7. The method for generating a dataset for an intelligent agent's handling task based on structured prompts as described in claim 1, characterized in that, In the automated data generation based on the local large model, the situation simulation generation process is as follows: the local large model parses the structured micro-scenario template instructions, calls the parameterized variable pool to randomly generate a logically consistent complete operational situation, verifies the compliance of the parameters, and outputs standardized situation text according to the problem description format.

8. The method for generating a dataset for an intelligent agent's handling task based on structured prompts as described in claim 1, characterized in that, The process involves using a locally deployed large language model to parse the structured micro-scene template, calling a parameterized variable pool to generate a work situation, combining embedded constraint rules for planning and reasoning, and outputting standardized question-answer pairs containing a complete reasoning chain, including: Based on the generated operational situation, the model combines embedded constraint rules and priority logic to carry out resource allocation, path calculation and time sequence planning, deduce executable action plans and externalize the complete reasoning process.

9. The method for generating a dataset for an intelligent agent's handling task based on structured prompts as described in claim 1, characterized in that, The method further includes: Batch question-answer pairs are generated by iterating variable values ​​through templates. The model verifies the consistency of parameter logic in real time. After generation, invalid samples with incomplete format, contradictory parameters, and logical conflicts are automatically removed, while compliant and valid data are retained.

Images