Data processing strategy generation method, model training method, equipment and storage medium

Abstract

The embodiment of the invention provides a data processing strategy generation method, a model training method, equipment and a storage medium. The method comprises the steps of obtaining an original data set and a plurality of operators; generating an initial strategy set based on the plurality of operators through a large language model, wherein each strategy in the initial strategy set comprises a plurality of operators in the plurality of operators and an execution sequence thereof; performing iterative adjustment on the initial strategy set based on the original data set until a target strategy suitable for the original data set is obtained, each round of iteration comprising: processing the original data set by using each strategy in the current strategy set, and obtaining an evaluation index of each strategy, the evaluation index represents a processing effect on the original data set, feeding back each strategy in the current strategy set and the evaluation index thereof to the large language model, and adjusting each strategy in the current strategy set through the large language model.

Description

TECHNICAL FIELD

[0001] The present specification relates to the technical field of data processing, and particularly relates to a data processing strategy generation method and system and a storage medium. BACKGROUND

[0002] With the development of data processing technology, the effect of using data to perform various tasks is increasingly required.

[0003] Generally, the collected raw data will contain a lot of low-quality data that affects the task execution. For example, in the model training scenario, sample data needs to be collected first to perform the model training task using the sample data. If the collected raw sample data is directly used for model training, the quality of the model obtained by training will be poor due to the low-quality data in the raw sample data.

[0004] In order to reduce the influence of low-quality data, before actually performing the task, relevant personnel need to manually analyze the raw data to determine the operator suitable for processing the raw data, and then call the corresponding operator to clean the raw data. However, the above method relies on manual experience, and the operator selected manually may not be suitable for processing the raw data, resulting in poor processing effect. Therefore, how to automatically determine the data processing strategy that is truly suitable for the raw data is a technical problem that needs to be considered.

[0005] The content of the background section only represents the inventor's own knowledge and does not mean that the above information has entered the public domain before the filing date of the present disclosure, nor does it mean that it can be prior art of the present disclosure. SUMMARY

[0006] The present specification provides a data processing strategy generation method, a model training method, an equipment and a storage medium, which can automatically and efficiently determine the truly suitable data processing strategy for the raw data set, thereby improving the data processing effect and ensuring the data privacy of the raw data set.

[0007] In a first aspect, the present specification provides a method for generating a data processing strategy, comprising: obtaining an original data set and a plurality of preset operators; generating an initial strategy set based on the plurality of operators by a large language model, each strategy in the initial strategy set comprising a plurality of operators in the plurality of operators and an execution order of the plurality of operators; and iteratively adjusting the initial strategy set based on the original data set until a target strategy suitable for the original data set is obtained, wherein each iteration comprises: processing the original data set by each strategy in a current strategy set to obtain an evaluation index of each strategy, the evaluation index representing a processing effect on the original data set, the current strategy set being the initial strategy set or a strategy set after adjustment in a previous iteration, and feeding each strategy in the current strategy set and the evaluation index of each strategy to the large language model, and adjusting each strategy in the current strategy set by the large language model.

[0008] In a second aspect, the present specification also provides a model training method, comprising: obtaining an original data set for training a target model; generating a target strategy suitable for the original data set, wherein the target strategy is generated based on the method of the first aspect; processing the original data set by the target strategy to obtain a training sample set; and training the target model by the training sample set.

[0009] In a third aspect, the present specification also provides an electronic device, comprising: at least one storage medium storing at least one instruction set; and at least one processor in communication connection with the at least one storage medium, wherein the at least one processor reads the at least one instruction set and executes the method of the first aspect or the second aspect according to an indication of the at least one instruction set.

[0010] In a fourth aspect, the present specification also provides a computer-readable non-volatile storage medium, wherein the computer-readable non-volatile storage medium stores at least one instruction set, and the at least one instruction set is executed by at least one processor to implement the method of the first aspect or the second aspect.

[0011] The data processing strategy generation method, model training method, device and storage medium provided by the specification combine operators to obtain a strategy, and optimize and adjust the strategy based on an original data set, which can better determine the influence of different combination modes of operators on the data processing effect, and ensure that the obtained target strategy is truly applicable to the original data set and has a good data processing effect on the original data set. Using a large language model to generate and adjust the strategy can more accurately and efficiently determine the target strategy applicable to the original data set. Moreover, such a method does not require manual checking of the specific content of the original data set, and can determine the data processing strategy suitable for the original data in a black box environment, ensuring the data privacy of the original data set.

[0012] Other functions of the data processing strategy generation method, model training method, device and storage medium provided by the specification will be partially listed in the following description. The creative aspects of the data processing strategy generation method, model training method, device and storage medium provided by the specification can be fully explained by practicing or using the methods, devices and combinations described in the following detailed examples. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the specification, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the specification, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0014] Figure 1 A schematic diagram of a model training scene is shown according to an embodiment of the specification;

[0015] Figure 2 A hardware structure diagram of an electronic device is shown according to an embodiment of the specification;

[0016] Figure 3 A flowchart of a data processing strategy generation method is shown according to an embodiment of the specification;

[0017] Figure 4 A simple block diagram of a strategy generation method is shown according to an embodiment of the specification;

[0018] Figure 5 A simple flowchart of strategy evaluation is shown according to an embodiment of the specification;

[0019] Figure 6 A simple flowchart of sampling an original data set is shown according to an embodiment of the specification;

[0020] Figure 7 A simple flowchart of processing an original dataset with a first policy is shown according to an embodiment of the present specification;

[0021] Figure 8 A schematic diagram of policy storage is shown according to an embodiment of the present specification; and

[0022] Figure 9 A flowchart of a model training method is shown according to an embodiment of the present specification. DETAILED DESCRIPTION

[0023] The following description provides specific applications and requirements of the present specification, in order to enable a person skilled in the art to manufacture and use the contents of the present specification. Various local modifications of the disclosed embodiments are obvious to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of the present specification. Therefore, the present specification is not limited to the embodiments shown, but is consistent with the widest scope of the claims.

[0024] The terms used herein are only for the purpose of describing specific example embodiments, and are not limiting. For example, unless the context clearly indicates otherwise, as used herein, the singular forms "a", "an" and "the" can also include the plural forms. When used in the present specification, the terms "comprise", "include" and / or "contain" mean that the associated whole, step, operation, element and / or component exists, but does not exclude the presence of one or more other features, whole, step, operation, element, component and / or group, or additional features, whole, step, operation, element, component and / or group can be added in the system / method.

[0025] These features of the present specification and other features, as well as the operation and function of related elements of the structure, and the economy of combination and manufacture of components can be significantly improved in view of the following description. With reference to the drawings, all of which form part of the present specification. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of the present specification. It should also be understood that the drawings are not drawn to scale.

[0026] The flowchart used in the present specification shows the operation of system implementation according to some embodiments of the present specification. It should be clearly understood that the operations of the flowchart can not be implemented in sequence. On the contrary, the operations can be implemented in reverse order or simultaneously. In addition, one or more other operations can be added to the flowchart. One or more operations can be removed from the flowchart.

[0027] In this specification, a large language model (LLM) can also be referred to simply as a large model. A large language model is a natural language processing model based on deep learning technology, with a parameter order of magnitude usually reaching tens of billions to hundreds of billions or even higher, and with strong language understanding and generation capabilities. A large language model can use a Transformer architecture or its variants (such as GPT, BERT, etc.), which uses attention mechanisms to model global sequence data and can efficiently handle long-range dependencies, thus performing well in natural language tasks. A large language model is pre-trained on a large corpus of text to learn statistical features and semantic relationships of language, making it have good generalization ability. The core capabilities of a large language model include but are not limited to understanding contextual semantics, generating coherent and grammatically correct text, performing logical reasoning, and handling multi-task scenarios. Its usage methods usually include direct inference and fine-tuning. In the direct inference mode, users guide the large language model to generate specific outputs by designing prompts. Prompts can be text-based task descriptions or instructions to stimulate the semantic understanding and generation capabilities of the large language model. In the fine-tuning mode, the large language model is further trained on a small dataset in a specific domain to optimize its performance on specific tasks. The strong generalization ability and flexibility of a large language model make it an important tool in the field of artificial intelligence technology, providing efficient and accurate solutions for automated text generation and understanding.

[0028] In some embodiments, a large language model can also have understanding and generation capabilities for other modalities (such as vision, audio, etc.) data, in which case the large language model can also be referred to as a multimodal large language model (MLLM). MLLMs provide a richer and more natural interactive experience by integrating text, images, sounds, and other types of input and output. The core advantage of MLLMs is their ability to process and understand information from different modalities and integrate these information to complete complex tasks. For example, MLLMs can analyze a picture and generate descriptive text, or generate corresponding images based on text descriptions. This cross-modal understanding and generation capability makes MLLMs have broad application prospects in many fields.

[0029] It should be noted that the key technologies of large language models can be found in the detailed description in the paper "A Survey of Large Language Models" (paper number: arXiv:2303.18223v16, published on March 11, 2025, public link: https: / / doi.org / 10.48550 / arXiv.2303.18223), and will not be repeated here.

[0030] This specification provides a method for generating data processing (DP) strategies. These strategies are used to generate suitable data processing strategies for the data to be processed, thereby improving the data processing effectiveness. The data processing strategies described here can be used to cleanse data, correcting or removing erroneous, incomplete, duplicate, or irrelevant parts, thus improving data quality and usability. This ensures that the processed data is of high quality and can be better used for subsequent tasks. In some scenarios, data processing strategies can be used to perform other data processing, such as data augmentation, data feature extraction, data transformation, or data analysis. The operators obtained when generating the data processing strategy may differ depending on the type of data processing.

[0031] The data processing strategy generation method provided in this specification can be applied to various scenarios. While the original datasets differ across scenarios, the execution process remains the same. For example, this method can be applied to model training scenarios, where the original dataset is the raw sample dataset required for model training, including multiple labeled data samples. This method can also be applied to device or environmental monitoring scenarios, where the original dataset includes data collected by sensors in the device or environment. This method can also be applied to report generation scenarios, where the original dataset includes data collected by scientific instruments or generated by simulation software. This method can also be applied to video post-processing scenarios, where the original dataset includes video directly captured by a camera.

[0032] The following example illustrates the generation method using the data processing strategy provided in this manual in a model training scenario.

[0033] In a model training scenario, an initial model needs to be trained by using a training sample set until the output of the model meets the requirements, and a trained model is obtained. Each training sample in the training sample set needs to match the function corresponding to the model to avoid a large amount of redundant, irrelevant, or incorrect data, so as to ensure good training effect of the model and high training efficiency. There are two problems to be solved in the model training scenario. One is that the original data set provided by the data provider may contain a large amount of noise and a large amount of low-quality and unusable data, which cannot be directly used for model training. The second is that the data provided by some data providers (such as confidential units or some special institutions) is high-privacy data, which needs to be processed in a ciphertext environment to avoid privacy leakage of the data provider.

[0034] The above two problems may cause the data to be introduced into the model training system only, and the data cannot be normally used out. Therefore, a processing strategy suitable for the original data set needs to be determined in a ciphertext environment to process the original data set to obtain clean, high-quality, and usable data. The ciphertext environment refers to a black box environment in which any information and content of the data cannot be obtained or viewed, that is, a data-invisible environment.

[0035] Generally, the original data set needs to be analyzed manually to determine the operators suitable for processing the original data set according to experience, and the operators are called to process the data. In this way, the data content in the original data set needs to be displayed, which cannot meet the privacy requirement of the original data set. Since the manual experience is difficult to accurately match all original data sets, the accuracy of selecting operators for the original data set is low, which may lead to poor data processing results of the original data set. In addition, manual data analysis and operator selection consume high labor cost, and the overall efficiency of data processing is low.

[0036] In a related technology, an original data set is preprocessed by using a model data processing framework containing a plurality of data processing operators. When data processing is performed, the original data set still needs to be analyzed manually, and the operators to be used in each processing stage are determined based on the analysis result, and then the model is instructed to call the corresponding operators in each stage. This data processing method still needs to be performed in a plaintext environment and cannot be used for high-privacy original data sets. In this related technology, the operators are isolated from each other and are not connected, and the data processing ability brought by each operator is difficult to be fully utilized by using the operators in isolation, and the data processing effect still needs to be improved.

[0037] The specification provides a data processing strategy generation method. For a to-be-processed original data set, different operators can be associated (combined) by using a large language model, and an applicable data processing strategy can be obtained by iterative adjustment. In this way, the operators are combined to obtain a strategy and the strategy is adjusted, which can better exert the overall data processing capability of each operator and ensure that the obtained target strategy is truly applicable to the original data set and has a good data processing effect on the original data set. The large language model is used to generate and adjust the strategy, which can accurately and efficiently determine the target strategy applicable to the original data set. Moreover, the specific content of the original data set does not need to be checked, and the data processing strategy applicable to the original data set can be determined in a black box environment, ensuring the data privacy of the original data set.

[0038] The specification also provides a model training method. The above-mentioned data processing strategy generation method can be applied to a model training scenario to generate a corresponding target strategy for an original data set used to train a target model. Furthermore, the original data set is processed by using the target strategy to obtain a training sample set, and the target model is trained by using the training sample set. Since the obtained target strategy is adapted to the original data set and can achieve a good data processing effect on the original data set, the training sample set obtained has high cleanliness and usability, and the training effect of the target model by using the training sample set is good, and the quality of the obtained target model is good.

[0039] Figure 1 A schematic diagram of a model training scenario, i.e., a schematic diagram of an application scenario 100 of the data processing strategy generation method, is shown according to an embodiment of the specification. As shown in the figure, the application scenario 100 can include a data providing device 110, a strategy generation device 120, a model training device 130, and a network 140. Figure 1

[0040] The data providing device 110 can be a terminal, and the strategy generation device 120 and the model training device 130 can be servers. The terminal is a device that performs a corresponding task in response to a user operation. The server can be a server that provides various services, such as a background server that provides support for data processing tasks required by the terminal. The server can be in communication connection with multiple terminals and receive data sent by the terminals. The data providing device 110 and the strategy generation device 120 can be connected through the network 140, and the strategy generation device 120 and the model training device 130 can also be connected through the network 140. The network 140 at the two places can be the same network or different networks.

[0041] ​The generation method of the data processing strategy performed by the strategy generation device 120 is taken as an example for illustration in the specification. The data providing device 110 sends an original data set for model training to the strategy generation device 120. The strategy generation device 120 can store instructions for performing the generation method of the data processing strategy and can execute or be configured to execute the instructions to obtain a target strategy applicable to the original data set. In some embodiments, the original data set can also be provided by a database. The strategy generation device 120 can send the obtained target strategy to the model training device 130. The model training device 130 can store instructions for performing the model training method and can execute or be configured to execute the instructions. The model training device 130 can preprocess the original data set based on the obtained target strategy and perform model training using the processed data to obtain a model with higher quality.

[0042] In some embodiments, the strategy generation device 120 and the model training device 130 can be integrated in one device, which executes the generation method of the data processing strategy provided in the specification and performs the model training method using the obtained target strategy to perform model training.

[0043] In some embodiments, the strategy generation device 120 can be integrated with the data providing device 110, which directly executes the generation method of the data processing strategy provided in the specification to obtain a target strategy. The data providing device 110 can also process the original data set using the obtained target strategy and provide the processed data set to the model training device 130 for model training. In this way, the data providing device 110 can store instructions for performing the generation method of the data processing strategy described in the specification and can execute or be configured to execute the instructions. The data providing device 110 can include a hardware device with data information processing function and necessary programs required to drive the hardware device to work.

[0044] The terminal described in the specification can include a mobile device, a tablet computer, a notebook computer, a built-in device of a motor vehicle, or the like, or any combination thereof. In some embodiments, the mobile device can include a smart home device, a smart mobile device, a virtual reality device, an augmented reality device, or the like, or any combination thereof. In some embodiments, the smart home device can include a smart television, a desktop computer, or the like, or any combination thereof. In some embodiments, the smart mobile device can include a smartphone, a personal digital assistant, a gaming device, a navigation device, or the like, or any combination thereof. In some embodiments, the virtual reality device or the augmented reality device can include a virtual reality headset, a virtual reality glasses, a virtual reality patch, an augmented reality headset, an augmented reality glasses, an augmented reality patch, or the like, or any combination thereof. For example, the virtual reality device or the augmented reality device can include Google glasses, a head-mounted display, VR, or the like. In some embodiments, the built-in device in the motor vehicle can include an on-board computer, an on-board television, or the like. In some embodiments, the terminal can include an image acquisition device for acquiring video or image information in a target space to obtain a target image. In some embodiments, the image acquisition device can be a two-dimensional image acquisition device (such as an RGB camera), or a two-dimensional image acquisition device (such as an RGB camera) and a depth image acquisition device (such as a 3D structured light camera, a laser detector, or the like). In some embodiments, the terminal can be a device with positioning technology for positioning the location of the terminal.

[0045] In some embodiments, the terminal can be installed with one or more application programs (APPs). The APPs can provide the user with the ability to interact with the outside world through the network and the interface. For example, the terminal can use the APP to provide the original data set to the policy generation device 120. The APPs include but are not limited to: web browser type APP programs, search type APP programs, chat type APP programs, shopping type APP programs, video type APP programs, financial management type APP programs, instant messaging tools, mailbox clients, social platform software, and the like. In some embodiments, the terminal can be installed with a target APP, and the user of the terminal can trigger a model training request through the target APP. The target APP can send the model training request and the original data set to the policy generation device 120 and the model training device 130 to trigger the policy generation device 120 to perform the generation method of the data processing policy described in the specification and trigger the model training device 130 to perform the model training method described in the specification.

[0046] The network described in the specification is a medium for providing a communication connection between devices (such as between a terminal and a server, between servers). The network can facilitate the exchange of information or data. For example, Figure 1As shown, the devices can be connected with a network, and transmit information or data to each other through the network. In some embodiments, the network can be any type of wired or wireless network, or a combination thereof. For example, the network can include a cable network, a wired network, a fiber-optic network, a telecommunication communication network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a public switched telephone network (PSTN), a Bluetooth network, a ZigBee network, a near-field communication (NFC) network, or the like. In some embodiments, the network can include one or more network access points. For example, the network can include wired or wireless network access points, such as base stations or Internet exchange points, through which one or more components of the terminal and the server can be connected to the network to exchange data or information.

[0047] It should be understood that Figure 1 The number of terminals, servers and networks in the above description is only illustrative. Any number of terminals, servers and networks can be provided according to the needs of implementation. The data processing strategy generation method described in the specification can be completely executed on the terminal, completely executed on the server, partially executed on the terminal and partially executed on the server.

[0048] It should be noted that the user data obtained by the specification is authorized by the user and does not involve user privacy. The data processing strategy generation scheme provided by the specification can be applied to various use scenarios, and the model training scenario described above is only one of the various use scenarios. In order to facilitate understanding, the following examples of the specification still take the model training scenario as an example for illustration. Those skilled in the art should understand that when the data processing strategy generation scheme provided by the specification is applied to other use scenarios, the implementation principle is similar, and the specification does not make one-by-one example description for other use scenarios. In addition, in different use scenarios, the data type of the original data set to which the data processing strategy generation scheme is directed can be different, and the data processing type corresponding to the generated target strategy can be different.

[0049] Figure 2 A hardware structure diagram of an electronic device according to an embodiment of the specification is shown. Figure 2 The electronic device 200 in the above description can be used as Figure 1 The electronic device 200 in the above description can be used as Figure 1 The electronic device 200 in the above description can be used as

[0050] As shown in Figure 2As shown, the electronic device 200 can include at least one storage medium 230 and at least one processor 220. In some embodiments, the electronic device 200 can also include a communication port 250 and an internal communication bus 210. The electronic device 200 can also include input / output components (I / O components) 260.

[0051] The internal communication bus 210 can connect different system components. For example, the internal communication bus 210 can connect the storage medium 230, the processor 220, the communication port 250, and the I / O components 260, etc.

[0052] The I / O components 260 support input / output between the electronic device 200 and other components.

[0053] The communication port 250 is used for data communication between the electronic device 200 and the outside world. For example, the communication port 250 can be used for data communication between the electronic device 200 and a network. The communication port 250 can be a wired communication port or a wireless communication port.

[0054] The storage medium 230 can include a data storage device. The data storage device can be a non-transitory storage medium or a transitory storage medium. For example, the data storage device can include one or more of a disk 232, a read-only memory (ROM) 234, or a random access memory (RAM) 236. The storage medium 230 also includes at least one instruction set stored in the data storage device. The instruction set can include computer program code, which can include programs, routines, objects, components, data structures, procedures, modules, etc.

[0055] The at least one processor 220 can be communicatively connected with the at least one storage medium 230. When the electronic device 200 is running, the at least one processor 220 reads the at least one instruction set and executes the generation method of the data processing strategy or the model training method provided in the present specification according to the indication of the at least one instruction set. The processor 220 can execute the steps included in the generation method of the data processing strategy or the steps included in the model training method. The processor 220 can be in the form of one or more processors, and in some embodiments, the processor 220 can include one or more hardware processors, such as a microcontroller, a microprocessor, a reduced instruction set computer (RISC), an application-specific integrated circuit (ASIC), an application-specific instruction set processor (ASIP), a central processing unit (CPU), a graphics processing unit (GPU), a physics processing unit (PPU), a microcontroller unit, a digital signal processor (DSP), a field programmable gate array (FPGA), an advanced RISC machine (ARM), a programmable logic device (PLD), any circuit or processor capable of executing one or more functions, or the like, or any combination thereof.

[0056] For the sake of illustration only, the electronic device 200 in the accompanying drawings shows only one processor 220. However, it should be noted that the electronic device 200 in the present specification can also include multiple processors, and therefore, the operations and / or method steps disclosed in the present specification can be executed by one processor or jointly executed by multiple processors. For example, if the processor 220 of the electronic device 200 in the present specification is described as executing step A and step B, it should be understood that step A and step B can also be executed jointly or separately by two different processors 220 (for example, a first processor executes step A, a second processor executes step B, or the first and second processors jointly execute steps A and B).

[0057] Figure 3 A flowchart of a generation method of a data processing strategy according to an embodiment of the present specification is shown. As before, the electronic device 200 can execute the generation method P300 of the data processing strategy of the present specification as the strategy generation device 120 in the Figure 1 The processor 220 in the electronic device 200 can read the instruction set stored in its local storage medium and then execute the generation method P300 of the data processing strategy of the present specification according to the provisions of the instruction set. For ease of description, the generation method of the data processing strategy will be referred to as the strategy generation method hereinafter. As shown in Figure 3 The method P300 can include the following steps 310 to 330.

[0058] Step 310: Obtain an original data set and obtain a plurality of preset operators.

[0059] The method for generating a strategy provided by the embodiments of the present specification is used to generate a data processing strategy applicable to a data set to be processed. Since the data processing strategies required by different data sets are basically different, the strategy generation device needs to obtain the data set to be processed first to generate the strategy accordingly. In the present specification, the data set to be processed is referred to as an original data set. The original data set can be provided by other devices other than the strategy generation device, such as by the data providing device 110 in Figure 1 In some embodiments, the data in the original data set can also be collected or generated by the strategy generation device itself.

[0060] Data processing needs to use data processing operators, which can also be referred to as data processing operators. The operator is a basic operation unit in the data processing process, which is an executable logic or algorithm unit. Executing a data processing operator on data is to perform a corresponding data processing process on the data. There are many existing available data processing operators, such as data deduplication operators, data optimization operators, data generation operators, data selection operators, and data denoising operators. Each operator is used to implement a corresponding data processing function. An operator of a function can also include multiple operators that can implement the function but are not exactly the same.

[0061] Different data has different data processing needs. Some data can be processed using only one operator to achieve its data processing needs; while some data needs to be processed using multiple operators in sequence to achieve its data processing needs. In the embodiments of the present specification, the strategy generation device generates a data processing strategy for the original data set, and the data processing strategy needs to be combined with the operator. The data processing strategy can include one operator or multiple operators.

[0062] The strategy generation device needs to obtain a plurality of preset operators to select and combine them in sequence to generate a data processing strategy. The plurality of preset operators are the operators available to the strategy generation device, which can be a part of the operators selected from all existing available operators, or all existing available operators. For example, the data processing strategy that the strategy generation device needs to generate is used for data cleaning of the original data set, and the plurality of preset operators can be operators that can help achieve data cleaning, and do not include operators of other functions. In this way, it is easier to determine the appropriate data processing strategy, reduce resource waste in the strategy search process, and improve the strategy search efficiency.

[0063] The strategy generation device obtains the operator, which is equivalent to configuring the corresponding data processing capability or data transformation rule for the strategy generation device. Then, the strategy generation device can execute the operator to implement the corresponding processing of the data, without considering the implementation details of the operator.

[0064] Step 320: generating, by the large language model, an initial strategy set based on the plurality of operators, each strategy in the initial strategy set including a number of operators in the plurality of operators and an execution order of the number of operators.

[0065] In the embodiments of the present specification, the policy generation device utilizes a large language model to generate a policy. The large language model can be deployed in the policy generation device, and the policy generation device directly uses the large language model. Alternatively, the large language model can also be deployed in other devices, and the policy generation device communicates with the other devices to call the large language model and work with the help of the capabilities of the large language model. The large language model has a high degree of automation, good processing capability for complex data, strong generalization ability, and high data processing efficiency. In the embodiments of the present specification, the policy generation device can utilize these capabilities of the large language model to quickly and accurately generate a data processing policy.

[0066] The policy generation device can generate an initial strategy set based on the obtained plurality of operators through the large language model, and then adjust until the final used strategy is obtained. The initial strategy set can include one strategy or multiple strategies. Each strategy can include at least one operator in the plurality of operators. For a strategy including multiple operators, the strategy further includes an execution order of the plurality of operators. Since the execution order of the operators is different, the processing effect on the data is also different, so different strategies can be generated by adjusting the execution order for the same operator.

[0067] The large language model is a prompt-based generative model, and its output is highly dependent on the guidance and control of the prompt text. In the present specification, a specific prompt text needs to be constructed to instruct the large language model to generate a policy. The prompt text is essentially a structured instruction or query information, which can convert the task intention into context information that the model can understand and execute, and is used to guide, constrain and customize the output behavior of the large language model.

[0068] For example, the prompt text can include one or more components of role definition, task instruction, context information, input / output format specification, and examples. The role definition part is used to assign a specific role to the model (e.g., "you are a data analysis expert"), so that it answers questions from the perspective of the corresponding role and professional knowledge. The task instruction part is used to describe the specific task that needs to be performed by the model (e.g., "please summarize the main points of the following text"). The context information part is used to provide background knowledge, reference information or specific data related to the task. The input / output format specification part is used to specify the format requirements of the model output (e.g., "output in JSON format"), to ensure the structured and parsable of the output results. The example part is used to provide one or more example pairs of input-output to demonstrate the expected task execution manner, so that the model can output the required results faster and more standardized.

[0069] In an implementation, the strategy generation method provided by the present specification further includes step 340: generating a prompt text based on the plurality of operators, and inputting the prompt text to the large language model. Wherein, the prompt text at least includes operator description text and first description text of the strategy generation task. The operator description text is used to describe the identification and capability information of the plurality of operators, and the first description text of the strategy generation task is used to guide the large language model to generate an initial strategy set by combining the plurality of operators in different ways.

[0070] The strategy generation device can generate a corresponding prompt text for the strategy generation task to be executed. The strategy generation device uses the prompt text to guide the large language model to execute the strategy generation task, and obtains the target strategy suitable for the original data set output by the large language model. The strategy generation device can generate the prompt text based on the filling mode of the static template, such as filling the information related to the strategy generation task into the preset prompt template. The strategy generation device can also rely on the system state to dynamically construct the prompt text, and use the corresponding prompt text structure to generate the prompt text under different conditions, and splice the historical record as context information into the new prompt text.

[0071] The operator description text can describe the information of each operator respectively. The information of each operator includes the identification and capability information of the operator. The identification of the operator can be the name of the operator, which can be represented by a summary word of its function. The capability information of the operator is an introduction to the function of the operator, which can list one or more functions that the operator can implement and the effects that can be achieved. In an implementation, the operator description text can first list the identification of each operator available to the large language model, and the capability information of the operator can be introduced in a separate introduction paragraph.

[0072] For example, an operator description text includes the following content: "You can use the following operators: 1. Data cleaning operator: responsible for basic data purification; 2. Data generation operator: responsible for supplementing new data; 3. Data optimization operator: responsible for improving data quality; 4. Data selection operator: responsible for filtering high-quality data." Among them, the data cleaning operator, data generation operator, data optimization operator and data selection operator are the identifiers of the four operators.

[0073] On the basis of the foregoing example of the operator description text, the operator description text can also include the following introduction of the capabilities of the operators:

[0074] "The data cleaning operator has the ability to remove duplicate content, special characters, and data that does not meet the length requirements. The data generation operator has the ability to generate missing questions or answers. The data optimization operator has the ability to optimize existing data to improve question and answer accuracy and improve the fluency and conciseness of the text. The data selection operator has the ability to select high-quality data."

[0075] In addition to the above introduction of the operators, the prompt text mainly includes content that guides the large language model to perform a strategy generation task. For example, through the first description text, the large language model is guided to first generate an initial strategy set. Then, through other description texts, the large language model is guided to gradually adjust on the basis of the initial strategy set to determine the final target strategy.

[0076] The first description text of the strategy generation task can describe the methods required to generate the initial strategy set, and can also describe the constraints required to be met by the initial strategy set. For example, the first description text guides the large language model to select operators in the plurality of operators for different combinations and ordering to generate the initial strategy set. For example, the number of strategies included in the initial strategy set is limited to no more than N. For example, the number of operators included in each strategy in the initial strategy set can be one or more. For example, the execution order between the operators in each strategy needs to be explicitly defined to ensure the generation of a diversified initial strategy set. For example, the execution order between the same operators in different strategies is different.

[0077] The following gives an example of the content of the first description text of a strategy generation task:

[0078] "According to your experience, initialize a plurality of different operator combinations as initial strategies. When initializing, each strategy can consist of a single operator or multiple operators, and the execution order of the operators needs to be explicitly defined. The number of initialized strategies is not more than four."

[0079] The above operator description text can belong to the content of the context information part of the prompt text, and the first description text of the strategy generation task can belong to the content of the task instruction part of the prompt text. The prompt text can further include a role definition part, an output format limitation part, and an example part for the large language model. The task instruction part can further include other content, such as a more general instruction described before the first description text.

[0080] The target to be achieved needs to be clearly defined in the prompt text. For example, design an optimized data processing strategy to improve data quality and achieve better model performance. Based on the above example operator description text and the first description text of the strategy generation task, the prompt text generated by the strategy generation device can further include the following content:

[0081] "You are an intelligent assistant specializing in processing model training data. Currently, you need to generate a data processing strategy to improve the data quality of the original data set used to train the model, thereby improving the training accuracy of the model, and you cannot view the specific content of the data. You can generate the strategy by selecting operators and planning their execution order, and find the appropriate strategy to process the original data set through continuous iteration.

[0082] The following are two strategy examples:

[0083] Strategy 1: Data cleaning operator.

[0084] Strategy 2: Data cleaning operator, data generation operator.

[0085] Please strictly follow the following format to output the results:

[0086] Strategy 1: Output operator name, separated by commas in execution order. ...

[0088] Strategy N: Output operator name, separated by commas in execution order."

[0089] The first two paragraphs of the above example prompt text content belong to the role definition of the large language model, and the remaining content belongs to the general instruction given to the large language model. This content also includes an output format limitation part and an example part for the large language model.

[0090] In an implementation, the first description text of the strategy generation task further serves to guide the large language model to combine no more than a preset number of operators when performing the operator combination. The preset number can be a specified constant. Alternatively, the preset number can not be a specific value, but a relatively vague description indicating that the large language model combines with fewer operators, to represent the meaning of "combining no more than a preset number of operators".

[0091] For example, the preset number is three. The first description text of the strategy generation task can further include the following content: "In the first round of initializing strategies, combine no more than three operators." Through this content, the large language model can be guided to combine no more than a preset number of operators.

[0092] In the embodiments of the present specification, the large language model is guided to combine with fewer operators in the early stage of strategy generation to generate strategies. The strategies generated in this way include various small operator combination units, and can obtain strategies containing the same operators but different execution orders of the operators. Based on these strategies, the strategy generation device can analyze and understand the influence of different combination modes and execution orders of the operators on the data processing effect. Based on these strategies, the strategy generation device can also collect preliminary insights about the effectiveness of each operator and its interaction, laying the foundation for further optimizing the strategy. If more operators are used to generate strategies in the future, the differences in the influence of the data processing effect of these small operator combination units can be referred to, to more conveniently and efficiently determine strategies with better data processing effect.

[0093] For example, the multiple operators obtained by the strategy generation device include A, B, C, and D. In the early stage, the strategy 1 generated by the large language model is "A→B", the strategy 2 is "B→A", and the strategy 3 is "A→C". The "→" in these strategies represents the execution order between the operators. Through strategy 1 and strategy 2, the influence of the execution order between A operator and B operator on the data processing effect can be determined. Through strategy 1 and strategy 3, the influence of the combination of A operator with different operators on the data processing effect can be determined. If it is determined that the combination of A operator and C operator achieves better data processing effect, then in the subsequent generation of strategies, compared with the combination of A operator and B operator, the combination of A operator and C operator can be preferentially selected.

[0094] In the embodiments of the present specification, when guiding the large language model to perform step 320, the prompt text input to the large language model can further include a start execution instruction on the basis of the above example content. For example, it can include "Now please execute the initialization of multiple strategies (no more than four)". Through this instruction, the large language model can be guided to generate the initial strategy set as required.

[0095] Step 330: iteratively adjust the initial strategy set based on the original data set until a target strategy suitable for the original data set is obtained.

[0096] After the strategy generation device generates the initial strategy set through the large language model, the strategy generation device needs to adjust the strategy based on the initial strategy set to search for a strategy more suitable for the original data set, that is, to determine a strategy that can achieve better processing effect on the original data set. In the embodiments of the present specification, the strategy search is performed in an iterative adjustment manner, and each adjustment continues to adjust on the basis of the previous adjustment result.

[0097] In the iterative adjustment process of step 330, each iteration includes the following steps 332 and step 334.

[0098] Step 332: use each strategy in the current strategy set to process the original data set respectively, and obtain the evaluation index of each strategy. The evaluation index represents the processing effect on the original data set. The current strategy set is the initial strategy set or the strategy set after the last adjustment.

[0099] Step 334: feed back each strategy in the current strategy set and its evaluation index to the large language model, and adjust each strategy in the current strategy set through the large language model.

[0100] The process of generating the initial strategy set by the large language model is the first round of work in the iterative adjustment of the strategy, that is, the initialization process. The large language model obtains a strategy set in each round of work, and the strategy generation device processes the strategy set obtained by the large language model in each round in the same way. In the present specification, the strategy set obtained by the large language model in the current round is referred to as the current strategy set. In the first round of work, the current strategy set is the initial strategy set; in any subsequent round of work, the current strategy set is the strategy set after the last adjustment of the large language model. The processing method of the current strategy set is introduced below, and the strategy set in each round can be referred to.

[0101] In the embodiments of the present specification, in each round of iterative adjustment, each strategy in the current strategy set needs to be evaluated to adjust the strategy based on the evaluation result and update the strategy set. The strategy generation device can apply each strategy generated by the large language model in the last round to the original data set to verify the processing effect of each strategy on the original data set, and then make the large language model adjust the strategy based on the processing effect of each strategy.

[0102] In the embodiments of the present specification, the evaluation index of a strategy is used to represent the processing effect of the strategy on the original data set. The evaluation index can be presented in the form of a numerical value, or in the form of a percentage, or in the form of a word. For example, the higher the evaluation index, the better the processing effect on the original data set. In some ways, a lower evaluation index can also represent a better processing effect. The present specification does not limit the presentation form of the evaluation index and the corresponding relationship with the processing effect.

[0103] The strategy generation device utilizes each strategy in the current strategy set to perform data processing on the original data set to obtain a data set processed by each strategy, and determines the evaluation index of the strategy by using the data set. For example, the strategy generation device can directly execute the strategy on the original data set by using a large language model to obtain a processed data set. The strategy generation device can also use other processing modules possessed by the strategy generation device to execute the strategy on the original data set to obtain a processed data set. The strategy generation device can also send the original data set and the strategies in the current strategy set to other processing devices to instruct the other processing devices to execute the strategy on the original data set, and obtain the processed data set sent by the other processing devices.

[0104] In some embodiments, different operators in one strategy can be executed on different devices or modules. The strategy generation device can execute the corresponding operator by using a device or module having an environment required for execution of each operator in the strategy. If two adjacent operators need to be executed in different environments, the result obtained after execution of the previous operator can be input into the execution environment of the next operator for processing.

[0105] After obtaining the data set processed by each strategy, the strategy generation device can determine the evaluation index corresponding to the strategy by using the determination method of the evaluation index. The determination process of the evaluation index can be performed by the strategy generation device, or the strategy generation device can call other modules or devices to perform the determination process. In an optional way, the evaluation index of the strategy can be directly obtained after the original data set is processed by the strategy by designing a flow of data processing by the strategy.

[0106] Some implementation ways of determining the evaluation index of each strategy (i.e., performing strategy evaluation) in step 332 will be introduced below.

[0107] In an implementation of the strategy evaluation, the strategy generation device evaluates the strategies based on the use and use scenario of the original data set itself. The strategy generation device tests the use of the data set obtained after processing the strategies in the use. The use effect of the processed data set in the use is used as an evaluation index of the strategies. In this way, the strategies are evaluated in the most direct way that fits the use scenario, which can ensure that the evaluation index obtained is highly accurate and can accurately reflect the processing effect on the original data set.

[0108] In the embodiments of the present specification, the original data set is a data set used for model training, and the original data set can be composed of multiple data samples. Accordingly, the step 332 of processing the original data set using each strategy in the current strategy set and obtaining the evaluation index of each strategy can include: for each strategy in the current strategy set, the following steps are performed to obtain the evaluation index of the strategy: processing the original data set using the strategy to obtain a target data set, obtaining first model accuracy information obtained by training a model using the target data set, and determining the evaluation index of the strategy based on the first model accuracy information.

[0109] For each strategy in the current strategy set, the strategy generation device can process the original data set using the strategy. In the present specification, the data set obtained after processing the original data set is referred to as a target data set. For example, a strategy includes A operator and B operator executed in sequence, the strategy generation device can execute A operator on the original data set to obtain an intermediate data set, and then execute B operator on the obtained intermediate data set to obtain a target data set.

[0110] The original data set is used for model training (such as for training a target model), and the better the model training effect, the better the quality of the original data set. Accordingly, in this strategy evaluation method, the strategy generation device can use the target data set obtained by using each strategy for model training, and determine the evaluation index of the strategy based on the obtained model training effect. The model training effect can be represented by first model accuracy information, therefore, the strategy generation device can determine the evaluation index of each strategy based on the first model accuracy information corresponding to each strategy. The first model accuracy information corresponding to the strategy represents the model training effect obtained by training a model using the target data set obtained by the strategy.

[0111] When the target data set obtained by using any strategy is used for model training, the target data set can be divided into a training set and a test set. The training set is used to drive the learning algorithm and adjust the model weight parameters in the specific training process of the model. The test set is used to measure the generalization ability of the model when facing new data as an evaluation benchmark after the end of the entire training process. The result obtained by evaluating the model through the test set after the model training can be used as the first model precision information to represent the model training effect. The above process of obtaining the first model precision information through model training can be performed by the strategy generation device itself or can be instructed by the strategy generation device to other devices to perform.

[0112] The original data set is used to train the target model. The target model can be any model, such as a large language model or a common model for implementing certain single functions. In an embodiment of the present specification, the target model can be directly trained using the target data set corresponding to each strategy to obtain the corresponding first model precision information. The parameter size of the model trained by the original data set itself can be large. For example, the parameter size of the target model is greater than a preset size. The preset size can be a size representation value specified by a staff or a strategy generation device, or can not be a fixed size representation value. If the target data set corresponding to each strategy is used to directly train the target model, more processing resources and longer training time can be consumed.

[0113] In order to avoid the above problems caused by directly training the target model, in an embodiment of the present specification, a smaller model can be trained using the target data set to implement strategy evaluation. The model can be a model with the same or similar functional effect as the target model, but with a smaller parameter size, which is equivalent to a simplified version of the target model. In this way, the first model precision information obtained based on model training can ensure that accurate evaluation indicators can be obtained, while reducing resource consumption, improving the determination efficiency of evaluation indicators, and accordingly improving the strategy search efficiency.

[0114] In the strategy evaluation manner, the strategy generation method provided in the embodiments of the present specification can further include: performing compression processing on the target model to obtain a test model, and a parameter size of the test model being smaller than a preset size. The step of obtaining the first model precision information trained by the target data set can include: training the test model by using the target data set, and taking the precision information of the trained test model as the first model precision information. For example, the compression processing performed on the target model can include operations such as knowledge distillation, model pruning, model quantization, or low-rank decomposition. The compression processing performed on the target model can be performed by the strategy generation device or by another device, and the strategy generation device can directly obtain or call the test model. The training of the test model and the determination of the precision information can refer to the above description of the model training and the first model precision information, and will not be repeated here.

[0115] In the embodiments of the present specification, the strategy generation device can determine the evaluation index of the strategy based on only the first model precision information. For example, the first model precision information corresponding to the strategy is directly taken as the evaluation index of the strategy, or only a certain format conversion is performed on the first precision information to obtain the evaluation index of the strategy.

[0116] In another embodiment, the strategy generation device determines the evaluation index of the strategy based on the first model precision and other information. The strategy generation device can also determine the evaluation index of the strategy based on the improvement of the use effect after the original data set is processed by using the strategy. For example, the more the use effect is improved, the better the data processing effect of the corresponding strategy can be represented by the evaluation index of the strategy. In this way, the evaluation index of the strategy can reflect the change after data processing, and better reflect the effect of the strategy.

[0117] In this embodiment, the strategy generation method provided in the embodiments of the present specification can further include: obtaining second model precision information trained by using the original data set. The step of determining the evaluation index of the strategy based on the first model precision information can include: comparing the first model precision information with the second model precision information, and determining the evaluation index of the strategy based on the comparison result. The comparison result can represent the improvement of the model training effect after the original data set is processed by using the strategy. The evaluation index determined based on the comparison result can better reflect the effect of the strategy, and the accuracy of the evaluation index is higher.

[0118] The process of obtaining the second model precision information by training the model with the original data set can refer to the foregoing description of obtaining the first model precision information by training the model with the target data set. The model to be trained with the original data set is consistent with the model to be trained with the target data set. For example, the test model is trained with the target data set and the original data set. The second model precision information is obtained in the same way as the first model precision information.

[0119] After obtaining the first model precision information corresponding to each strategy and the second model precision information corresponding to the original data set without processing, the strategy generation device can compare the first model precision information corresponding to each strategy with the second model precision information. For example, the model precision information is represented by a numerical value, and the larger the numerical value, the higher the model precision and the better the model training effect. For each strategy, the strategy generation device can calculate the difference between the second model precision information and the first model precision information corresponding to the strategy, and determine the evaluation index of the strategy based on the difference. For example, the difference is directly determined as the evaluation index of the strategy, or the difference is transformed to obtain a result as the evaluation index.

[0120] In another implementation manner of strategy evaluation, a data quality scoring model can be trained, or a trained data quality scoring model can be obtained. The strategy generation device can input the target data set processed by each strategy in the current strategy set into the data quality scoring model to obtain the feedback score output by the model for each target data set corresponding to each strategy. The strategy generation device can take the feedback score corresponding to each target data set as the evaluation index of the strategy used to obtain the target data set.

[0121] In each iteration, after obtaining the evaluation index of each strategy in the current strategy set by step 332, the strategy generation device performs step 334 to feed back each strategy in the current strategy set and its evaluation index to the large language model, and adjusts the strategy based on the evaluation index through the large language model. During the strategy adjustment process, the large language model still needs to work based on the guidance of the prompt text. In each iteration, the strategies and their evaluation indexes obtained in the last iteration can be added to the prompt text and input into the large language model. For example, the strategy and the evaluation index are added to the prompt text as context information. The large language model can understand the advantages of each strategy based on the evaluation index of each strategy to adjust the strategy and gradually adjust to a better strategy that can achieve better data processing effect.

[0122] The prompt text generated by the policy generation device in the above introduction can also include a second description text of a policy adjustment task, which guides the large language model to adjust the policy in each iteration. In an embodiment, the second description text of the policy adjustment task is used to guide the large language model to adjust each policy in the current policy set with the goal of improving the processing effect on the original data set. The adjustment manner includes at least one of the following manners: adjusting the operator combination contained in each policy in the current policy set, adjusting the operator execution order of each policy in the current policy set, and adjusting the number of policies in the current policy set.

[0123] For example, during the policy adjustment process, the large language model can replace, add, or delete some operators contained in the policies in the current policy set to adjust the operator combination contained in the policies. The large language model can also change the order of the operators in some policies included in the current policy set. The large language model can also reselect and sort the operators from the preset plurality of operators to generate new policies and add them to the policy set. The large language model can also delete all or part of the policies in the current policy set. The number of policies in the policy sets generated in different iterations can be the same or different, which can be determined by the large language model according to the understanding.

[0124] In an embodiment, the second description text of the policy adjustment task is also used to guide the large language model to infer and determine the adjustment manner based on the comparison between the evaluation indicators of different policies during the iterative adjustment. For example, the second description text can guide the large language model to compare the evaluation indicators of the policies in the current policy set, determine the top n policies with the highest evaluation indicators, and n can be an integer greater than or equal to 1. For example, the large language model determines a policy with the highest evaluation indicator. The second description text can also guide the large language model to adjust the policy in the direction close to the top n policies, and the adjusted policy has not been generated before. For another example, the second description text can also guide the large language model to recombine the operators in the top n policies to adjust the policy. The specific policy adjustment manner obtained according to the comparison between the evaluation indicators is determined by the large language model itself, and here are only two examples, which do not constitute a limitation.

[0125] In the embodiments of the present specification, the large language model is guided to determine the adjustment manner based on the comparison between the evaluation indicators of the policies. In this way, the large language model can adjust the policy based on the previous better policy as a reference, which can speed up the process of obtaining the target policy and avoid searching in inappropriate policies without direction, which affects the search efficiency.

[0126] In the embodiments of the present specification, the policy generation device performs a plurality of iteration processes using the large language model until the large language model outputs a target policy meeting the condition. The target policy can be a data processing policy suitable for the target original data according to the large language model. The policy generation device can use the prompt text to guide the large language model to finally output only one target policy that is truly suitable for the target data set, or can also guide the large language model to output a plurality of target policies that are better.

[0127] An exemplary content of the second description text of the policy adjustment task is given below:

[0128] "You need to iteratively adjust the policy to improve the processing effect of the policy on the original data set. At each adjustment, you can compare the evaluation indicators of the current policies, and determine how to adjust the policy according to the comparison. The higher the evaluation indicator of a policy, the better the data processing effect of the policy. When adjusting, you can change the policy combination and working order in the policy, and also can change the number of generated policies. Repeat the adjustment until you think you have searched for a policy that achieves a good processing effect on the original data set, without adjusting any operator combination and its working order."

[0129] In some embodiments, the prompt text input to the large language model can also include some special content (such as notes) that need to be specified to guide the large language model to comply with during work. For example, the notes can supplement the description of the functions of the operators operable by the large language model, specify the task execution mode of the large language model, and also describe some possible situations during the execution process and the execution mode under the situation.

[0130] Figure 4 A simple block diagram of a policy generation method according to an embodiment of the present specification is shown. The policy generation method provided by the present specification will be introduced in detail below Figure 4 and the foregoing introduction to Figure 3 . As shown in Figure 4 , the policy generation device generates an initial prompt text. The content of the initial prompt text can refer to the content included in the foregoing example of the prompt text. The policy generation device performs step 410 to input the initial prompt text to the large model agent, i.e., the large language model described above. The large model agent uses the large language model as a brain to solve the user's problem. The large model agent can also obtain a plurality of available data processing operators, Figure 4 Only four types of operators, data cleaning operators, data generation operators, data optimization operators, and data selection operators, are taken as examples for introduction.

[0131] The large model agent can execute step 420, generating N different data processing (DP) strategies based on the obtained multiple operators. For example, strategy 1 includes two operators executed sequentially, while strategy N includes only one operator. Then, the strategy generation device executes step 430, using the different DP strategies generated by the large model agent to process the original datasets respectively, obtaining the corresponding target datasets. Based on each target dataset, the strategy generation device executes step 440 to evaluate the strategies and obtain evaluation metrics for each DP strategy.

[0132] Figure 5 A simplified flowchart of a strategy evaluation method provided according to an embodiment of this specification is shown. Figure 5 As shown, the policy generation device can train a test model using the target dataset corresponding to each policy, train a test model using the unprocessed raw dataset, and perform inference based on the training results to calculate the policy's evaluation metric. For example, the policy generation device obtains the corresponding first model accuracy information by training the test model using the target dataset, and obtains the corresponding second model accuracy information by training the test model using the unprocessed raw dataset. Then, the policy generation device compares the first model accuracy information with the second model accuracy information corresponding to each policy to determine the policy's evaluation metric based on the comparison results.

[0133] After step 440, the policy generation device executes step 450, grouping each DP policy and its evaluation metric into a set, filling the prompt text with the contents of this set, and then inputting the optimized prompt text back into the large model agent. Steps 420 to 450 constitute one iteration in the policy generation process. The optimized prompt text can guide the large model agent to execute this iteration process again, allowing it to adjust its policies based on the evaluation metrics of each policy in the previous iteration. This policy adjustment process is the same as the policy regeneration process in step 420. The policy generation device can use the large model agent to iterate and adjust the policies multiple times until a policy truly suitable for the original dataset is output. For example, this policy can be a policy whose evaluation metric is greater than a preset threshold, or a policy that the large model agent cannot increase further after multiple iterations.

[0134] In the embodiments of the present specification, the policy generation device iteratively adjusts the initial policy set based on the original data set, which can be based on all data in the original data set or based on part of the data in the original data set. In an embodiment, the step 330 of iteratively adjusting the initial policy set based on the original data set until the target policy suitable for the original data set is obtained comprises: sampling the original data set to obtain a sampled data set while maintaining the data distribution characteristics; and iteratively adjusting the initial policy set based on the sampled data set until the target policy suitable for the original data set is obtained.

[0135] In the process of iteratively adjusting the policy based on the original data set, the original data set is used to evaluate the data processing performance of each generated policy to obtain the most effective policy for the original data set. The original data set can include many types of data, and a data processing policy has similar processing effects on each data in the same type of data. For a policy, processing various types of data in the original data set with the policy can obtain the processing effect on the entire data set without repeated processing of similar data. Therefore, only a data subset that can sufficiently approximate the basic data distribution of the original data set can achieve performance evaluation of the policy.

[0136] In the embodiments of the present specification, the original data set is sampled while maintaining the data distribution characteristics, which can ensure that the sampled data set well represents the original data set. Based on the sampled data set, the policy can be accurately evaluated. Based on the sampled data set, the initial policy set is iteratively adjusted, which only needs to process part of the data in the original data set, thereby reducing the amount of data to be processed and improving the data processing efficiency, and further improving the efficiency of obtaining the target policy.

[0137] In an example, maintaining the data distribution characteristics needs to ensure that the proportion of each type of data in the sampled data set is substantially the same as the proportion of each type of data in the original data set. For example, the data in the original data set can be classified by type to determine the proportion of each type of data. When sampling, each type of data is sampled at the same proportion. For example, the original data set includes 100 data, of which 40 are L1 type data and 60 are L2 type data. If a total of 10 data is to be sampled, the sampling ratio of each type of data can be 10:1. Four data are sampled from the L1 type data and six data are sampled from the L2 type data. The type classification process can be performed by the model to ensure the data privacy of the original data set.

[0138] The original dataset contains data of varying quality, including some low-quality data with significant noise. Therefore, processing of the original dataset is necessary. Different strategies exhibit different processing effects depending on the level of noise. In the embodiments described in this specification, data types can be divided based on the level of noise, dividing the original dataset into multiple subsets. Sampling is then performed on each of these subsets to obtain a sampled dataset, ensuring that the proportion of different data types remains unchanged after sampling.

[0139] In this approach, the steps of sampling the original dataset to obtain a sampled dataset while maintaining the data distribution characteristics include: dividing the original dataset into multiple sub-datasets based on the noise level of the data; sampling each sub-dataset separately while maintaining the data distribution characteristics of each sub-dataset to obtain the sampling results corresponding to each of the multiple sub-datasets; and merging the sampling results corresponding to each of the multiple sub-datasets to obtain the sampled dataset.

[0140] Figure 6 A simplified flowchart illustrating a sampling of an original dataset according to an embodiment of this specification is shown below, in conjunction with... Figure 6 The process of obtaining the sampled dataset is described. The strategy generation device can classify the data in the original dataset using a classification model F. This data classification model F can classify the data in the original dataset based on the noise level, grouping data with the same level of noise into a subset.

[0141] For example, the classification model F can be a binary classification model (also known as a binary selection model). For instance... Figure 6 As shown, classification model F can divide the original dataset into two subsets: one subset includes data with noise levels no higher than a preset level, and the other subset includes data with noise levels higher than the preset level. For example, using D... clean and D noise This represents the two subsets of data. In some implementations, the classification model F can also be a multi-class model. For example, the noise level can be divided into three different levels, each corresponding to one of the three sampling subsets. The policy generation device uses the three-class model to determine the noise level of each data point in the original dataset and assigns the data to the corresponding sampling subset based on its noise level.

[0142] After dividing the original dataset into multiple subsets, the policy generation device can sample data from each subset while maintaining its data distribution characteristics. The collection of data obtained by sampling from each subset is called the sampled subset. Figure 6 As shown, the policy generation device targets D clean and Dnoise Respectively performing the sampling operation to obtain a corresponding sampling subset D s,clean and D s,noise Further, the policy generation device can combine the two sampling subsets to obtain the final sampling dataset.

[0143] In one sampling manner, the total sampling number M can be set for the original dataset. After the original dataset is divided into multiple sub-datasets, the data proportion of the multiple sub-datasets can be determined. Based on the data proportion, the number of data to be sampled in each sub-dataset is determined, and then the number of data in the corresponding sub-dataset is sampled. It is assumed that m data need to be sampled for each sub-dataset.

[0144] Next, a sampling manner for maintaining data distribution characteristics is introduced. For any one of the multiple sub-datasets divided from the original dataset, the sub-dataset includes multiple data samples. The sampling process for the sub-dataset includes repeatedly performing the following steps 610 to 630 until the preset sampling scale is reached. The preset sampling scale can be represented by the set total sampling number.

[0145] Step 610: Each data sample in the sub-dataset is traversed, and for the current data sample being traversed, the similarity between the embedding vector of the current data sample and the embedding vectors of other data samples is determined, and the similarity is taken as the representative degree of the current data sample to the sub-dataset.

[0146] The policy generation device can use the embedding model to map each data sample in the sub-dataset to a high-dimensional embedding vector to obtain a corresponding vector subset. Each embedding vector in the vector subset is traversed to determine the similarity between the currently traversed embedding vector (referred to as the current embedding vector, that is, the embedding vector of the current data sample) and other embedding vectors. The similarity is the overall similarity between the embedding vector and all remaining embedding vectors. The similarity is used to represent the representative degree of the current data sample to the sub-dataset.

[0147] For example, there are 100 data samples in a sub-dataset, each of which can be mapped to an embedding vector to obtain a vector subset composed of 100 embedding vectors. The policy generation device calculates the similarity between each embedding vector in the 100 embedding vectors and the remaining 99 embedding vectors.

[0148] Step 620: Determine the target data sample with the maximum representative degree from the sub-dataset.

[0149] The policy generation device can select the target data sample with the maximum representative degree in the sub-dataset based on the representative degrees of the data samples, as the data sample sampled in this round.

[0150] Step 630: add the target data sample into the sampling result corresponding to the sub-data set, and delete the target data sample from the sub-data set.

[0151] For example, the policy generation device selects a target data sample after one round of sampling, and then updates the sub-data set for sampling, so that the sub-data set is adjusted from including 100 data to including the remaining 99 data that are not selected. Then the above steps 610 to 630 can be repeatedly executed to determine the similarity of the embedding vectors of each data and the remaining 98 data, and then determine another target data sample based on the similarity.

[0152] In the embodiments of the present specification, the similarity is determined by converting the data samples into embedding vectors. The embedding vectors can deeply mine and quantify the semantic association between the data, can better identify the data samples, and ensure the accuracy of the similarity determination; and the efficiency of vector operation is relatively high, which can improve the efficiency of similarity determination. By calculating the similarity between each data sample and other data samples in the data set, data sampling is performed. Each target data sample sampled in this way can well represent the overall data distribution of the data set, and ensure that each data sample sampled is most similar to the original data distribution characteristics of the data set.

[0153] The above is an example of using the iterative sampling method of the policy generation device to sample each sub-data set obtained by dividing the original data set. In some embodiments, the policy generation device can directly perform the above iterative sampling method on the original data set to obtain a sampled data set. For example, the policy generation device can first establish an empty sampling data set D s , and fill the target data sample sampled in each iteration into the sampling data set until the size of the sampling data set D s reaches the preset sampling size.

[0154] In an optional implementation, after the policy generation device obtains the similarity of each sample data and other sample data in the data set in step 610, it can directly select the top m data with the highest similarity as the target data sample, and then directly obtain the sampling result for the data set.

[0155] In the embodiments of the present specification, the policy generation device performs step 332 in the process of each iteration, and processes the original data set by using each policy in the current policy set respectively to perform policy evaluation based on the obtained target data set. The policy generated by the policy generation device is used to reduce the noise contained in the data, and to process the noise data into clean data. Each data in the original data set contains different degrees of noise. For clean data with relatively high quality and low noise degree in the original data set, the difference before and after processing by using the policy is small, and the influence on policy evaluation is small, so the data can be processed without using the policy. The policy generation device can skip the processing process of the policy on these data, and select part of the data in the original data set for processing. In this way, the processing time of these data can be saved, and the policy search efficiency can be improved.

[0156] In step 332, the policy generation device needs to process the original data set by using the first policy. It is assumed that the first policy is any one of the policies in the current policy set. The following describes an implementation of step 332 by taking the first policy as an example. The same processing process as the first policy can be performed for each policy in the current policy set. The first policy includes a first operator and a second operator executed in sequence. The first operator is the first operator executed in the first policy, and the second operator is the last operator executed in the first policy. For other policies including more operators, the process performed before the execution of the operator can refer to the process before the first operator, the process performed after the execution of all operators can refer to the process after the second operator, and the process of executing each operator can refer to the execution process of the first operator and the second operator.

[0157] The process of processing the original data set by using the first policy can include the following steps 710 to 740. Figure 7 An easy flowchart of processing the original data set by using the first policy according to an embodiment of the present specification is shown. As shown in Figure 7 The first policy includes a first operator S1 and a second operator S2. The process of processing data by using the first policy is introduced below in combination with Figure 7 the example.

[0158] Step 710: dividing the original data set into a first clean data subset and a first noise data subset.

[0159] Each policy generated by the policy generation device is used to process the data into clean data. The policy generation device can screen the clean data and noise data in the original data set, and divide the original data set into two disjoint data subsets, such as a first clean data subset and a first noise data subset. For example, as shown in Figure 7 The policy generation device can divide the original data set by using a classification model F to obtain a first clean data subset Dcur,clean and the first noisy data subset D cur,noisy The method for partitioning the data subset here can refer to the aforementioned methods for partitioning multiple datasets. The classification model F can be found in the previous introduction and will not be repeated here. The policy generation device uses the operators in the policy to process the data in the first noisy data subset; the data in the first clean data subset can be skipped.

[0160] Step 720: Execute the first operator on the first noisy data subset to obtain the first result data subset, and combine the first result data subset with the first clean data subset to obtain the intermediate dataset.

[0161] like Figure 7 As shown, the policy generation device targets the first noisy data subset D. cur,noisy Execute the first operator S1 in the first strategy on the first clean subset D. cur,clean No action taken. Then merge D. cur,noisy The execution result and D cur,clean Obtain the intermediate dataset.

[0162] Step 730: Divide the intermediate dataset into a second clean data subset and a second noisy data subset.

[0163] Step 730 can be referred to the above description of step 710, and will not be repeated here.

[0164] Step 740: Execute the second operator on the second noisy data subset to obtain the second result data subset, and combine the second result data subset with the second clean data subset to obtain the target dataset.

[0165] Step 740 can be referred to the aforementioned description of step 720, and will not be repeated here. The target dataset represents the final processing result obtained after applying a complete policy to the original dataset.

[0166] Each operator in the strategy also optimizes the data to reduce its noise level. Some noisy data may become clean data after the first few operators in the strategy. Therefore, during processing using subsequent operators in the strategy, this clean data can be skipped, further shortening data processing time and improving data processing efficiency. Accordingly, in the embodiments of this specification, the first operator is executed on the first noisy data subset obtained from the first partition of the original dataset. Before executing the next operator, the existing dataset is partitioned again, and the next operator is executed only on the noisy data obtained from the second partition, and so on.

[0167] In an implementation, all operators in the first strategy can be executed based on the first noisy data subset, and the processing result is combined with the first clean data subset to obtain the target data set.

[0168] In the embodiments of the present specification, when the strategy is executed on the original data set, the processing of the clean data in the original data set can be skipped. In this way, the data processing time can be reduced, and the efficiency of executing the strategy on the original data set can be improved without affecting the evaluation effect of the strategy, and the overall strategy search efficiency can be improved.

[0169] The above-mentioned way of skipping the processing of clean data provided by the embodiments of the present specification can improve the strategy search efficiency. In order to facilitate description, the above-mentioned way of sampling will be referred to as the first efficiency improvement way, and the way of skipping the processing of clean data will be referred to as the second efficiency improvement way. The two efficiency improvement ways can be used alone or together. When the two efficiency improvement ways are used together, the original data set in step 710 can be replaced by the sampled data set. Figure 7 For example, the two ways are used together, as shown in the following figure: Figure 7 The original data set D s The classification model F is used.

[0170] The third efficiency improvement way is also provided below, which is to reuse the target data set corresponding to the generated strategy. The strategy generation device generates different strategies by combining and sorting different operators. In the process of iteratively adjusting the strategy, there can be some operator combinations in the strategies generated in the later rounds that coincide with the strategies generated in the previous rounds. When each strategy is evaluated, the same data set is processed. For example, each strategy is used to process the original data set or the sampled data set in the original data set. The same operator combination is used to process the same data set, and the same result is obtained. In this way, the processing result obtained in the previous round can be reused in the subsequent round, so that the subsequent round can continue to process directly based on the processing result, so as to improve the data processing efficiency.

[0171] In the application of the third efficiency improving manner, the strategy generation method provided by the embodiments of the present specification includes, for any first strategy in the current strategy set, the step of processing the original data set by using the first strategy, including: querying whether the historical data set is stored in the preset storage space, wherein the historical data set is obtained by processing the original data set by using the second strategy, and the operator combination contained in the second strategy is a prefix of the operator combination contained in the first strategy; and if yes, obtaining the historical data set from the preset storage space, and executing the operators in the latter part of the first strategy in sequence based on the historical data set, until the processed target data set is obtained. If the historical data set corresponding to the first strategy is not queried in the preset storage space, the original data set is processed by using each operator in the first strategy in sequence.

[0172] The strategy generation device can store each strategy generated in each iteration and the data processing result obtained by executing the strategy on the original data set (i.e., the target data set obtained by processing the original data set by using the strategy) in the preset storage space. In the embodiments of the present specification, the target data set stored in the preset storage space is referred to as a historical data set. Figure 8 A schematic diagram of strategy storage is shown according to an embodiment of the present specification. As shown in Figure 8 DP strategy 1 includes sequentially executed first operator S1 and second operator S2, DP strategy 2 includes sequentially executed third operator S3, fourth operator S4 and first operator S1, and DP strategy N includes first operator S1.

[0173] In the execution of the first strategy, the strategy generation device can first query whether the operator combination contained in a certain strategy is a prefix of the operator combination contained in the first strategy based on the operators contained in the first strategy and the execution order thereof. For example, it is queried that the operator combination contained in the second strategy is a prefix of the operator combination contained in the first strategy. The historical data set corresponding to the second strategy is also stored in the preset storage space. The strategy generation device can obtain the historical data set corresponding to the second strategy from the preset storage space, and execute the operators different from the second strategy in the first strategy in sequence based on the historical data set.

[0174] For example, the first strategy generated in the current round is "S1→S2→S3", and for the first strategy, it is queried in the preset storage space that the second strategy "S1→S2" exists, which coincides with the operator combination "S1→S2" in the prefix of the first strategy. The strategy generation device can obtain the historical data set corresponding to the second strategy, and execute the operator S3 based on the historical data set to obtain the target data set corresponding to the first strategy.

[0175] In some cases, the plurality of strategies in the preset storage space contain operator combinations that are prefixes of the operator combination contained by the first strategy, but are respectively prefixes of different lengths in the first strategy. At this time, the strategy generation device can select the strategy corresponding to the longest prefix as the second strategy, and execute the operators of the subsequent part on the basis of the corresponding historical data set thereof. In this way, the data processing process can be reduced to the greatest extent, and the data processing efficiency can be improved. For example, for the first strategy "S1→S2→S3", it is found in the preset storage space that there are DP strategy 1 ("S1→S2") and DP strategy N ("S1") that conform to the prefix of the first strategy. The strategy generation device can select DP strategy 1 as the second strategy.

[0176] After the target data set is obtained by processing the original data set by using the first strategy, the strategy generation method provided in the embodiments of the present specification further includes storing the first strategy and the target data set to a preset storage space. The strategy generation device can store each strategy generated in each iteration and the corresponding target data set, so as to facilitate subsequent reuse, reduce redundant calculation, and accelerate strategy evaluation.

[0177] In the third efficiency improvement mode and the first and second efficiency improvement modes described above, any one, any two, or all three can be selected and used as needed, which is not limited here.

[0178] In the strategy generation method provided in the embodiments of the present specification, a large language model is used as an intelligent agent, without the need for human intervention, to automatically generate and optimize data processing strategies. Starting from the initial prompt text, the large language model is guided to generate a plurality of candidate strategies, and the large language model is guided to compare and analyze the strategies by using the evaluation indicators of the strategies, to iteratively improve the strategies and obtain a target strategy suitable for the original data set. This iterative context learning mechanism enables the intelligent agent to gradually converge to a high-quality strategy generation process without the need for direct human intervention or access to the original data set, and does not cause human privacy leakage problems. In the embodiments of the present specification, three efficiency improvement modes are used to reduce data processing time, improve data processing efficiency, reduce redundant calculation, and further improve strategy search efficiency.

[0179] In summary, the data processing strategy generation method provided in this specification combines operators to obtain the strategy and optimizes and adjusts the strategy based on the original dataset. This allows for a better determination of the impact of different combinations of operators on the data processing effect, ensuring that the obtained target strategy is truly applicable to the original dataset and has a good data processing effect on it. Utilizing a large language model for strategy generation and adjustment allows for a more accurate and efficient determination of the target strategy suitable for the original dataset. Furthermore, this method eliminates the need for manual review of the original dataset's contents, enabling the determination of a suitable data processing strategy within a black-box environment, thus ensuring the data privacy of the original dataset.

[0180] Figure 9 A flowchart of a model training method provided according to an embodiment of this specification is shown. This method can be applied to... Figure 1 The model training device 130 shown is an example. Figure 9 As shown, the model training method P900 includes steps 910 to 940.

[0181] Step 910: Obtain the original dataset used to train the target model.

[0182] Regarding the target model and the original dataset, please refer to the aforementioned [reference needed]. Figure 3 The relevant content described in the introduction will not be repeated here. For example, the model training device can obtain the raw dataset used to train the target model from the data providing device. The data providing device can also send a model training request to the model training device, which can carry relevant requirements for model training. Accordingly, the model training device performs subsequent processing based on the model training request.

[0183] Step 920: Generate a target policy applicable to the original dataset, wherein the target policy is generated based on the data processing strategy generation method described above.

[0184] Step 920 can refer to the aforementioned steps. Figure 3 The details of the provided policy generation method are omitted here. The aforementioned policy generation device can be integrated into the model training device, which then executes the policy generation method to obtain the target policy. Alternatively, the model training device can call the policy generation device to execute the policy generation method to obtain the target policy.

[0185] Step 930: Process the original dataset using the target strategy to obtain the training sample set.

[0186] Step 920 can refer to the foregoing description of processing the original data set by using the generated strategy (e.g., the first strategy). For example, the model training device sequentially executes each operator in the target strategy for all data in the original data set to obtain a final processing result, which is taken as the training sample set.

[0187] Step 940: training the target model by using the training sample set.

[0188] After obtaining the training sample set, the model training device can perform specific model training operations by using the training sample set. For example, the model training device can input each training sample in the training sample set into an initial target model to obtain the predicted output of the target model for each training sample. The model training device obtains a loss value based on the difference between the predicted output and the true label of the training sample, iteratively adjusts the parameters of the target model to minimize the loss value, and obtains the trained target model.

[0189] After the target model is trained, it can be used to perform a corresponding task. For example, the target model can be called by a terminal to implement a processing function on corresponding data. For example, if the target model is a text recognition model, it can implement text recognition on data input by the terminal. For another example, if the target model is a generative dialogue model, it can implement a function of answering questions input by the terminal.

[0190] In the embodiments of the present specification, the model training device generates a target strategy suitable for the original data set by using the strategy generation method. Thus, the original data set is optimally processed by using a suitable strategy, which can ensure that the quality of the training samples in the obtained training sample set is high and contains less noise. Moreover, the processing of the original data set can be performed without manually checking the specific content of the original data set, which ensures data privacy. Training the target model by using the training sample set can improve the model training effect and efficiency, facilitate faster training of a usable target model, and ensure that the use effect of the target model is good.

[0191] In another aspect of the present specification, a computer-readable non-transitory storage medium storing at least one set of instructions is provided, which when executed by a processor, directs the processor to perform the steps of the method P300 of generating a data processing strategy or the method P900 of training a model described in the present specification. In some possible implementation, the various aspects of the present specification can also be implemented as a program product in the form of a computer readable medium having computer readable program code embodied in the medium. When the program product is executed by the electronic device 200, the program code directs the electronic device 200 to perform the steps of the method P300 of generating a data processing strategy or the method P900 of training a model described in the present specification. The program product for implementing the above method can include the program code in a portable compact disc read-only memory (CD-ROM) and can be run on the electronic device 200. However, the program product of the present specification is not limited to this, and in the present specification, the readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system. The program product can employ any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any suitable combination of the above. More specific examples of the readable storage medium include a portable disc, a hard disc, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. The computer readable storage medium can include a data signal carried in a carrier wave as part of a baseband in which readable program code is borne. Such a propagated data signal can take on many forms, including but not limited to electro-magnetic, optical, or any suitable combination thereof. The readable storage medium can also be any readable medium that is not a storage medium that can send, communicate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The program code contained in the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical, RF, or the like, or any suitable combination of the above. The program code for performing the operations of the present specification can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C++, or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can be executed entirely on the electronic device 200, partially on the electronic device 200, as an independent software package, partially on the electronic device 200 and partially on a remote computing device, or entirely on a remote computing device.

[0192] The foregoing description of specific embodiments of the disclosure has been presented for the purposes of illustration and description. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order and still accomplish the desired results. Additionally, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous.

[0193] In light of the above, it should be appreciated that the foregoing detailed description of the disclosure and the examples and specific embodiments described therein have been presented for purposes of illustration and description only. It is therefore, not intended to be limited, except as is in the spirit and scope of the disclosure. Many modifications and variations will no doubt be apparent to those of ordinary skill in the art once the disclosure above is known, i.e., once the concepts involved are appreciated. It is therefore, contemplated that the disclosure embodied in the present specification and examples is intended to cover all such modifications and variations as fall within the spirit and scope of the disclosure. It is also to be understood that features recited in the claims can be combined in any combination desired.

[0194] Furthermore, some of the terms used in the present specification are used in their broadest form. For example, the terms “one embodiment”, “an embodiment” and / or “some embodiments” means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the disclosure. The appearance of the phrases “in one embodiment” or “in some embodiments” in various places in the specification are not necessarily all referring to the same embodiment, and may, but do not necessarily, refer to different embodiments. Furthermore, the particular features, structures or characteristics can be combined in any suitable manner in one or more embodiments.

[0195] It should be understood that, in the foregoing description of embodiments of the disclosure, various features are sometimes grouped together in a single embodiment, figure, or description of a related concept. However, this is not mandatory, and alternatively, groups of features can be spread over several embodiments, figures, or descriptions of related concepts. Again, it should be emphasized that the concepts of the present disclosure can comprise, among other things, various combinations of the features described in the foregoing disclosure, even if the various combinations are not explicitly described in the foregoing disclosure.

[0196] Each patent, patent application, publication of a patent application, and other material, for example articles, books, specifications, publications, documents, things, and / or the like which can have been cited or referred to in this document, are incorporated by reference into the present document, and are hereby made part of the present document, for all purposes, to the same extent as if each individual publication or item had been individually cited or incorporated by reference. Additionally, where a definition or use of a term in this document is provided to overcome a definition or use of the term in an individually cited document that is incorporated by reference, the definition or use of the term in this document prevails.

[0197] Finally, it should be understood that the embodiments of the application disclosed herein are illustrative of the principles of the present specification. Other modifications that fall within the scope of the present specification can also be made. Accordingly, the present specification discloses embodiments only as examples. Substantially any arrangement, which is neither specifically described nor explicitly illustrated, can be substituted for the specific embodiments disclosed, without departing from the scope of the present specification. Accordingly, the present specification discloses embodiments only as examples.

Claims

1. A method for generating a data processing strategy, comprising: Obtain the original dataset and several pre-defined operators; An initial policy set is generated based on the multiple operators using a large language model. Each policy in the initial policy set includes several operators from the multiple operators and their execution order. as well as The initial policy set is iteratively adjusted based on the original dataset until a target policy suitable for the original dataset is obtained, wherein each iteration includes: The original dataset is processed using each strategy in the current strategy set, and an evaluation metric for each strategy is obtained. This evaluation metric characterizes the processing effect on the original dataset. The current strategy set is either the initial strategy set or the strategy set adjusted in the previous round. Each strategy in the current strategy set and its evaluation index are fed back to the large language model, and the strategies in the current strategy set are adjusted through the large language model.

2. The method according to claim 1, further comprising: Based on the plurality of operators, a prompt text is generated and input into the large language model, wherein the prompt text includes at least: Operator description text, used to describe the identifiers and capabilities of the plurality of operators, and The first descriptive text for the policy generation task is used to guide the large language model to generate the initial policy set by combining the multiple operators in different ways.

3. The method according to claim 2, wherein, The first descriptive text is also used to: guide the large language model to use no more than a preset number of operators for combination when performing operator combination.

4. The method according to claim 2, wherein, The prompt text also includes: The second descriptive text for the strategy adjustment task is used to guide the large language model to adjust each strategy in the current strategy set with the goal of improving the processing performance on the original dataset, wherein the adjustment method includes at least one of the following: Adjust the operator combinations contained in each strategy in the current strategy set. Adjust the execution order of operators for each policy in the current policy set, and Adjust the number of strategies in the current strategy set.

5. The method according to claim 4, wherein the second descriptive text is further used to: guide the large language model to infer the adjustment method based on the comparison between evaluation metrics of different strategies during the iterative adjustment process.

6. The method according to claim 1, wherein, The original dataset is used for model training. The process involves processing the original dataset using each policy from the current policy set to obtain evaluation metrics for each policy, including: For each strategy in the current strategy set, the following steps are performed to obtain the evaluation metric for that strategy: The target dataset is obtained by processing the original dataset using the described strategy. Obtain the first model accuracy information obtained by training the model using the target dataset, and Based on the accuracy information of the first model, the evaluation index of the strategy is determined.

7. The method according to claim 6, wherein, The original dataset is used to train the target model, and the parameter size of the target model is larger than a preset size. The method further includes: compressing the target model to obtain a test model, and the parameter size of the test model is smaller than the preset size. The step of obtaining the first model accuracy information obtained by training the model using the target dataset includes: The test model is trained using the target dataset, and the accuracy information of the trained test model is used as the accuracy information of the first model.

8. The method according to claim 6, further comprising: Obtain the second model accuracy information obtained by training the model using the original dataset; The step of determining the evaluation index of the strategy based on the accuracy information of the first model includes: The accuracy information of the first model is compared with the accuracy information of the second model, and the evaluation index of the strategy is determined based on the comparison result.

9. The method according to claim 1, wherein, The iterative adjustment of the initial policy set based on the original dataset until a target policy suitable for the original dataset is obtained includes: The original dataset is sampled while maintaining its data distribution characteristics to obtain a sampled dataset; and The initial policy set is iteratively adjusted based on the sampled dataset until a target policy suitable for the original dataset is obtained.

10. The method according to claim 9, wherein, The step of sampling the original dataset while maintaining the data distribution characteristics to obtain a sampled dataset includes: Based on the level of noise in the data, the original dataset is divided into multiple sub-datasets; While maintaining the data distribution characteristics of each subset, sampling is performed on each subset to obtain the sampling results corresponding to each of the multiple subsets; and The sampling results corresponding to the multiple sub-datasets are merged to obtain the sampling dataset.

11. The method according to claim 10, wherein, For any one of the plurality of subsets, wherein the subset includes multiple data samples, the sampling process includes repeatedly performing the following steps until a preset sampling size is reached: Iterate through each data sample in the subset, and for the current data sample, determine the similarity between the embedding vector of the current data sample and the embedding vectors of other data samples, and use the similarity as the representativeness of the current data sample to the subset. From the subset of data, identify the target data sample with the highest representativeness; as well as The target data sample is added to the sampling result corresponding to the sub-dataset, and the target data sample is deleted from the sub-dataset.

12. The method according to claim 1, wherein, For any first policy in the current policy set, when the first policy includes a first operator and a second operator executed sequentially, the original dataset is processed using the first policy, including: The original dataset is divided into a first clean data subset and a first noisy data subset; The first operator is applied to the first noisy data subset to obtain a first result data subset, and the first result data subset is combined with the first clean data subset to obtain an intermediate dataset. The intermediate dataset is divided into a second clean data subset and a second noisy data subset; and The second operator is applied to the second noisy data subset to obtain a second result data subset, and the second result data subset is combined with the second clean data subset to obtain the target dataset.

13. The method according to claim 1, wherein, For any first policy in the current policy set, the original dataset is processed using the first policy, including: The system queries whether a historical dataset is stored in a preset storage space. This historical dataset is obtained by processing the original dataset using a second strategy, where the operator combination included in the second strategy is a prefix of the operator combination included in the first strategy. If so, the historical dataset is retrieved from the preset storage space, and the operators in the subsequent part of the first strategy are executed sequentially based on the historical dataset until the processed target dataset is obtained.

14. The method of claim 13, wherein after processing the original dataset using the first strategy, the method further comprises: The first strategy and the target dataset are stored in the preset storage space.

15. A model training method, comprising: Obtain the original dataset used to train the target model; Generate a target policy applicable to the original dataset, wherein the target policy is generated based on the method described in any one of claims 1 to 14; The original dataset is processed using the target strategy to obtain a training sample set; and The target model is trained using the training sample set.

16. An electronic device comprising: At least one storage medium storing at least one instruction set; as well as At least one processor is communicatively connected to the at least one storage medium, wherein the at least one processor reads the at least one instruction set and executes the method as described in any one of claims 1-15 according to the instructions of the at least one instruction set.

17. A computer-readable non-volatile storage medium, wherein, The computer-readable non-volatile storage medium stores at least one instruction set, which, when executed by at least one processor, implements the method as described in any one of claims 1-15.

Images