Model self-evolution method and device, storage medium and electronic equipment

Abstract

The invention relates to the technical field of large models, and discloses a model self-evolution method and device, a storage medium and electronic device.The model self-evolution method comprises the steps that a context data packet is acquired, and a to-be-evolved large model is utilized to generate a candidate self-editing data set according to the context data packet; updating the to-be-evolved large model by using candidate self-editing data in the candidate self-editing data set, evaluating a downstream task model of the to-be-evolved large model based on each updated temporary large model, and constructing a target training data set according to an evaluation result; and training the to-be-evolved large model by using the target training data set, and recording the trained large model as the to-be-evolved large model. According to the method, a static pre-training model is converted into a dynamic system which can continuously learn and evolve, so that the model can continuously iterate itself in continuous interaction with new data, and near-real-time, automatic and continuous learning and evolution of a large model to be evolved can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large model, and particularly relates to a model self-evolution method, device, storage medium and electronic equipment. BACKGROUND

[0002] With the development of artificial intelligence technology, traffic multi-modal large models represented by TransGPT have shown strong capabilities in understanding complex traffic scenarios. These large models can integrate image, text, radar point cloud and other data, providing an unprecedented perception and analysis basis for intelligent transportation systems (ITS).

[0003] However, one of the core limitations of these large models is their static nature. After large-scale pre-training and deployment, the internal knowledge and reasoning ability of the large model is fixed. However, the real-world traffic environment is highly dynamic and constantly evolving, with new traffic patterns, road facilities and unexpected events emerging. Static models cannot adapt to these changes, and their performance will deteriorate over time, making it difficult to meet the stringent requirements for real-time and adaptability.

[0004] The main way of related update models relies on human-driven offline retraining, that is, human experts continuously collect and label new data, and periodically fine-tune or retrain the entire model. This process is costly and time-consuming. SUMMARY

[0005] The present application aims to at least partially solve one of the technical problems in the related art. To this end, one object of the present application is to propose a model self-evolution method that can realize near-real-time, automated and continuous learning and evolution of the model.

[0006] A second object of the present application is to propose a model self-evolution device.

[0007] A third object of the present application is to propose a computer-readable storage medium.

[0008] A fourth object of the present application is to propose an electronic device.

[0009] To achieve the above objectives, a first aspect of the present invention proposes a model self-evolution method, the method comprising: acquiring a context data packet; generating a candidate self-editing dataset using the large model to be evolved based on the context data packet; updating the large model to be evolved using the candidate self-editing data in the candidate self-editing dataset; evaluating downstream task models of the large model to be evolved based on the updated temporary large models; and constructing a target training dataset based on the evaluation results; training the large model to be evolved using the target training dataset; and recording the trained large model as the large model to be evolved.

[0010] According to the model self-evolution method of the present invention, the large model to be evolved actively generates structured, high-level semantic self-edited data based on the original context data it receives, and uses its own understanding ability. The downstream task model of the large model to be evolved is used to detect the quality of the candidate self-edited data generated by the large model to be evolved. The large model to be evolved uses the high-quality self-edited data and its corresponding context data for iterative updates, thereby realizing near real-time, automated and continuous learning and evolution of the large model to be evolved.

[0011] In addition, the model self-evolution method proposed in the above embodiments of the present invention may also have the following additional technical features: According to one embodiment of the present invention, the temporary large model corresponds one-to-one with the candidate self-edited data in the candidate self-edited dataset. The evaluation of the downstream task model of the large model to be evolved based on the updated temporary large models includes: for each temporary large model, inputting validation data from a preset validation dataset into the corresponding temporary large model; inputting the semantic feature vector corresponding to the validation data output by the corresponding temporary large model and the quantization parameters corresponding to the validation data in the preset validation dataset into the downstream task model to obtain a prediction result generated by the downstream task model based on the output of the corresponding temporary large model, wherein the preset validation dataset includes validation samples and real results, and the validation samples include validation data and quantization parameters; evaluating the downstream task model based on the prediction result generated by the downstream task model based on the output of the corresponding temporary large model and the real results to obtain a corresponding temporary performance score; comparing each temporary performance score with a baseline performance score, and determining the reward value corresponding to each candidate self-edited data based on the comparison result.

[0012] According to one embodiment of the present invention, the baseline performance score is obtained by evaluating the downstream task model based on the prediction result generated by the downstream task model based on the output result of the large model to be evolved and the actual result, wherein the output result of the large model to be evolved is the result obtained by inputting the verification data of the verification samples in the preset verification dataset into the large model to be evolved.

[0013] According to one embodiment of the present invention, constructing the target training dataset based on the evaluation results includes: Data pairs are generated based on the reward value and the preceding data for each candidate self-edited data; the reward value in each data pair is compared with a preset reward threshold, and data pairs with reward values ​​greater than the preset threshold are selected; the dataset composed of the preceding data and candidate self-edited data in the selected data pairs is denoted as the target training dataset.

[0014] According to an embodiment of the present invention, training the large model to be evolved using the target training dataset includes: using the context data in the target training dataset as training input and the self-edited data corresponding to the context data as training target to train the large model to be evolved.

[0015] According to an embodiment of the present invention, the method further includes: training the large model to be evolved using the target training dataset when the amount of target training data in the target training dataset reaches a preset data volume threshold, or when the time since the last iteration update of the large model to be evolved reaches a preset time threshold.

[0016] According to one embodiment of the present invention, the large model to be evolved is a multimodal large model in the transportation domain, the context data packet includes context data containing multimodal traffic information, and the downstream task model includes a traffic risk prediction model.

[0017] To achieve the above objectives, a second aspect of the present invention provides a model self-evolution apparatus, the apparatus comprising: a generation module, configured to acquire a context data packet and generate a candidate self-edited dataset using the large model to be evolved based on the context data packet; a screening and construction module, configured to update the large model to be evolved using the candidate self-edited data in the candidate self-edited dataset, evaluate the downstream task models of the large model to be evolved based on the updated temporary large models, and construct a target training dataset based on the evaluation results; and an update module, configured to train the large model to be evolved using the target training dataset, and record the trained large model as the large model to be evolved.

[0018] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the model self-evolution method as described above.

[0019] To achieve the above objectives, a fourth aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the model self-evolution method as described above.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] Figure 1 This is a flowchart of a model self-evolution method according to an embodiment of the present invention; Figure 2 This is a flowchart of evaluating a downstream task model according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the construction of a target training dataset based on evaluation results, according to one embodiment of the present invention. Figure 4 This is a flowchart of a model self-evolution method according to a specific embodiment of the present invention; Figure 5 This is a schematic diagram of a model self-evolution device according to an embodiment of the present invention; Figure 6 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0023] The model self-evolution method, apparatus, storage medium, and electronic device of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Figure 1 This is a flowchart of a model self-evolution method according to an embodiment of the present invention. Figure 1 As shown, model self-evolution methods may include: S101, Obtain the context data packet, and use the large model to be evolved to generate a candidate self-editing dataset based on the context data packet.

[0025] In this embodiment of the invention, the large model to be evolved is a pre-trained large model, which can be obtained through construction and pre-training.

[0026] The large model to be evolved in this embodiment of the invention can be a single-modal (or multi-modal) large model, that is, it has the ability to receive single-modal (or multi-modal) input and generate natural language text, and at the same time has basic scene understanding and text generation capabilities.

[0027] Implementably, a basic multimodal large model T (large model to be evolved) can be constructed and pre-trained. This basic multimodal large model T has the ability to receive multimodal input and generate natural language text, and also has basic scene understanding and text generation capabilities, such as TransGPT (multimodal large model in the transportation domain).

[0028] To improve the efficiency of large model self-evolution, relevant context data packets are obtained based on the data type processed by the large model to be evolved. For example, when the large model to be evolved is a multimodal model in the transportation domain, such as the TransGPT model, one or more context data packets containing multimodal traffic information are obtained. The obtained context data packets contain multimodal data, such as radar point clouds, images, and text.

[0029] The acquired context data is input into the large model to be evolved. Based on this context data, the large model to be evolved generates a set of candidate self-edit datasets containing k candidate self-edit data. ).

[0030] It should be noted that each candidate self-edited data in the embodiments of the present invention is a structured, high-level semantic summary of the corresponding context data.

[0031] In one embodiment of the present invention, the self-edited data can be question-and-answer pairs. For example, the data package is a large dataset including radar point clouds, images, text, and other data. The large model to be evolved generates multiple question-and-answer pairs based on the data dataset, and each question-and-answer pair is a candidate self-edited data. , of which 1 .

[0032] It should be noted that the embodiments of the present invention do not limit the form of candidate self-edited data, and corresponding self-edited data can be generated according to actual needs.

[0033] S102, using candidate self-edited data from the candidate self-edited dataset to update the large-scale model to be evolved, and evaluating the downstream task models of the large-scale model to be evolved based on the updated temporary large-scale models, and constructing the target training dataset based on the evaluation results.

[0034] Specifically, a lightweight parameter update technique, such as LoRA, can be used based on the candidate self-editing dataset ( Candidate self-edited data in ) By updating the large evolutionary model, k temporary large models can be obtained. .

[0035] Based on the updated temporary large models The downstream task model M of the large evolutionary model T is evaluated to obtain the temporary performance score of the downstream task model. Based on the evaluation results, the target self-edited data and its corresponding context data are selected, and the target training dataset is constructed based on the selected target self-edited data and its corresponding context data.

[0036] It should be noted that the large model to be evolved and its downstream task model in the embodiments of the present invention are two independent models. The downstream task model processes the downstream task based on the results generated by the large model to be evolved.

[0037] S103, use the target training dataset to train the large model to be evolved, and record the trained large model as the large model to be evolved.

[0038] Specifically, using the selected target training dataset, supervised fine-tuning training is performed on the large model to be evolved, and the trained large model is recorded as the large model to be evolved, so as to iteratively update the large model to be evolved again.

[0039] The model self-evolution method in this embodiment of the invention involves the large model to be evolved actively generating structured, high-level semantic self-edited data based on the original context data it receives, utilizing its own understanding capabilities. The downstream task model of the large model to be evolved is used to detect the quality of the candidate self-edited data generated by the large model to be evolved. The large model to be evolved uses the high-quality self-edited data and its corresponding context data for iterative updates, thereby achieving near real-time, automated, and continuous learning and evolution of the large model to be evolved.

[0040] In one embodiment of the present invention, the temporary large model corresponds one-to-one with the candidate self-edited data in the candidate self-edited dataset.

[0041] In this embodiment of the invention, a lightweight parameter update technique is employed, based on candidate self-editing data in the candidate self-editing dataset. When updating large evolutionary models, a candidate self-edited data Corresponding to a temporary large model .

[0042] In one embodiment of the present invention, such as Figure 2 As shown, the evaluation of downstream task models of the large model to be evolved, based on the updated temporary large models, may include: S201, For each temporary large model, input the validation data in the preset validation dataset into the corresponding temporary large model, and input the semantic feature vector corresponding to the validation data output by the corresponding temporary large model and the quantization parameters corresponding to the validation data in the preset validation dataset into the downstream task model to obtain the prediction result generated by the downstream task model based on the output result of the corresponding temporary large model. The preset validation dataset may include validation samples and real results, and the validation samples may include validation data and quantization parameters. S202, evaluate the downstream task model based on the prediction results and actual results generated by the downstream task model based on the output results of the corresponding temporary large model, and obtain the corresponding temporary performance score; S203 compares each provisional performance score with the baseline performance score and determines the reward value corresponding to each candidate self-edited data based on the comparison results.

[0043] It should be noted that the validation data in the validation sample can be single-modal (or multi-modal) data, which is determined by the data type processed by the large model T to be evolved.

[0044] To evaluate the utility of candidate self-edited data, embodiments of the present invention utilize a temporary large model and a downstream task model to determine the reward value of candidate self-edited data.

[0045] Specifically, using a temporary large model Execute the predefined downstream task (downstream task model M), that is, input the validation data of the validation samples in the preset validation dataset into the temporary large model. Temporary large model The validation data is processed to output the semantic feature vectors corresponding to the validation data. The temporary large model is then used. The semantic feature vector and quantization parameters corresponding to the output validation data are input into the downstream task model. The downstream task model outputs the prediction result based on the semantic feature vector and quantization parameters corresponding to the validation data. The prediction result is then generated based on the temporary large model from the downstream task model. The output results (semantic feature vectors corresponding to the validation data) are used to evaluate the predicted and actual results of the downstream task model, resulting in a quantified provisional performance score for the downstream task model M. Based on each temporary performance score Compared with baseline performance score The size relationship is used to calculate the reward value corresponding to each candidate self-edited data. .

[0046] In practice, the difference between each provisional performance score and the baseline performance score is calculated, and the corresponding difference is recorded as the reward value for each candidate self-edited data.

[0047] For example, when the performance score is 0-100 and the baseline performance score is 90, if the temporary performance score is 80, then the difference between the temporary performance score and the baseline performance score, 80-90=-10, is recorded as the reward value corresponding to the candidate self-edited data. If the temporary performance score is 95, then the difference between the temporary performance score and the baseline performance score, 95-90=5, is recorded as the reward value corresponding to the candidate self-edited data.

[0048] It should be noted that, in this embodiment of the invention, the reward value corresponding to each candidate self-edited data is not calculated based on each temporary performance score and the baseline performance score. The specific implementation method is limited, and can be set according to actual needs.

[0049] In embodiments of the present invention, the reward value corresponding to the candidate self-edited data is positively correlated with the temporary performance score. That is, the higher the temporary performance score, the greater the reward value corresponding to the candidate self-edited data. Conversely, the lower the temporary performance score, the smaller the reward value corresponding to the candidate self-edited data.

[0050] In one embodiment of the present invention, the baseline performance score can be obtained by evaluating the downstream task model based on the prediction results and the actual results generated by the downstream task model based on the output results of the large model to be evolved, wherein the output results of the large model to be evolved are the results obtained by inputting the verification data of the verification samples in the preset verification dataset into the large model to be evolved.

[0051] In this embodiment of the invention, the downstream task model M is a pre-trained task model, which can be obtained through independent construction and pre-training.

[0052] For example, when the large model to be evolved is a multimodal large model T in the transportation field, the downstream task model M is used to perform a specific, quantifiable task related to transportation business objectives, such as traffic congestion risk prediction.

[0053] In this embodiment of the invention, a fixed validation dataset is preset. The performance of the downstream task model is evaluated using this validation dataset to obtain a baseline performance score. .

[0054] This invention illustrates the specific process of obtaining baseline performance scores using a concrete example: In this specific example, the basic multimodal large model T functions to receive real-time, unstructured multimodal data (e.g., real-time video stream and radar point cloud data of a specific intersection) from a specific intersection, and outputs a structured high-level semantic understanding of the intersection scene, such as: "{"Main Event": "A scrape accident occurred in the eastbound straight lane", "Queue Length": "80 meters at the east entrance", "Traffic Light Status": "Red light at the east entrance"}.

[0055] In this specific example, the downstream task model (M) is a congestion prediction model (e.g., a spatiotemporal graph neural network, STGNN), which has been independently trained using historical data from this particular intersection. Its input is a fused feature vector that incorporates: (1) the high-level semantic understanding generated by the large model T (e.g., converted into an embedding vector through encoding); and (2) quantified traffic parameters from other data sources (e.g., raw traffic flow, average speed, etc., as statistically analyzed by upstream detectors). Its output is the probability that the intersection will enter a state of "severe congestion" (e.g., defined as congestion when the average speed is below 10 km / h) within a certain period of time (e.g., within 10 minutes).

[0056] In this specific example, the preset verification dataset is obtained by pre-collecting and solidifying historical data of the specific intersection, such as the original multimodal data of all historical peak periods in the past three months and the corresponding real traffic conditions, and processing the historical data into labeled verification samples (assuming 10,000 samples).

[0057] It should be noted that the sample structure of the verification samples in the preset verification dataset can include two parts: verification samples and real results. The verification samples can include real-time video streams, radar point cloud data, and other data for the intersection, along with the corresponding quantified traffic parameters for the relevant time period (e.g., "{"Traffic Flow": 80, "Average Speed": 25}"). The real labels (real results) are the actual congestion status labels for a subsequent period (e.g., 10 minutes) within the relevant time period (e.g., "Not Congested" or "Severely Congested").

[0058] For the imbalanced classification task of "severe congestion," the F1 score is used as a quantitative performance metric. It's important to note that the F1 score is a core metric in machine learning for evaluating the performance of classification models. It calculates the harmonic mean of precision and recall, comprehensively reflecting the model's predictive accuracy and coverage. It is often used to address the difficulty in evaluating a model when precision and recall conflict.

[0059] In calculating baseline performance scores At that time, the basic multimodal large model T processes all multimodal data in the preset validation dataset, extracts them into semantic feature vectors, and provides them to the downstream task model M along with the quantized traffic parameters for the corresponding time period. The downstream task model M infers from the sample input and outputs a congestion probability prediction value for each sample (10,000 samples). The prediction results of the downstream task model M are compared with the true labels in the validation dataset to calculate the F1 score. Assuming the F1 score is 0.82, then this F1 score of 0.82 is established as the baseline performance score of the downstream task model. .

[0060] In one embodiment of the present invention, such as Figure 3 As shown, the target training dataset can be constructed based on the evaluation results and may include: S301, Generate corresponding data pairs based on the reward values ​​corresponding to each candidate's self-edited data and the data above; S302, compare the reward value of each data pair with the preset reward threshold, and filter out the data pairs whose reward value is greater than the preset threshold; S303, the dataset consisting of the preceding data and candidate self-edited data from the selected data pairs is denoted as the target training dataset.

[0061] According to the preset screening rules and the reward values ​​corresponding to the candidate self-edited data, the embodiments of the present invention screen high-quality self-edited data and construct a high-quality dataset.

[0062] Specifically, the reward values ​​corresponding to each candidate self-edited data and the preceding data are aggregated to generate data pairs containing the context data package, candidate self-edited data, and reward values ​​(< , >, ),in, Indicates the context data packet, This indicates candidate self-edited data. This represents the reward value. The reward value of each data pair is compared with a preset reward threshold, and data pairs with reward values ​​greater than the preset threshold are selected. The preceding data and candidate self-edited data from the selected data pairs are added to a pre-established and initialized target data list L to obtain the target training dataset.

[0063] When filtering data pairs whose reward values ​​are greater than a preset threshold, the reward values ​​in each data pair can be compared with the preset reward threshold one by one to filter out the data pairs with reward values ​​greater than the preset threshold. Alternatively, the data pairs can be sorted in descending order of their reward values, and then one or more data pairs with reward values ​​greater than the preset reward threshold can be filtered out.

[0064] It should be noted that the embodiments of the present invention do not limit the method of selecting data pairs whose reward value is greater than a preset threshold.

[0065] In one embodiment of the present invention, training the large evolutionary model using the target training dataset may include: The preceding data in the target training dataset is used as the training input, and the self-edited data corresponding to the preceding data packet is used as the training target to train the large model to be evolved.

[0066] Specifically, using the target training dataset constructed above, a supervised fine-tuning training of the large model to be evolved is performed. During training, the context data packets contained in the target training dataset are used as training input, and their corresponding candidate self-edited data are used as training targets. That is, the context data in the target training dataset is input into the large model to be evolved, obtaining the predicted data output by the large model. The loss function of the large model to be evolved is calculated based on the predicted data and the candidate self-edited data corresponding to the context data, and the parameters of the large model to be evolved are updated and adjusted according to the loss function.

[0067] Once the large-scale evolutionary model has been trained, an evolved large-scale model with updated parameters will be obtained. This evolved large-scale model will serve as the next-generation large-scale model T' for processing subsequent new context data packets, and the above model self-evolution method will be repeatedly executed, thus forming a continuous self-evolutionary loop.

[0068] It should be noted that after the large model to be evolved has been trained, the target data list L should be cleared.

[0069] In one embodiment of the present invention, the model self-evolution method may further include: When the amount of target training data in the target training dataset reaches a preset data volume threshold, or when the time since the last iteration update of the large model to be evolved reaches a preset time threshold, the target training dataset is used to train the large model to be evolved.

[0070] In practice, the amount of target training data in the target training dataset is detected, and it is determined whether the amount of target training data has reached a preset data volume threshold. When the amount of target training data in the target training dataset reaches the preset data volume threshold, the large model to be evolved is triggered to undergo training iterations.

[0071] In practice, the time since the last iteration update of the large model to be evolved is calculated, and it is detected whether the time since the last iteration update has reached a preset time threshold. When the time since the last iteration update of the large model to be evolved reaches the preset time threshold, the large model to be evolved is triggered to perform training iteration.

[0072] It should be noted that the embodiments of the present invention do not limit the method of triggering the training and updating of the large model to be evolved, and can be set according to actual needs.

[0073] In one specific embodiment of the present invention, the large model to be evolved is a multimodal large model in the transportation domain, the context data packet includes context data containing multimodal traffic information, and the downstream task model includes a traffic risk prediction model.

[0074] It should be noted that the embodiments of the present invention do not specifically limit the large model to be evolved, the context data package, and the downstream task model, which can be set according to actual needs.

[0075] The following describes the model self-evolution method in this invention, taking the large model to be evolved as a multimodal large model in the transportation domain, the context data packet as a context data packet containing multimodal traffic information, and the downstream task model as a traffic risk prediction model (see [link]). Figure 4 ): Before performing self-evolution on a large multimodal model in the transportation domain, the following preparatory work is carried out: Specifically, a multimodal large-scale model for the transportation domain is initialized, and a multimodal large-scale model T for the transportation domain is constructed and pre-trained. This multimodal large-scale model T for the transportation domain has the ability to receive multimodal inputs and generate natural language text, and also has basic scene understanding and text generation capabilities.

[0076] Independently construct and train a traffic risk prediction model M. This traffic risk prediction model M is used to perform a quantifiable traffic congestion risk prediction task related to traffic business objectives.

[0077] After constructing the traffic risk prediction model M, a baseline performance score can be pre-generated based on the traffic risk prediction model M. Specifically, using a fixed validation dataset, the performance of the traffic risk prediction model M is evaluated to obtain a baseline performance score. .

[0078] It should be noted that an empty target data list L can also be initialized to store the context data and candidate self-editing data for subsequent filtering.

[0079] When performing self-evolution on a large multimodal model T in the transportation domain, one or more context data packets containing multimodal traffic information are acquired. These context data packets contain multimodal data, such as radar point clouds, images, and text.

[0080] Based on this context data packet, a candidate self-edited dataset containing k candidate self-edited data is generated using a multimodal large model T in the transportation domain. It should be noted that each candidate self-edited data is a structured, high-level semantic summary of the context data package.

[0081] Evaluate each candidate self-edited data in the candidate self-edited dataset The utility of the candidate self-edited data is calculated, and its corresponding reward value is determined. Specifically, this is based on the candidate self-edited data. A corresponding temporary model is generated using a lightweight parameter update technique (such as LoRA). .

[0082] The provisional results generated by the provisional model Ti are input into the traffic risk prediction model M. Traffic risk prediction is performed based on the provisional results generated by the provisional model Ti. The results output by the traffic risk prediction model M are then evaluated to obtain a quantified provisional performance score. .

[0083] Temporary performance score Compared with the pre-generated baseline performance score mentioned above Compare the data and calculate the self-edited data based on the comparison results. Corresponding reward value .

[0084] Filtering high-quality data and constructing the target training dataset: By summarizing the reward values ​​and the preceding data for each candidate self-edited data, we can obtain data pairs containing the context data packet, candidate self-edited data, and reward values ​​(< , >, ).

[0085] According to the preset filtering rules, one or more data pairs are selected from the data pairs in descending order of reward value. The preceding data and candidate self-edited data in the selected data pairs are added to the target data list L initialized above.

[0086] The system checks whether the multimodal large model T in the transportation domain meets the model evolution conditions (the amount of target training data in the target training dataset reaches a preset data volume threshold, or the time since the last iteration update of the multimodal large model T in the transportation domain reaches a preset time threshold). If the conditions are met, the evolution training iteration of the multimodal large model T in the transportation domain is triggered; otherwise, the target training data is acquired.

[0087] During training iterations, the traffic domain multimodal large model T uses the constructed target training dataset described above to perform supervised fine-tuning training. In this training, the context data packets contained in the target training dataset serve as training input, and their corresponding candidate self-edited data serve as training targets. After training, an evolved traffic domain multimodal large model T with updated parameters is obtained. This evolved traffic domain multimodal large model T will serve as the next-generation traffic domain multimodal large model T' for processing subsequent new context data packets, and the above steps are repeated to form a continuous self-evolutionary loop. Before the next iteration update, the target data list L is cleared to enter the next loop.

[0088] The model self-evolution method of this invention is applied to large-scale multimodal models in the transportation domain. After the trained large-scale multimodal transportation model is deployed, it establishes an intrinsic mechanism that can respond to real-time multimodal data streams and continuously update its internal knowledge and reasoning capabilities, breaking the limitation of static weights. This allows the model to break free from heavy reliance on continuous manual annotation and data processing, autonomously extracting and generating high-quality, high-information-density training data from raw, unstructured multimodal inputs. An objective feedback mechanism is established, ensuring that each self-update of the model is directed towards improving the performance of downstream real-world tasks (such as traffic congestion risk prediction), rather than aimlessly changing, thus ensuring that the "self-evolution" of the large-scale multimodal transportation model is effective and beneficial.

[0089] The model self-evolution method of this invention proposes an innovative model in which the large model to be evolved acts as a "knowledge creator." That is, based on the raw context data it receives, the large model actively generates structured, high-level semantic self-edited data using its own understanding capabilities. This transforms the model from a passive "data consumer" into an active "knowledge creator," achieving autonomous transformation from raw data to high-value training signals.

[0090] This invention proposes a self-evolving closed-loop scheme with the ultimate goal of improving practical application performance. The quality of each "self-edited data" is not determined subjectively, but rather through an objective quantitative process: it is used to temporarily fine-tune an independent downstream task model and the performance improvement of that model is detected. This performance improvement serves as a reward signal, which is used to optimize the main model's "self-editing" generation strategy through a reinforcement learning algorithm. This "generation → verification → reward → evolution" cycle ensures that the model's evolutionary direction always aligns with the goal of improving practical application value.

[0091] The model self-evolution method in this embodiment of the invention transforms a static pre-trained model into a dynamic system that can continuously learn and evolve, enabling the model to iterate itself continuously through continuous interaction with new data, thus overcoming the problem of performance degradation in dynamic environments.

[0092] The model self-evolution method in this embodiment of the invention creates an automated data generation engine by autonomously generating "self-editing data" through the model, which greatly reduces the long-term dependence on expensive and time-consuming manual data annotation and solves the bottleneck problem of data iteration.

[0093] The model self-evolution method in this embodiment of the invention uses the performance improvement of specific downstream tasks as the reward signal for reinforcement learning, ensuring that each "evolution" of the model is purposeful and beneficial, and guaranteeing that the iteration of the model always proceeds in the right direction of solving practical problems and creating greater application value, thus avoiding invalid or harmful updates.

[0094] This invention provides a model self-evolution device.

[0095] Figure 5 This is a schematic diagram of a model self-evolution device according to an embodiment of the present invention, as shown below. Figure 5 As shown, the model self-evolution device 100 may include a generation module 10, a screening and construction module 20, and an update module 30.

[0096] The generation module 10 is used to obtain context data packets and generate candidate self-editing datasets based on the context data packets using the large model to be evolved.

[0097] In this embodiment of the invention, the large model to be evolved is a pre-trained large model, which can be obtained through construction and pre-training.

[0098] The large model to be evolved in this embodiment of the invention can be a single-modal (or multi-modal) large model, that is, it has the ability to receive single-modal (or multi-modal) input and generate natural language text, and at the same time has basic scene understanding and text generation capabilities.

[0099] Implementably, a basic multimodal large model T (large model to be evolved) can be constructed and pre-trained. This basic multimodal large model T has the ability to receive multimodal input and generate natural language text, and also has basic scene understanding and text generation capabilities, such as TransGPT (multimodal large model in the transportation domain).

[0100] To improve the efficiency of large model self-evolution, relevant context data packets are obtained based on the data type processed by the large model to be evolved. For example, when the large model to be evolved is a multimodal model in the transportation domain, such as the TransGPT model, one or more context data packets containing multimodal traffic information are obtained. The obtained context data packets contain multimodal data, such as radar point clouds, images, and text.

[0101] The acquired context data is input into the large model to be evolved. Based on this context data, the large model to be evolved generates a set of candidate self-edit datasets containing k candidate self-edit data. ).

[0102] It should be noted that each candidate self-edited data in the embodiments of the present invention is a structured, high-level semantic summary of the corresponding context data.

[0103] In one embodiment of the present invention, the self-edited data can be question-and-answer pairs. For example, the data package is a large dataset including radar point clouds, images, text, and other data. The large model T generates multiple question-and-answer pairs based on the data dataset, and each question-and-answer pair is a candidate self-edited data. , of which 1 .

[0104] It should be noted that the embodiments of the present invention do not limit the form of candidate self-edited data, and corresponding self-edited data can be generated according to actual needs.

[0105] The screening and building module 20 is used to update the large model to be evolved using candidate self-edited data in the candidate self-edited dataset, evaluate the downstream task models of the large model to be evolved based on the updated temporary large models, and build the target training dataset based on the evaluation results.

[0106] Specifically, a lightweight parameter update technique, such as LoRA, can be used based on the candidate self-editing dataset ( Candidate self-edited data in ) By updating the large evolutionary model, k temporary large models can be obtained. .

[0107] Based on the updated temporary large models The downstream task model M of the large evolutionary model T is evaluated to obtain the temporary performance score of the downstream task model. Based on the evaluation results, the target self-edited data and its corresponding context data are selected, and the target training dataset is constructed based on the selected target self-edited data and its corresponding context data.

[0108] It should be noted that the large model to be evolved and its downstream task model in the embodiments of the present invention are two independent models. The downstream task model processes the downstream task based on the results generated by the large model to be evolved.

[0109] The update module 30 is used to train the large model to be evolved using the target training dataset, and the trained large model is recorded as the large model to be evolved.

[0110] Specifically, using the selected target training dataset, supervised fine-tuning training is performed on the large model to be evolved, and the trained large model is recorded as the large model to be evolved, so as to iteratively update the large model to be evolved again.

[0111] The model self-evolution method in this embodiment of the invention involves the large model to be evolved actively generating structured, high-level semantic self-edited data based on the original context data it receives, utilizing its own understanding capabilities. The downstream task model of the large model to be evolved is used to detect the quality of the candidate self-edited data generated by the large model to be evolved. The large model to be evolved uses the high-quality self-edited data and its corresponding context data for iterative updates, thereby achieving near real-time, automated, and continuous learning and evolution of the large model to be evolved.

[0112] In one embodiment of the present invention, the temporary large model corresponds one-to-one with the candidate self-edited data in the candidate self-edited dataset.

[0113] A candidate self-editing data in an embodiment of the present invention Corresponding to a temporary large model .

[0114] In one embodiment of the present invention, the screening and construction module 20 is used to input the verification data in the preset verification dataset into the corresponding temporary large model for each temporary large model, and input the semantic feature vector corresponding to the verification data output by the corresponding temporary large model and the quantization parameters corresponding to the verification data in the preset verification dataset into the downstream task model to obtain the prediction result generated by the downstream task model based on the output result of the corresponding temporary large model. The preset verification dataset may include verification samples and real results, and the verification samples may include verification data and quantization parameters. The downstream task model is evaluated based on the prediction result generated by the downstream task model based on the output result of the corresponding temporary large model and the real result to obtain the corresponding temporary performance score. Each temporary performance score is compared with the baseline performance score, and the reward value corresponding to each candidate self-edited data is determined based on the comparison result.

[0115] To evaluate the utility of candidate self-edited data, embodiments of the present invention utilize a temporary large model and a downstream task model to determine the reward value of candidate self-edited data.

[0116] Specifically, using a temporary large model Execute the predefined downstream task (downstream task model M), that is, input the validation data of the validation samples in the preset validation dataset into the temporary large model. Temporary large model The validation data is processed to output the semantic feature vectors corresponding to the validation data. The temporary large model is then used. The semantic feature vector and quantization parameters corresponding to the output validation data are input into the downstream task model. The downstream task model outputs the prediction result based on the semantic feature vector and quantization parameters corresponding to the validation data. The prediction result is then generated based on the temporary large model from the downstream task model. The output results (semantic feature vectors corresponding to the validation data) are used to evaluate the predicted and actual results of the downstream task model, resulting in a quantified provisional performance score for the downstream task model M. Based on each temporary performance score Compared with baseline performance score The size relationship is used to calculate the reward value corresponding to each candidate self-edited data. .

[0117] In practice, the difference between each provisional performance score and the baseline performance score is calculated, and the corresponding difference is recorded as the reward value for each candidate self-edited data.

[0118] It should be noted that, in this embodiment of the invention, the reward value corresponding to each candidate self-edited data is not calculated based on each temporary performance score and the baseline performance score. The specific implementation method is limited, and can be set according to actual needs.

[0119] In embodiments of the present invention, the reward value corresponding to the candidate self-edited data is positively correlated with the temporary performance score. That is, the higher the temporary performance score, the greater the reward value corresponding to the candidate self-edited data. Conversely, the lower the temporary performance score, the smaller the reward value corresponding to the candidate self-edited data.

[0120] In one embodiment of the present invention, the baseline performance score is obtained by evaluating the downstream task model based on the prediction results and the actual results generated by the downstream task model based on the output results of the large model to be evolved. The output results of the large model to be evolved are the results obtained by inputting the verification data of the verification samples in the preset verification dataset into the large model to be evolved.

[0121] In this embodiment of the invention, the downstream task model M is a pre-trained task model, which can be obtained through independent construction and pre-training.

[0122] For example, when the large model to be evolved is a multimodal large model T in the transportation field, the downstream task model M is used to perform a specific, quantifiable task related to transportation business objectives, such as traffic congestion risk prediction.

[0123] In this embodiment of the invention, a fixed validation dataset is preset. The performance of the downstream task model is evaluated using this validation dataset to obtain a baseline performance score. .

[0124] In one embodiment of the present invention, the screening construction module 20 is used to generate corresponding data pairs based on the reward value and the preceding text data of each candidate self-edited data; compare the reward value of each data pair with a preset reward threshold, and screen out data pairs with reward values ​​greater than the preset threshold; and denot the dataset composed of the preceding text data and candidate self-edited data in the screened data pairs as the target training dataset.

[0125] According to the preset screening rules and the reward values ​​corresponding to the candidate self-edited data, the embodiments of the present invention screen high-quality self-edited data and construct a high-quality dataset.

[0126] Specifically, the reward values ​​corresponding to each candidate self-edited data and the preceding data are aggregated to generate data pairs containing the context data package, candidate self-edited data, and reward values ​​(< , >, ),in, Indicates the context data packet, This indicates candidate self-edited data. This represents the reward value. The reward value of each data pair is compared with a preset reward threshold, and data pairs with reward values ​​greater than the preset threshold are selected. The preceding data and candidate self-edited data from the selected data pairs are added to a pre-established and initialized target data list L to obtain the target training dataset.

[0127] When filtering data pairs whose reward values ​​are greater than a preset threshold, the reward values ​​in each data pair can be compared with the preset reward threshold one by one to filter out the data pairs with reward values ​​greater than the preset threshold. Alternatively, the data pairs can be sorted in descending order of their reward values, and then one or more data pairs with reward values ​​greater than the preset reward threshold can be filtered out.

[0128] It should be noted that the embodiments of the present invention do not limit the method of selecting data pairs whose reward value is greater than a preset threshold.

[0129] In one embodiment of the present invention, the update module 30 is used to train the large model to be evolved by taking the context data in the target training dataset as the training input and the self-edited data corresponding to the context data packet as the training target.

[0130] Specifically, using the target training dataset constructed above, a supervised fine-tuning training of the large model to be evolved is performed. During training, the context data packets contained in the target training dataset are used as training input, and their corresponding candidate self-edited data are used as training targets. That is, the context data in the target training dataset is input into the large model to be evolved, obtaining the predicted data output by the large model. The loss function of the large model to be evolved is calculated based on the predicted data and the candidate self-edited data corresponding to the context data, and the parameters of the large model to be evolved are updated and adjusted according to the loss function.

[0131] Once the large-scale evolutionary model has been trained, an evolved large-scale model with updated parameters will be obtained. This evolved large-scale model will serve as the next-generation large-scale model T' for processing subsequent new context data packets, and the above model self-evolution method will be repeatedly executed, thus forming a continuous self-evolutionary loop.

[0132] It should be noted that after the large model to be evolved has been trained, the target data list L should be cleared.

[0133] In one embodiment of the present invention, the update module 30 is used to train the large model to be evolved using the target training dataset when the amount of target training data in the target training dataset reaches a preset data volume threshold, or when the time since the last iteration update of the large model to be evolved reaches a preset time threshold.

[0134] In practice, the amount of target training data in the target training dataset is detected, and it is determined whether the amount of target training data has reached a preset data volume threshold. When the amount of target training data in the target training dataset reaches the preset data volume threshold, the large model to be evolved is triggered to undergo training iterations.

[0135] In practice, the time since the last iteration update of the large model to be evolved is calculated, and it is detected whether the time since the last iteration update has reached a preset time threshold. When the time since the last iteration update of the large model to be evolved reaches the preset time threshold, the large model to be evolved is triggered to perform training iteration.

[0136] It should be noted that the embodiments of the present invention do not limit the method of triggering the training and updating of the large model to be evolved, and can be set according to actual needs.

[0137] In one embodiment of the present invention, the large model to be evolved is a multimodal large model in the transportation domain, the context data packet includes context data containing multimodal traffic information, and the downstream task model includes a traffic risk prediction model.

[0138] It should be noted that other specific embodiments of the model self-evolution device provided in the embodiments of the present invention can be found in other specific embodiments of the model self-evolution method of the above embodiments of the present invention.

[0139] The model self-evolution device of this invention transforms a static pre-trained model into a dynamic system capable of continuous learning and evolution, enabling the model to iterate itself continuously through continuous interaction with new data, thus overcoming the problem of performance degradation in dynamic environments.

[0140] The model self-evolution device in this embodiment of the invention creates an automated data generation engine by autonomously generating "self-editing data" through the model, which greatly reduces the long-term dependence on expensive and time-consuming manual data annotation and solves the bottleneck problem of data iteration.

[0141] The model self-evolution device in this embodiment of the invention uses the performance improvement of specific downstream tasks as the reward signal for reinforcement learning, ensuring that each "evolution" of the model is purposeful and beneficial, and guaranteeing that the iteration of the model always proceeds in the right direction of solving practical problems and creating greater application value, thus avoiding invalid or harmful updates.

[0142] This invention provides a computer-readable storage medium.

[0143] In this embodiment, a computer program is stored on a computer-readable storage medium, and when the computer program is executed by a processor, it implements the model self-evolution method as described above.

[0144] This invention provides an electronic device.

[0145] In this embodiment, the electronic device may include a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, it implements the model self-evolution method as described above.

[0146] Figure 6 This is a structural block diagram of an electronic device according to an embodiment of the present invention.

[0147] like Figure 6 As shown, the electronic device 500 includes a processor 501 and a memory 503. The processor 501 and the memory 503 are connected, for example, via a bus 502. Optionally, the electronic device 500 may also include a transceiver 504. It should be noted that in practical applications, the transceiver 504 is not limited to one type, and the structure of this electronic device 500 does not constitute a limitation on the embodiments of the present invention.

[0148] Processor 501 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 501 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0149] Bus 502 may include a pathway for transmitting information between the aforementioned components. Bus 502 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 502 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0150] The memory 503 is used to store a computer program corresponding to the model self-evolution method of the above embodiments of the present invention. This computer program is controlled and executed by the processor 501. The processor 501 is used to execute the computer program stored in the memory 503 to implement the content shown in the foregoing method embodiments.

[0151] Among them, electronic devices 500 include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 6 The electronic device 500 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0152] The computer-readable storage medium and electronic device of the present invention, based on the above-described model self-evolution method, can transform a static pre-trained model into a dynamic system capable of continuous learning and evolution, enabling the model to continuously iterate itself through continuous interaction with new data, thus overcoming the problem of performance degradation in dynamic environments.

[0153] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0154] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0155] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0156] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0157] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0158] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0159] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0160] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method of model self-evolution, characterized by, The method comprises: acquiring a context data package, and generating a candidate self-editing data set by a large model to be evolved according to the context data package; updating the large model to be evolved by using candidate self-editing data in the candidate self-editing data set, evaluating a downstream task model of the large model to be evolved based on each temporary large model after updating, and constructing a target training data set according to an evaluation result; training the large model to be evolved by using the target training data set, and recording the large model after training as the large model to be evolved.

2. The method of model self-evolution of claim 1, wherein, The temporary large model corresponds to the candidate self-editing data in the candidate self-editing data set one by one, and the evaluation of the downstream task model of the large model to be evolved based on each temporary large model after updating comprises: for each temporary large model, inputting verification data in a preset verification data set into the corresponding temporary large model, inputting semantic feature vectors corresponding to the verification data output by the corresponding temporary large model and quantization parameters corresponding to the verification data in the preset verification data set into the downstream task model, and obtaining a prediction result generated by the downstream task model based on the output result of the corresponding temporary large model, wherein the preset verification data set comprises verification samples and real results, and the verification samples comprise verification data and quantization parameters; evaluating the downstream task model based on the prediction result generated by the downstream task model based on the output result of the corresponding temporary large model and the real result, and obtaining a corresponding temporary performance score; comparing each temporary performance score with a baseline performance score, and determining a reward value corresponding to each candidate self-editing data according to a comparison result.

3. The method of model self-evolution of claim 2, wherein, The baseline performance score is obtained by evaluating the downstream task model based on a prediction result generated by the downstream task model based on an output result of the large model to be evolved and the real result, wherein the output result of the large model to be evolved is a result obtained by inputting verification data of a verification sample in the preset verification data set into the large model to be evolved.

4. The method of model self-evolution of claim 1, wherein, The construction of the target training data set according to the evaluation result comprises: generating a corresponding data pair according to the reward value corresponding to each candidate self-editing data and the context data; comparing the reward value in each data pair with a preset reward threshold, and screening out a data pair with a reward value greater than the preset threshold; a data set composed of the context data and the candidate self-editing data in the screened data pair is recorded as the target training data set.

5. The method of model self-evolution of claim 4, wherein, The training of the large model to be evolved by using the target training data set comprises: training the large model to be evolved by using the context data in the target training data set as a training input and the self-editing data corresponding to the context data package as a training target.

6. The method of model self-evolution of claim 5, wherein, The method further comprises: training the large model to be evolved by using the target training data set when the number of target training data in the target training data set reaches a preset data amount threshold or the time interval from the previous iteration update of the large model to be evolved reaches a preset time threshold.

7. The method of model evolution of any of claims 1-6, wherein, The large model to be evolved is a multi-modal large model in the transportation field, the context data packet includes context data containing multi-modal transportation information, and the downstream task model includes a traffic risk prediction model.

8. A model self-evolution device characterized by comprising: The device comprises: A generation module is configured to obtain a context data packet and generate a candidate self-editing dataset from the context data packet by using a large model to be evolved; A screening and construction module is configured to update the large model to be evolved by using candidate self-editing data in the candidate self-editing dataset, evaluate a downstream task model of the large model to be evolved based on each temporary large model after the update, and construct a target training dataset according to an evaluation result; An update module is configured to train the large model to be evolved by using the target training dataset, and record the large model after the training as the large model to be evolved.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the model self-evolution method according to any one of claims 1-7.

10. An electronic device comprising a memory, a processor, the memory having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the model self-evolution method according to any one of claims 1-7.

Images