Fine-tuning method and device for swarm-oriented radiation reconnaissance

Abstract

This application provides a method and apparatus for fine-tuning a large model for swarm radiation reconnaissance. The method includes: acquiring a base model and configuring a two-layer low-rank adaptation architecture for the base model to obtain a pre-trained large model; the two-layer low-rank adaptation architecture includes a shared low-rank adaptation layer and multiple private low-rank adaptation layers; acquiring multiple key tasks in the radiation reconnaissance domain and setting initial weights for security priorities for each key task; constructing a multi-task joint training dataset based on each key task and performing multi-task fine-tuning training on the pre-trained large model to update the parameters of the pre-trained large model; calculating the weighted sum of security priorities of the losses of each key task and dynamically adjusting the initial weights of security priorities until the weighted sum of security priorities of the losses of each key task is minimized, thereby obtaining the target large model. This application enables the deployment of lightweight large models for multi-task scenarios of swarm radiation reconnaissance, significantly improving inference computation efficiency and reducing memory consumption, and has flexible scalability.

Description

Technical Field

[0001] This application belongs to the field of large model technology and relates to a method and device for fine-tuning large models for swarm radiation reconnaissance. Background Technology

[0002] Swarm radiation reconnaissance is a collaborative reconnaissance system composed of multiple small drones or biomimetic robots, specifically designed for rapid detection and dynamic monitoring of areas contaminated by nuclear radiation. Each "swarm" is equipped with a miniature gamma / neutron detector and an energy spectrum analysis module. Through swarm intelligence algorithms, it achieves autonomous formation, area scanning, and data fusion, accurately locating radiation sources and generating three-dimensional radiation field maps. This significantly reduces the risk of radiation exposure to personnel and is suitable for scenarios such as nuclear accident emergencies and border checks.

[0003] In swarm systems, multiple parallel intelligent tasks heavily rely on the support of large language models (LLMs). For example, radiation anomaly detection, dynamic path planning, and radiation source type identification require LLMs for data fusion, semantic understanding, and logical reasoning. Therefore, how to deploy lightweight large models for multi-task scenarios of swarm radiation reconnaissance is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This application provides a method and apparatus for fine-tuning a large model for swarm radiation reconnaissance, which solves the technical problem of how to deploy a lightweight large model for multi-task scenarios of swarm radiation reconnaissance.

[0005] Firstly, this application provides a method for fine-tuning a large model for swarm radiation reconnaissance, the method comprising:

[0006] A base model is obtained, and a two-layer low-rank adaptation architecture is configured for the base model to obtain a pre-trained large model; the two-layer low-rank adaptation architecture includes a shared low-rank adaptation layer and multiple private low-rank adaptation layers.

[0007] Acquire multiple key tasks in the field of radiation reconnaissance and assign initial weights for security priorities to each of these key tasks;

[0008] A multi-task joint training dataset is constructed based on each of the key tasks, and the pre-trained large model is fine-tuned using the multi-task joint training dataset to update the parameters of the pre-trained large model.

[0009] Calculate the weighted sum of the security priorities of the losses of each of the critical tasks, and dynamically adjust the initial weights of the security priorities until the weighted sum of the security priorities of the losses of each of the critical tasks is minimized, and then obtain the target large model.

[0010] In some embodiments of the first aspect of this application, multi-task fine-tuning training of the pre-trained large model using the multi-task joint training dataset includes:

[0011] The training data in the multi-task joint training dataset is input into the pre-trained large model, which then performs inference calculations to obtain prediction outputs. Specifically, the parameters of the corresponding private low-rank adaptation layer and the parameters of the shared low-rank adaptation layer are selected according to the key task to which the training data belongs, and the weight parameters of the base model are kept frozen.

[0012] In some embodiments of the first aspect of this application, the number of parameters of the two-layer low-rank adaptive architecture is not greater than a preset ratio range of the number of parameters of the base model, and the parameters of the shared low-rank adaptive layer are greater than the parameters of the private low-rank adaptive layer.

[0013] In some embodiments of the first aspect of this application, acquiring multiple key tasks in the field of radiation reconnaissance and setting initial weights for security priorities for each key task includes:

[0014] The key tasks in the field of radiation reconnaissance include radiation anomaly detection, radiation source type identification, natural language command understanding, radiation situation report generation, and intelligent trajectory planning assistance.

[0015] A first security priority initial weight is set for the radiation anomaly detection task;

[0016] A second security priority initial weight is set for the radiation source type discrimination task;

[0017] A third security priority initial weight is set for the natural language instruction understanding task;

[0018] A fourth security priority initial weight is set for the radiation situation report generation task;

[0019] The fifth safety priority initial weight is set for the intelligent trajectory planning assistance task.

[0020] In some embodiments of the first aspect of this application, when dynamically adjusting the initial weight of the security priority, the initial weight of the first security priority of the radiation anomaly detection task shall not be lower than a preset proportion range of the initial value, and the initial weight of the second security priority of the radiation source type discrimination task shall not be lower than a preset proportion range of the initial value.

[0021] In some embodiments of the first aspect of this application, the method further includes:

[0022] The base model in the target large model is quantized and compressed so that the weight parameters of the base model are quantized from high precision to low precision; and the parameters of the two-layer low-rank adaptive architecture are kept at high precision.

[0023] In some embodiments of the first aspect of this application, the method further includes:

[0024] The target large model is deployed in an edge-cloud collaborative manner to perform inference calculations for swarm radiation reconnaissance missions; wherein,

[0025] When communication is normal, the swarm radiation reconnaissance tasks are classified to allocate tasks meeting the first criterion to the edge for inference computation, and tasks meeting the second criterion to the cloud for inference computation; and

[0026] When the communication status is abnormal, the swarm radiation reconnaissance mission is rolled back to local inference calculation.

[0027] In some embodiments of the first aspect of this application, reasoning calculations for swarm radiation reconnaissance missions include:

[0028] The parameters of the corresponding private low-rank adaptation layer are loaded in the front-end buffer according to the current swarm radiation reconnaissance mission, so as to perform inference calculations on the current mission through the target large model;

[0029] The parameters of the corresponding private low-rank adaptation layer are loaded in the background buffer according to the subsequent swarm radiation reconnaissance mission, and the subsequent swarm radiation reconnaissance mission is switched to the current mission through atomic swap so as to perform inference calculation through the target large model.

[0030] In some embodiments of the first aspect of this application, the method further includes:

[0031] The parameters of the target large model are frozen to incrementally fine-tune the private low-rank adaptation layer according to the new key task to obtain an incremental model;

[0032] The incremental model is used to perform multi-task security assessment in a security sandbox, and the target large model is updated using the incremental model when the security threshold is met.

[0033] Secondly, this application provides a large-scale model fine-tuning device for swarm radiation reconnaissance, the device comprising:

[0034] A building module is configured to acquire a base model and configure a two-layer low-rank adaptation architecture for the base model to acquire a pre-trained large model; the two-layer low-rank adaptation architecture includes a shared low-rank adaptation layer and multiple private low-rank adaptation layers.

[0035] The security module is configured to acquire multiple key tasks in the radiation reconnaissance field and set initial weights for security priorities for each of the key tasks;

[0036] The training module is configured to construct a multi-task joint training dataset based on each of the key tasks, and to perform multi-task fine-tuning training on the pre-trained large model using the multi-task joint training dataset to update the parameters of the pre-trained large model.

[0037] The loss calculation module is configured to calculate the safety priority weighted sum of the losses of each of the key tasks, and dynamically adjust the initial weights of the safety priorities until the safety priority weighted sum of the losses of each of the key tasks is minimized, thereby obtaining the target large model.

[0038] As described above, the large-scale model fine-tuning method and apparatus for swarm radiation reconnaissance provided in this application have the following beneficial effects:

[0039] This application configures a two-layer low-rank adaptive architecture through a base model, which can simultaneously serve multi-task scenarios of swarm radiation reconnaissance with only a few parameters added, reducing the GPU memory usage from approximately 84GB to approximately 5GB. Simultaneously, this application ensures that the performance of critical tasks is not compromised through joint training, and utilizes a shared low-rank adaptive layer to capture general knowledge in the radiation reconnaissance domain for reuse across all tasks, achieving efficient reuse of domain knowledge.

[0040] In terms of deployment, this application employs a combination of hybrid quantization, a double-buffering mechanism, and communication degradation fallback to make the lightweight large model friendly to edge environments. Furthermore, in online applications, adding new tasks only requires training a new set of private low-rank adaptation layer parameters, demonstrating flexible scalability. In addition, this application uses a secure sandbox verification mechanism to ensure that incremental fine-tuning during online adaptation does not lead to performance degradation, ensuring the system remains stable and reliable. Attached Figure Description

[0041] Figure 1 The diagram shown is a flowchart illustrating the large model fine-tuning method for swarm radiation reconnaissance provided in this application embodiment.

[0042] Figure 2 The diagram shown is a forward propagation schematic of the pre-trained large model provided in the above embodiments of this application.

[0043] Figure 3 The diagram shown is a flowchart illustrating the process of dynamically adjusting the initial weight of security priority according to an embodiment of this application.

[0044] Figure 4 The diagram shown is an application illustration of edge-cloud collaborative inference deployment of a target large model provided in the embodiments of this application.

[0045] Figure 5The diagram shown is a structural schematic of the large model fine-tuning device for swarm radiation reconnaissance as described in an embodiment of this application. Detailed Implementation

[0046] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0047] It should be noted that in the following description, reference is made to the accompanying drawings, which illustrate several embodiments of this application. It should be understood that other embodiments may also be used, and changes in mechanical composition, structure, electrical system, and operation may be made without departing from the spirit and scope of this application. The following detailed description should not be considered limiting, and the scope of the embodiments of this application is defined only by the claims of the published patent. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. Spatial terms such as “upper,” “lower,” “left,” “right,” “below,” “below,” “lower part,” “above,” “upper part,” etc., may be used herein to illustrate the relationship between one element or feature shown in the figures and another element or feature.

[0048] Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are to be interpreted inclusively, or mean any one or any combination thereof.

[0049] In the field of radiation reconnaissance, with the development of unmanned systems and intelligent swarm technologies, swarm radiation reconnaissance systems have gradually become an important tool. These systems typically consist of multiple unmanned platforms and need to perform multiple parallel intelligent tasks in complex, dynamic, and potentially high-risk environments, such as radiation field construction, radionuclide localization, nuclide identification, threat assessment, and path planning. These tasks place high demands on capabilities such as language understanding, knowledge reasoning, and command following, thus generally requiring the use of large language models to provide underlying cognitive and decision-making support. However, current technologies still face a series of core challenges in meeting the parallel requirements of swarm systems for large language model capabilities.

[0050] First, the conflict between resource constraints and multi-model deployment is significant. Swarm platforms typically have limited computing power, storage, and power consumption, making it difficult to simultaneously support multiple independent large language models. Deploying a complete model for each intelligent task alone will quickly exhaust system resources, severely impacting real-time performance and endurance. Second, general-purpose large language models are insufficient for zero-shot scenarios to meet the specialized needs of radiation reconnaissance. Standard pre-trained models lack detailed knowledge in radiation physics, nuclide characteristics, and reconnaissance procedures, making it difficult to directly perform highly reliable radiation data analysis and task decision-making.

[0051] Furthermore, existing multi-task low-rank adaptation methods have significant limitations. While low-rank adaptation and similar schemes support multi-task sharing of the base model to some extent, they often struggle to account for the specificities of each task within a limited parameter space when handling multiple heterogeneous reconnaissance tasks, leading to performance degradation or unstable adaptation effects between tasks. Meanwhile, most current edge deployment schemes for large language models lack system-level optimizations for radiation reconnaissance tasks, including inference acceleration, model pruning, quantization strategies, and task scheduling, failing to meet the low latency and high energy efficiency requirements of swarm systems.

[0052] Finally, the gradient conflict problem, which is prevalent in multi-task joint fine-tuning, further exacerbates the aforementioned difficulties. Different intelligent tasks update the large model parameters in different directions, and directly sharing the large model parameters can lead to mutual interference between tasks, and even damage the overall performance of the model.

[0053] To at least address the aforementioned technical issues, this application proposes a method and apparatus for fine-tuning a large model for swarm radiation reconnaissance. This method can deploy lightweight large models for multi-task scenarios of swarm radiation reconnaissance, significantly improving inference computation efficiency and reducing memory usage, while also offering flexibility and scalability.

[0054] The technical solutions in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0055] Figure 1 The diagram shown illustrates a flowchart of a large-model fine-tuning method for swarm radiation reconnaissance provided in an embodiment of this application. Figure 1 As shown, the large model fine-tuning method for swarm radiation reconnaissance includes steps S1 to S4.

[0056] Step S1: Obtain the base model and configure a two-layer low-rank adaptation architecture for the base model to obtain a pre-trained large model; the two-layer low-rank adaptation architecture includes a shared low-rank adaptation layer and multiple private low-rank adaptation layers.

[0057] In some embodiments, the base model can be an open-source lightweight LLM (such as Qwen3-8B). Qwen3-8B has approximately 8 billion parameters and is significantly optimized for inference, instruction following, multilingual support, and long text processing (context length can reach tens of thousands of words), while maintaining a low inference deployment threshold and can run efficiently on consumer-grade graphics cards and even some edge devices.

[0058] In some embodiments, a two-layer low-rank adaptation architecture is configured for the base model to obtain a pre-trained large model. The two-layer low-rank adaptation architecture includes a shared low-rank adaptation layer and multiple private low-rank adaptation layers. Each private low-rank adaptation layer corresponds to a key task in the radiation reconnaissance domain and is used to capture task-specific knowledge, while the shared low-rank adaptation layer is general to all tasks and is used to capture general knowledge in the radiation reconnaissance domain.

[0059] In some embodiments, the rank of the shared low-rank adaptation layer is 16, applied to the bottom to middle layers of the Transformer in the base model; the rank of the private low-rank adaptation layer is 4, applied to the higher layers of the Transformer in the base model.

[0060] Furthermore, the number of parameters in the two-layer low-rank adaptive architecture does not exceed a preset proportion of the number of parameters in the base model, and the parameters of the shared low-rank adaptive layer are greater than the parameters of the private low-rank adaptive layer. In some embodiments, the number of parameters in the two-layer low-rank adaptive architecture is controlled within 3% of the base model. That is, the base model only needs to increase its parameters by a maximum of 3% to simultaneously serve multiple key tasks in the radiation reconnaissance field, effectively reducing memory consumption.

[0061] Step S2: Obtain multiple key tasks in the field of radiation reconnaissance and set initial weights for security priority for each key task.

[0062] In some embodiments, the key tasks in the field of radiation reconnaissance are: radiation anomaly detection, radiation source type identification, natural language command understanding, radiation situation report generation, and trajectory intelligent planning assistance. Among these tasks, the radiation anomaly detection task refers to automatically identifying abnormal signals deviating from normal background levels from continuous or discrete radiation monitoring data to determine whether there is a potential radioactive threat; the radiation source type identification task refers to classifying and identifying radiation sources based on acquired energy spectrum or dose rate characteristics, determining whether they are natural nuclides, medical nuclides, industrial nuclides, or special nuclear materials; the natural language command understanding task refers to parsing reconnaissance commands issued by operators in natural language (such as "focus on scanning the northeast direction area") into executable semantic representations and task parameters; the radiation situation report generation task refers to automatically generating structured or text-based radiation environment situation descriptions and assessment reports for command personnel or technical experts by integrating multi-source data and analysis results during the reconnaissance process; and the intelligent trajectory planning assistance task refers to generating or recommending optimized reconnaissance trajectories for the swarm platform by combining current radiation field distribution predictions, environmental obstacles, and mission objectives to improve detection efficiency and reduce the radiation risk to personnel and the platform.

[0063] Furthermore, in multi-task joint training in the field of radiation reconnaissance, the gradient directions of different tasks may conflict. That is, when the model shares underlying parameters, the parameter update direction calculated by task A and the update direction of task B are at an obtuse angle or even opposite in vector space. In this case, updating parameters simultaneously along two directions can produce mutual cancellation or even negative effects, that is, the optimization of one task will impair the performance of another task. To address this, this application proposes a gradient reconciliation strategy based on the security priority of radiation reconnaissance tasks.

[0064] In some embodiments, a first security priority initial weight is set for the radiation anomaly detection task. A second security priority initial weight is set for the radiation source type discrimination task. ; Set a third security priority initial weight for the natural language instruction understanding task. A fourth security priority initial weight is set for the radiation situation report generation task. ; Set the initial weight of the fifth safety priority for the intelligent trajectory planning assistance task. .

[0065] Step S3: Construct a multi-task joint training dataset based on each of the key tasks, and use the multi-task joint training dataset to perform multi-task fine-tuning training on the pre-trained large model in order to update the parameters of the pre-trained large model.

[0066] In some embodiments, a corresponding multi-task joint training dataset is constructed based on the category of the key task. For example, when the key tasks are radiation anomaly detection, radiation source type identification, natural language instruction understanding, radiation situation report generation, and intelligent trajectory planning assistance, a corresponding multi-task joint dataset is constructed, including radiation anomaly detection data, radiation source classification data, natural language instruction understanding data, radiation situation report generation data, and intelligent trajectory planning suggestion data. All data uses a unified Prompt template system, with task type identifiers distinguishing different tasks.

[0067] Furthermore, Monte Carlo simulations of radiation fields can be used to generate simulated data, which can then be combined with real data to construct a multi-task joint training dataset. The Monte Carlo method for generating simulated radiation field data is a computational technique based on probability statistics and particle transport theory. This method simulates the random transport processes of a large number of radiating particles, such as gamma photons and neutrons, in a medium (e.g., air, soil, buildings), including emission, scattering, absorption, and detection response, thereby statistically determining the energy deposition spectrum, dose rate, or flux distribution at the detection point. In the field of radiation reconnaissance, this can generate labeled datasets covering different nuclide types, geometric layouts, shielding conditions, and background environments without the need for actual radioactive sources. These datasets can be used to train large models and also provide a controllable and low-cost benchmark for evaluating the performance of multi-task algorithms in complex radiation fields.

[0068] In some embodiments, a pre-trained large model is fine-tuned using a multi-task joint training dataset to update its parameters. Specifically, training data from the multi-task joint training dataset is input into the pre-trained large model, enabling it to perform inference computation, obtain predicted outputs, and update its parameters based on the loss between the predicted outputs and the true labels.

[0069] Furthermore, during inference computation of the pre-trained large model, it selects the parameters of the corresponding private low-rank adaptation layer and the parameters of the shared low-rank adaptation layer according to the key task to which the training data belongs, and keeps the weight parameters of the base model frozen. The mathematical expression for the forward propagation of the pre-trained large model is shown in Equation (1).

[0070]

[0071] in, For input, For the frozen base model weights; To share low-rank adaptation layer parameters (shared by all missions) and capture general knowledge in the radiation reconnaissance domain, rank = 16, applied to the bottom to middle layers of the Transformer; For the parameters of the private low-rank adaptation layer of the i-th critical task, capture task-specific knowledge, rank = 4, applied to higher levels of the Transformer.

[0072] Figure 2 The diagram shown illustrates the forward propagation of a pre-trained large model provided in an embodiment of this application. Figure 2 As shown, when the critical tasks are radiation anomaly detection, radiation source type identification, natural language command understanding, radiation situation report generation, and intelligent trajectory planning assistance, the two-layer low-rank adaptive architecture includes a shared low-rank adaptive layer and five private low-rank adaptive layers. During forward propagation, the pre-trained large model loads the parameters of the corresponding private low-rank adaptive layers and the parameters of the shared low-rank adaptive layers according to the critical task to which the input training data belongs, performs inference calculations, and keeps the weight parameters of the base model frozen. Finally, the results are summed and output.

[0073] Step S4: Calculate the weighted sum of the security priorities of the losses of each critical task, and dynamically adjust the initial weights of the security priorities until the weighted sum of the security priorities of the losses of each critical task is minimized, and then obtain the target large model.

[0074] In some embodiments, during multi-task joint training, let the gradient of the k-th key task at training step t be . The security priority weight is The fusion gradient is then shown in equation (2):

[0075]

[0076] During training, the direction and scale of gradients generated by different tasks during multi-task training are adjusted by calculating fusion gradients, thereby stabilizing and optimizing the joint parameters of the model, so that the joint gradients used to update the parameters can benefit all key tasks at the same time.

[0077] Furthermore, when training the pre-trained large model using a multi-task joint training dataset, this application adopts a task alternation sampling strategy, with different learning rates for the shared low-rank adaptation layer and the private low-rank adaptation layer. The training objective is to minimize the safety priority weighted sum of the losses of each task. That is, the initial weights of the safety priority are dynamically adjusted during training, and the fusion gradient is calculated to update the model parameters until the safety priority weighted sum of the losses of each key task is minimized, at which point the target large model is obtained. The safety priority weighted sum of the losses of each key task can be expressed by equation (3):

[0078]

[0079] in, To share low-rank adaptation layer parameters, For the first Private low-rank adaptation layer parameters for each task Let be the loss function for the k-th task.

[0080] Furthermore, the loss function measures the difference between the model's predicted output and the true label, outputting a scalar value (i.e., the loss value) to guide the model's parameter updates. In multi-task joint training, the loss function for each task can be customized, and this application does not impose any restrictions on this.

[0081] Figure 3 This is a schematic diagram illustrating the process of dynamically adjusting the initial weight of security priorities according to an embodiment of this application. Figure 3 As shown, a first security priority initial weight is set for the radiation anomaly detection task. A second security priority initial weight is set for the radiation source type discrimination task. ; Set a third security priority initial weight for the natural language instruction understanding task. A fourth security priority initial weight is set for the radiation situation report generation task. ; Set the initial weight of the fifth safety priority for the intelligent trajectory planning assistance task. Then, the fusion gradient is calculated using equation (2), and the model parameters are updated accordingly. Next, the loss of each key task is calculated, and the safety priority weighted sum of the loss of each key task is calculated using equation (3). The initial weight of the safety priority is dynamically adjusted continuously until the safety priority weighted sum of the loss of each key task is minimized, at which point the target large model is obtained.

[0082] It should be noted that when dynamically adjusting the initial weights of the safety priorities, the initial weight of the first safety priority for the radiation anomaly detection task shall not be lower than a preset percentage range of the initial value, and the initial weight of the second safety priority for the radiation source type determination task shall not be lower than a preset percentage range of the initial value. In some embodiments, the initial weight of the first safety priority for the radiation anomaly detection task shall not be lower than 80% of the initial value, and the initial weight of the second safety priority for the radiation source type determination task shall not be lower than 80% of the initial value.

[0083] Furthermore, after achieving the aforementioned training objective, namely minimizing the weighted sum of the safety priorities of the losses for each of the key tasks, the target large model is obtained. At this point, the base model in the target large model is quantized and compressed, so that the weight parameters of the base model are quantized from high precision to low precision; and the parameters of the two-layer low-rank adaptive architecture are kept at high precision.

[0084] In some embodiments, the weights of the base model in the target large model are quantized from FP16 to INT4 precision, compressing the model size from approximately 14GB to approximately 4GB. Meanwhile, the weight parameters of the two-layer low-rank adaptive architecture remain unchanged at FP16 precision. The AWQ algorithm can be used to quantize the model's weight parameters. AWQ (Activation-aware Weight Quantization) is an efficient weight quantization algorithm for large language models. It analyzes the distribution characteristics of model activation values ​​to identify the most significant weights (typically accounting for about 1% of the total parameters) that have the most impact on inference accuracy. During the quantization process, these key weights are retained at high precision (e.g., FP16), while other ordinary weights are quantized at low precision (e.g., INT4).

[0085] In some embodiments, after the accuracy quantization of the target large model is completed, the target large model is deployed collaboratively from the edge to the cloud for inference calculation of the swarm radiation reconnaissance task, i.e., an application of edge-cloud collaborative inference deployment. Specifically, when communication is normal, the swarm radiation reconnaissance task is classified, with tasks meeting a first criterion assigned to the edge for inference calculation and tasks meeting a second criterion assigned to the cloud for inference calculation; and when communication is abnormal, the swarm radiation reconnaissance task is rolled back to local inference calculation. The first criterion is a real-time requirement, and the second criterion is a complexity requirement.

[0086] Figure 4 The diagram shown illustrates the application of edge-cloud collaborative inference deployment of a target large model provided in this embodiment of the application. Figure 4 As shown, after receiving the input from the swarm radiation reconnaissance mission, when communication is normal, the target large model classifies the swarm radiation reconnaissance mission through the mission router. Tasks with high real-time requirements are assigned to the edge for inference calculations, including radiation anomaly detection, radiation source type identification, natural language command understanding, and intelligent trajectory planning assistance. Simultaneously, tasks with high complexity are uploaded to the cloud for processing, including situational analysis and report generation. Furthermore, when communication is abnormal, the swarm radiation reconnaissance mission is reverted to local inference calculations.

[0087] Furthermore, when the communication status is normal, when the target large model performs inference calculations for the swarm radiation reconnaissance task, it includes loading the parameters of the corresponding private low-rank adaptation layer in the front-end buffer according to the current swarm radiation reconnaissance task, so as to perform inference calculations for the current task through the target large model; loading the parameters of the corresponding private low-rank adaptation layer in the back-end buffer according to the subsequent swarm radiation reconnaissance task, and switching the subsequent swarm radiation reconnaissance task to the current task through atomic swaps so as to perform inference calculations through the target large model.

[0088] In some embodiments, when performing inference calculations for swarm radiation reconnaissance missions, the system employs an efficient double-buffering mechanism to optimize the switching process of the private low-rank adaptation layer when loading the corresponding task's private low-rank adaptation layer weights into the target large model. The model pre-loads the private low-rank adaptation layer weights of subsequent new tasks into the back buffer and completes the necessary verification and preparation work, while forward inference continues continuously from the front buffer, without blocking each other. When switching to a subsequent task is required, the model only needs to perform one atomic swap operation to update the pointer of the active buffer, instantly activating the new back buffer weights. The entire switching process can be stably controlled within 1 millisecond, truly achieving zero-latency task switching.

[0089] Taking the deployment of a multi-task model of a swarm radiation reconnaissance system in a nuclear emergency response scenario as an example, this paper further illustrates the technical solution of this application.

[0090] (1) Training datasets for five tasks were constructed, totaling approximately 11,300 data points, with radiation anomaly detection, radiation source type identification, natural language command understanding, radiation situation report generation, and intelligent trajectory planning assistance as key tasks. Radioactive source energy spectrum data were generated using Monte Carlo simulation (MCNP / Geant4) to expand the training set. Furthermore, all data were formatted using the Prompt template.

[0091] (2) Qwen3-8B was selected as the base model (32-layer Transformer) to construct a two-layer low-rank adaptive architecture: a shared low-rank adaptive layer r=16 was applied to layers 1-16 (approximately 4.2M parameters), and private low-rank adaptive layers r=4 for each of the 5 tasks were applied to layers 17-32 (approximately 1.05M per group). The total number of parameters in the two-layer low-rank adaptive architecture was approximately 9.45M, accounting for 0.135% of the base model.

[0092] (3) Joint training is performed with safety priority weights. The loss of the validation set of each task is checked every 10 epochs. If the loss of a task recovers, its weight is increased by 10%, but the safety priority weight of the radiation anomaly detection task and the radiation source type discrimination task shall not be less than 80% of the initial value.

[0093] (4) The base model was quantized to INT4 using the AWQ algorithm, compressed from 14GB to approximately 3.8GB. It was deployed on a ground station workstation equipped with an NVIDIA RTX 4090. The anomaly detection inference latency was approximately 35ms / line.

[0094] (5) Private low-rank adaptation layer hot switching realizes seamless switching between reconnaissance phases: during the wide-area scanning phase, private low-rank adaptation layer_anomaly detection + private low-rank adaptation layer_command understanding are loaded; after anomaly is detected, private low-rank adaptation layer_track planning is preloaded in the background, and the switching time is <1ms.

[0095] Furthermore, this application also supports incremental fine-tuning after acquiring new data in actual reconnaissance missions. In some embodiments, the parameters of the target big model are frozen to incrementally fine-tune the private low-rank adaptation layer according to the new key mission to obtain an incremental model; the incremental model is subjected to multi-task security evaluation in a security sandbox, and the target big model is updated using the incremental model when a security threshold is met.

[0096] In some embodiments, when the target large model needs to be deployed in an aquatic environment, 200 aquatic environmental radiation data points are collected, and only the private low-rank adaptation layer is fine-tuned to obtain an incremental model. The incremental model is then subjected to multi-task security evaluation in a security sandbox using a standard validation set. When a security threshold, such as the accuracy of the core task, is met, the target large model is updated using the incremental model, thereby flexibly expanding to other new application scenarios. Simultaneously, this application retains the private low-rank adaptation layer weights from previous versions. If problems are found with the updated model in practical applications, it can be quickly rolled back to any stable historical version, thus achieving agile model iteration while ensuring security and stability.

[0097] The scope of protection for the large model fine-tuning method for swarm radiation reconnaissance described in this application is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this application is included within the scope of protection of this application.

[0098] Therefore, this application configures a two-layer low-rank adaptive architecture through a base model, which can simultaneously serve multi-task scenarios of swarm radiation reconnaissance with only a few parameters added, reducing the GPU memory usage from approximately 84GB to approximately 5GB. Simultaneously, this application ensures that the performance of critical tasks is not compromised through joint training, and utilizes a shared low-rank adaptive layer to capture general knowledge in the radiation reconnaissance domain for reuse by all tasks, achieving efficient reuse of domain knowledge.

[0099] In terms of deployment, this application employs a combination of hybrid quantization, a double-buffering mechanism, and communication degradation fallback to make the lightweight large model friendly to edge environments. Furthermore, in online applications, adding new tasks only requires training a new set of private low-rank adaptation layer parameters, demonstrating flexible scalability. In addition, this application uses a secure sandbox verification mechanism to ensure that incremental fine-tuning during online adaptation does not lead to performance degradation, ensuring the system remains stable and reliable.

[0100] Figure 5The diagram shown is a structural schematic of the large-model fine-tuning device for swarm radiation reconnaissance described in an embodiment of this application. Figure 5 As shown, the large model fine-tuning device 100 for swarm radiation reconnaissance includes a construction module 110, a security module 120, a training module 130, and a loss calculation module 140.

[0101] Module 110 is configured to acquire a base model and configure a two-layer low-rank adaptation architecture for the base model to acquire a pre-trained large model; the two-layer low-rank adaptation architecture includes a shared low-rank adaptation layer and multiple private low-rank adaptation layers.

[0102] Security module 120 is configured to acquire multiple key tasks in the radiation reconnaissance field and set initial weights for security priorities for each of the key tasks;

[0103] Training module 130 is configured to construct a multi-task joint training dataset based on each of the key tasks, and to use the multi-task joint training dataset to perform multi-task fine-tuning training on the pre-trained large model in order to update the parameters of the pre-trained large model.

[0104] The loss calculation module 140 is configured to calculate the safety priority weighted sum of the losses of each of the key tasks, and dynamically adjust the initial weight of the safety priority until the safety priority weighted sum of the losses of each of the key tasks is minimized, and then obtain the target large model.

[0105] It should be noted that the principles of the construction module 110, security module 120, training module 130, and loss calculation module 140 provided in this application correspond to the above method steps, and will not be repeated here.

[0106] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or modules or units may be electrical, mechanical, or other forms.

[0107] The modules / units described as separate components may or may not be physically separate. The components shown as modules / units may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected to achieve the objectives of the embodiments of this application, depending on actual needs. For example, the functional modules / units in the various embodiments of this application may be integrated into one processing module, or each module / unit may exist physically separately, or two or more modules / units may be integrated into one module / unit.

[0108] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0109] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0110] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for fine-tuning a large model for swarm radiation reconnaissance, characterized in that, The method includes: A base model is obtained, and a two-layer low-rank adaptation architecture is configured for the base model to obtain a pre-trained large model; the two-layer low-rank adaptation architecture includes a shared low-rank adaptation layer and multiple private low-rank adaptation layers. Acquire multiple key tasks in the field of radiation reconnaissance and assign initial weights for security priorities to each of these key tasks; A multi-task joint training dataset is constructed based on each of the key tasks, and the pre-trained large model is fine-tuned using the multi-task joint training dataset to update the parameters of the pre-trained large model. Calculate the weighted sum of the security priorities of the losses of each of the critical tasks, and dynamically adjust the initial weights of the security priorities until the weighted sum of the security priorities of the losses of each of the critical tasks is minimized, and then obtain the target large model.

2. The method for fine-tuning a large model for swarm radiation reconnaissance according to claim 1, characterized in that, Multi-task fine-tuning of the pre-trained large model using the multi-task joint training dataset includes: The training data in the multi-task joint training dataset is input into the pre-trained large model, which then performs inference calculations to obtain prediction outputs. Specifically, the parameters of the corresponding private low-rank adaptation layer and the parameters of the shared low-rank adaptation layer are selected according to the key task to which the training data belongs, and the weight parameters of the base model are kept frozen.

3. The method for fine-tuning a large model for swarm radiation reconnaissance according to claim 1, characterized in that, The number of parameters in the dual-layer low-rank adaptive architecture is not greater than a preset ratio range of the number of parameters in the base model, and the parameters of the shared low-rank adaptive layer are greater than the parameters of the private low-rank adaptive layer.

4. The method for fine-tuning a large model for swarm radiation reconnaissance according to claim 1, characterized in that, Acquire multiple key tasks in the field of radiation reconnaissance and assign initial weights to security priorities for each of these key tasks, including: The key tasks in the field of radiation reconnaissance include radiation anomaly detection, radiation source type identification, natural language command understanding, radiation situation report generation, and intelligent trajectory planning assistance. A first security priority initial weight is set for the radiation anomaly detection task; A second security priority initial weight is set for the radiation source type discrimination task; A third security priority initial weight is set for the natural language instruction understanding task; A fourth security priority initial weight is set for the radiation situation report generation task; The fifth safety priority initial weight is set for the intelligent trajectory planning assistance task.

5. The method for fine-tuning a large model for swarm radiation reconnaissance according to claim 4, characterized in that, When dynamically adjusting the initial weight of the safety priority, the initial weight of the first safety priority of the radiation anomaly detection task shall not be lower than the preset proportion range of the initial value, and the initial weight of the second safety priority of the radiation source type discrimination task shall not be lower than the preset proportion range of the initial value.

6. The method for fine-tuning a large model for swarm radiation reconnaissance according to claim 1, characterized in that, The method further includes: The base model in the target large model is quantized and compressed, so that the weight parameters of the base model are quantized from high precision to low precision; and Maintain high accuracy of parameters in the two-layer low-rank adaptive architecture.

7. The method for fine-tuning a large model for swarm radiation reconnaissance according to claim 1, characterized in that, The method further includes: The target large model is deployed in an edge-cloud collaborative manner to perform inference calculations for swarm radiation reconnaissance missions; wherein, When communication is normal, the swarm radiation reconnaissance tasks are classified to allocate tasks meeting the first criterion to the edge for inference computation, and tasks meeting the second criterion to the cloud for inference computation; and When the communication status is abnormal, the swarm radiation reconnaissance mission is rolled back to local inference calculation.

8. The method for fine-tuning a large model for swarm radiation reconnaissance according to claim 1, characterized in that, The reasoning and calculation for the swarm radiation reconnaissance mission includes: The parameters of the corresponding private low-rank adaptation layer are loaded in the front-end buffer according to the current swarm radiation reconnaissance mission, so as to perform inference calculations on the current mission through the target large model; The parameters of the corresponding private low-rank adaptation layer are loaded in the background buffer according to the subsequent swarm radiation reconnaissance mission, and the subsequent swarm radiation reconnaissance mission is switched to the current mission through atomic swap so as to perform inference calculation through the target large model.

9. The method for fine-tuning a large model for swarm radiation reconnaissance according to claim 1, characterized in that, The method further includes: The parameters of the target large model are frozen to incrementally fine-tune the private low-rank adaptation layer according to the new key task to obtain an incremental model; The incremental model is used to perform multi-task security assessment in a security sandbox, and the target large model is updated using the incremental model when the security threshold is met.

10. A large-scale model fine-tuning device for swarm radiation reconnaissance, characterized in that, The device includes: A building module is configured to acquire a base model and configure a two-layer low-rank adaptation architecture for the base model to acquire a pre-trained large model; the two-layer low-rank adaptation architecture includes a shared low-rank adaptation layer and multiple private low-rank adaptation layers. The security module is configured to acquire multiple key tasks in the radiation reconnaissance field and set initial weights for security priorities for each of the key tasks; The training module is configured to construct a multi-task joint training dataset based on each of the key tasks, and to perform multi-task fine-tuning training on the pre-trained large model using the multi-task joint training dataset to update the parameters of the pre-trained large model. The loss calculation module is configured to calculate the safety priority weighted sum of the losses of each of the key tasks, and dynamically adjust the initial weights of the safety priorities until the safety priority weighted sum of the losses of each of the key tasks is minimized, thereby obtaining the target large model.

Images