Task flow processing method and device in privacy computing scenario

Abstract

Embodiments of the present specification relate to a task flow processing method and device in a privacy computing scenario. The method comprises: obtaining first data to be calculated and first prompt text for describing user computing requirements; inputting the first data and the first prompt text into a large language model to obtain a plurality of first privacy computing components selected from a privacy computing component library and a first dependency relationship between each first privacy computing component output by the large language model; the large language model is fine-tuned by privacy computing related training data. Then, at least based on the each first privacy computing component and the first dependency relationship, a first DAG task flow is constructed.

Description

TECHNICAL FIELD

[0001] One or more embodiments of the present specification relate to the field of artificial intelligence, and in particular to a task flow processing method and device in a private computing scenario. BACKGROUND

[0002] Private Computing is an emerging computing mode that aims to protect personal data and privacy while enabling effective analysis and utilization of data. It can analyze and process data without revealing privacy through encryption, anonymization, differential privacy, and other technical means.

[0003] A private computing platform is a collection of system environments and tools built to implement private computing. It integrates various algorithms, protocols, libraries, and tools required for private computing, providing a convenient and efficient private computing development environment for users. However, due to the complexity of the principles of private computing, even if the private computing platform provides various ready-to-use tools for users, ordinary users still find it difficult to reasonably use these tools to achieve their own purposes. Therefore, a method is needed to better help users use various functions in the private computing platform to perform private computing tasks. SUMMARY

[0004] One or more embodiments of the present specification describe a task flow processing method and device in a private computing scenario to help users better use various functions on the private computing platform.

[0005] In a first aspect, a task flow processing method in a private computing scenario is provided, comprising:

[0006] obtaining first data to be calculated and a first prompt text describing user computing requirements;

[0007] inputting the first data and the first prompt text into a large language model to obtain a plurality of first private computing components selected from a private computing component library and a first dependency relationship between each first private computing component output by the large language model; the large language model is fine-tuned by private computing related training data;

[0008] constructing a first DAG task flow based at least on the plurality of first private computing components and the first dependency relationship.

[0009] In one possible implementation, the method further comprises:

[0010] send the first DAG task flow and the first data to an execution scheduler, so that the execution scheduler schedules each first privacy computing component to perform privacy computation on the first data according to the first dependency relationship, to obtain a first component product of each first privacy computing component and a first computation result of the first DAG task flow on the first data.

[0011] In a possible implementation, the parameter variable of each of the plurality of first privacy computing components has a default first parameter value; and the method further comprises:

[0012] obtaining second prompt text containing improvement requirements raised by the user based on the first component product and / or the first computation result;

[0013] inputting the second prompt text into the large language model to obtain second parameter values set by the large language model for the parameter variables of the plurality of first privacy computing components;

[0014] setting attribute values of corresponding component nodes in the first DAG task flow based on the second parameter values, to update the first DAG task flow, to obtain a second DAG task flow.

[0015] In a possible implementation, the method further comprises:

[0016] send the second DAG task flow and the first data to an execution scheduler, so that the execution scheduler updates the parameter variables of each first privacy computing component according to the second parameter values, and schedules each first privacy computing component to perform privacy computation on the first data according to the first dependency relationship, to obtain a second component product of each first privacy computing component and a second computation result of the second DAG task flow on the first data.

[0017] In a possible implementation, the parameter variable of each of the plurality of first privacy computing components has a default first parameter value; and the large language model further outputs an expected computation result of performing privacy computation on the first data; and the method further comprises:

[0018] inputting the first computation result and the expected computation result into the large language model to obtain third parameter values set by the large language model for the parameter variables of the plurality of first privacy computing components according to a difference between the first computation result and the expected computation result;

[0019] setting attribute values of corresponding component nodes in the first DAG task flow based on the third parameter values, to update the first DAG task flow, to obtain a third DAG task flow.

[0020] In a possible implementation, the parameter variable of each of the plurality of first privacy computing components has a default first parameter value; and the method further includes:

[0021] inputting the first calculation result and a user-input expected calculation result into the large language model to obtain fourth parameter values output by the large language model for parameter variable settings of the plurality of first privacy computing components according to a difference between the first calculation result and the expected calculation result;

[0022] setting attribute values of corresponding component nodes in the first DAG task flow based on the fourth parameter values to update the first DAG task flow, to obtain a fourth DAG task flow.

[0023] In a possible implementation, the large language model further outputs fifth parameter values of the parameter variables in each of the plurality of first privacy computing components; and a first DAG task flow is constructed based at least on the plurality of first privacy computing components and the first dependency relationship, including:

[0024] a first DAG task flow is constructed based on the first dependency relationship between the plurality of first privacy computing components, and attribute values of corresponding component nodes in the first DAG task flow are set based on the fifth parameter values.

[0025] In a possible implementation, the method further includes:

[0026] obtaining target data to be calculated, a target DAG task flow, and target prompt text; the target DAG task flow includes a plurality of target privacy computing components and target dependency relationships between the plurality of target privacy computing components; and the target prompt text is used to enable the large language model to perform rationality diagnosis on the target DAG task flow;

[0027] inputting the target data, the target DAG task flow, and the target prompt text into the large language model to obtain a first diagnosis result output by the large language model; the first diagnosis result is used to describe rationality of performing privacy computation on the target data according to the target DAG task flow.

[0028] In a possible implementation, the method further includes:

[0029] sending the target DAG task flow and the target data to an execution scheduler, so that the execution scheduler schedules the plurality of target privacy computing components to perform privacy computation on the target data according to the target dependency relationships, to obtain target component products of the plurality of target privacy computing components and a target calculation result of the target DAG task flow on the target data;

[0030] The target component product and the target calculation result are sent to at least a large language model to obtain a second diagnosis result output by the large language model; and the second diagnosis result is used to describe rationality of the target component product and the target calculation result.

[0031] In a possible implementation, the execution scheduler also outputs a target log in a privacy calculation process; and sending the target component product and the target calculation result to at least a large language model to obtain a second diagnosis result output by the large language model includes:

[0032] The target component product, the target calculation result and the target log are sent to a large language model to obtain a second diagnosis result output by the large language model.

[0033] In a second aspect, a large language model fine-tuning method is provided, including:

[0034] A training set is constructed, including first type training samples and second type training samples; input data of the first type training samples includes question texts for functions of each privacy calculation component in a privacy calculation component library, and labels include function description texts of each privacy calculation component; input data of the second type training samples includes data to be calculated and texts describing user calculation requirements, and labels include multiple privacy calculation components in the privacy calculation component library and dependency relationships between the privacy calculation components.

[0035] Based on the training set, a large language model is fine-tuned.

[0036] In a possible implementation, the training set further includes third type training samples, input data of the third type training samples includes texts describing improvement result requirements, and labels include parameter values of parameter variables of the multiple privacy calculation components.

[0037] In a possible implementation, the training set further includes fourth type training samples, input data of the fourth type training samples includes real calculation results and expected calculation results, and labels include parameter values of parameter variables of the multiple privacy calculation components.

[0038] In a possible implementation, the training set further includes fifth type training samples, input data of the fifth type training samples includes data to be calculated and texts describing user calculation requirements, and labels include multiple privacy calculation components in the privacy calculation component library, dependency relationships between the privacy calculation components, and parameter values of parameter variables of the multiple privacy calculation components.

[0039] In a possible implementation, the training set further includes a sixth type of training sample, the input data of the sixth type of training sample includes data to be calculated and text describing user computing requirements, and the label includes an expected calculation result of privacy calculation on the input data.

[0040] In a possible implementation, the training set further includes a seventh type of training sample, the input data of the seventh type of training sample includes data to be calculated and a DAG task flow, and the label includes a rationality diagnosis result of privacy calculation on the target data according to the DAG task flow.

[0041] In a possible implementation, the training set further includes an eighth type of training sample, the input data of the eighth type of training sample includes data to be calculated, a DAG task flow, a privacy calculation component product, and a privacy calculation result, and the label includes a rationality diagnosis result of the privacy calculation component product and the privacy calculation result.

[0042] In a third aspect, a task flow processing apparatus in a privacy calculation scenario is provided, and includes:

[0043] A first obtaining unit is configured to obtain first data to be calculated and first prompt text for describing user computing requirements.

[0044] A first dependency determining unit is configured to input the first data and the first prompt text into a large language model to obtain a plurality of first privacy calculation components selected from a privacy calculation component library and a first dependency relationship between the first privacy calculation components output by the large language model, wherein the large language model is fine-tuned by privacy calculation related training data.

[0045] A first constructing unit is configured to construct a first DAG task flow based at least on the first privacy calculation components and the first dependency relationship.

[0046] In a possible implementation, the apparatus further includes:

[0047] A first calculating unit is configured to send the first DAG task flow and the first data to an execution scheduler, so that the execution scheduler schedules the first privacy calculation components to perform privacy calculation on the first data according to the first dependency relationship, to obtain a first component product of each first privacy calculation component and a first calculation result of the first DAG task flow on the first data.

[0048] In a fourth aspect, a large language model fine-tuning apparatus is provided, and includes:

[0049] The construction unit is configured to construct a training set containing first-class training samples and second-class training samples; the input data of the first-class training samples contains question texts for the functions of each privacy computing component in the privacy computing component library, and the label contains the function description text of each privacy computing component; the input data of the second-class training samples contains data to be calculated and text describing user computing requirements, and the label contains multiple privacy computing components in the privacy computing component library and the dependency relationship between the privacy computing components.

[0050] The fine-tuning unit is configured to fine-tune the large language model based on the training set.

[0051] In a fifth aspect, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed in a computer, the computer program causes the computer to execute the method of the first aspect or the second aspect.

[0052] In a sixth aspect, a computing device is provided, which includes a memory and a processor. The memory stores executable code, and the processor executes the executable code to implement the method of the first aspect or the second aspect.

[0053] The task flow processing method and device in the privacy computing scenario provided by the embodiments of the present specification fine-tune a large language model based on a training set related to privacy computing, so that the large language model has knowledge related to privacy computing. Then, the fine-tuned large language model selects multiple privacy computing components from a privacy computing component library of a privacy computing platform based on user input data to be calculated and specific computing requirements of the user, and arranges the execution order of the privacy computing components. Then, according to the output content of the large language model, a corresponding DAG (Directed Acyclic Graph, directed acyclic graph) task flow is constructed to meet the computing requirements of the user. Further, the parameter values of the parameter variables of each computing component can be adjusted according to the specific execution of the DAG task flow and further requirements of the user, to generate a DAG task flow that is more suitable for the computing requirements of the user. In addition, the rationality of the DAG task flow itself can be diagnosed, and the rationality of the privacy computing result can be diagnosed, so that the user can more clearly understand the execution of the privacy computing task. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the multiple embodiments disclosed in the specification, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only a part of the embodiments disclosed in the specification, and other drawings can be obtained by those skilled in the art without creative labor.

[0055] Figure 1 An implementation scenario diagram showing a task flow processing method in a privacy computing scenario according to one embodiment is shown.

[0056] Figure 2 A flowchart showing a task flow processing method in a privacy computing scenario according to one embodiment is shown.

[0057] Figure 3 A diagram showing a DAG task flow according to one example is shown.

[0058] Figure 4 A diagram showing a DAG task flow according to another example is shown.

[0059] Figure 5 A flowchart showing a large language model fine-tuning method according to one embodiment is shown.

[0060] Figure 6 A schematic block diagram showing a task flow processing apparatus in a privacy computing scenario according to one embodiment is shown.

[0061] Figure 7 A schematic block diagram showing a large language model fine-tuning apparatus according to one embodiment is shown. DETAILED DESCRIPTION

[0062] The schemes provided in the present specification will be described below with reference to the accompanying drawings.

[0063] As mentioned previously, users often face very complex use scenarios in a privacy computing scenario. Due to the numerous privacy computing scenarios, there are also a large number of privacy computing components (hereinafter also referred to as “components”) involved. Privacy computing components are a key part of privacy computing technology, and are used to compute and analyze data while protecting data privacy. For example, feature engineering involves a large number of components such as missing value filling, one-hot encoding, WOE encoding (Weight of Evidence Encoding), feature standardization, feature derivation, and outlier processing; data preprocessing involves components such as conditional filtering, stratified sampling, and weighted sampling; statistical analysis involves components such as basic distribution, correlation coefficient, variance inflation factor, and feature importance; and machine learning involves components such as decision trees, deep neural networks, and convolutional neural networks. A single privacy computing component is often used to complete a relatively independent and complete subtask. Meanwhile, each privacy computing component contains a number of parameter variables, which are used to control the specific functions of the component. For example, the parameter variables in the missing value filling component can be used to control which numerical value is used to fill in the missing values, and the parameter variables in the deep neural network component can be the number of layers in the deep neural network, the number of nodes in each layer, and other hyperparameters.

[0064] In addition to the above examples, each of the various computing scenarios (e.g., modeling, prediction, result evaluation, etc.) corresponds to one or more privacy computing components, respectively. At the same time, each component in the privacy computing platform can involve a large number of implementation details and differences, and each is applicable to different scenarios. In addition, a complete computing process involves a complex and lengthy computing link, and a complete set of computing processes can use dozens of components or more. For example, a computing link can be: data reading, data preprocessing, feature engineering processing, modeling, prediction, and evaluation.

[0065] After designing the computing process, the computing process can be sent to the execution scheduler of the privacy computing platform. The execution scheduler will schedule the corresponding components according to the selected components in the computing process and the execution order between the components to perform privacy computing on the to-be-computed data, and finally complete the computing task.

[0066] After designing the computing process and completing the computing task, it is often necessary to evaluate whether the output of the entire link process meets the expectations and whether the overall effect of the computing process meets the expectations. On this basis, if it does not meet the expectations, some parameters of the components in the link may need to be modified for tuning, and then how to tune also requires the user to have corresponding knowledge reserves.

[0067] As can be seen from the above description, the complexity of privacy computing itself brings a high entry threshold to users. If a user wants to normally use the privacy computing platform to complete a privacy computing task and achieve the expected computing result, the user needs to be familiar with the functional characteristics of dozens or hundreds of components, component implementation details, related parameter tuning methods, and the like. This brings great difficulty to users. Based on this, the embodiments of the present specification provide a self-service intelligent process solution in a general privacy computing scenario, so that users can simply, efficiently and conveniently use the privacy computing platform.

[0068] Figure 1 An implementation scenario schematic diagram of a task flow processing method in a privacy computing scenario according to an embodiment is shown. In the implementation scenario, a user can use the privacy computing platform to complete a privacy computing task. Figure 1In the example, the privacy computing platform responsible for privacy computing at least includes a component library and an execution scheduler, wherein the component library contains a plurality of privacy computing components, and the execution scheduler is used to schedule each component to perform a related privacy computing task. A large language model is connected to the privacy computing platform and can access the components in the component library and interact with the execution scheduler to obtain the execution status of the privacy computing task, such as execution logs, computing results, and the like. The large language model is fine-tuned by privacy computing related training data and has various knowledge related to privacy computing, and can complete various tasks related to privacy computing according to the prompt text input by the user. The method of fine-tuning the large language model will be described in the subsequent part of the specification.

[0069] First, the to-be-computed data and the requirement prompt text describing the privacy computing requirement are input into the large language model. The requirement prompt text may be, for example: “I have a series of training data. Please first preprocess the training data according to its characteristics, and then train a binary classification model according to the training data.” The large language model will select one or more components from the component library according to the characteristics of the to-be-computed data and the specific content in the requirement prompt text, for example, in Figure 1 The selected components are component 1 to component n. The dependency relationship between each component is arranged.

[0070] Then, according to the n selected components and the dependency relationship between each component, a DAG task flow is constructed. The nodes in the DAG task flow can be each component, and the directed edges between the nodes represent the execution order and dependency relationship between the components represented by the two nodes connected by the edges. The DAG task flow can be used to describe the execution process of the privacy computing task.

[0071] After obtaining the corresponding DAG task flow, the DAG task flow and the to-be-computed data can be sent to the execution scheduler, so that the execution scheduler schedules each component to perform privacy computing on the to-be-computed data according to the execution process described by the DAG task flow, obtains the component product of each output of each component, and the computing result of the DAG task flow as a whole on the to-be-computed data.

[0072] The above describes the basic process of using a large language model to assist in privacy computing. Further, the large language model can also evaluate the rationality (whether it meets the expectations) of the privacy computing result according to the component product and / or the computing result. Or, the user himself gives a DAG task flow and describes his computing requirement through a prompt text, and the large language model judges the rationality (whether it can meet the computing requirement) of the DAG task flow according to the computing requirement. These contents will be described in detail in the subsequent part of the embodiment of the specification.

[0073] The specific implementation steps of the task flow processing method in the privacy computing scenario are described below in combination with specific embodiments.

[0074] Figure 2 A flowchart of a task flow processing method in a privacy computing scenario according to an embodiment is shown, and the execution subject of the method can be any platform or server or device cluster with computing and processing capabilities, etc. As shown in the figure, the method at least includes: step 202, obtaining first data to be calculated and first prompt text for describing user computing requirements; step 204, inputting the first data and the first prompt text into a large language model to obtain a plurality of first privacy computing components selected from a privacy computing component library and a first dependency relationship between each first privacy computing component output by the large language model; the large language model is fine-tuned by privacy computing related training data; step 206, constructing a first DAG task flow based on at least the first privacy computing components and the first dependency relationship. Figure 2

[0075] The specific execution process of each step is described below.

[0076] First, in step 202, the first data to be calculated and the first prompt text for describing the user's computing requirements are obtained.

[0077] The first data can be single-source data, for example, it can come from a single data table of a single database, or from a single file; it can also be multi-source data, for example, it can come from multiple data tables of a single database, or multiple data tables of multiple databases, or from multiple files. Here, no limitation is made.

[0078] The first prompt text is used to describe the computing requirements, which can include a description of the first data, a description of the privacy computing task. Further, it can also include a description of specific requirements for certain privacy computing components, and a description of the expected standards achieved for the privacy computing results.

[0079] The description of the first data can include the type, attribute, unit, etc. of each field of the first data. If the first data contains image data, the description of the first data can also include a general summary of the content contained in the image. The description of the privacy computing task can include the task type, the method to be used, etc. The description of the specific requirements for certain privacy computing components can include specifying the use of certain components, specifying the value of a certain parameter of a certain component, etc. The expected standards for the privacy computing results can include making a specific measurement index (accuracy, similarity, etc.) reach a set target value, etc.

[0080] ​According to the above, in one embodiment, the first prompt text can be: "I have some training data saved in the xxx path, and then train a classification model according to these data."

[0081] In another embodiment, the first prompt text can be: "I have some training data saved in the xxx database xxx data table, the first column of the training data represents xxx, the second column represents xxx… The last column is the label. These data are noisy, please preprocess them first, then divide the training data into training set and test set according to the ratio of 7:3, train a multi-classification model based on deep neural network. You need to search for hyperparameters according to the performance of the model, and constantly adjust the hyperparameter combination in the model for training and evaluation, so that the final prediction accuracy of the model reaches 90%."

[0082] Next, in step 204, the first data and the first prompt text are input into the large language model to obtain a plurality of first privacy computing components selected from the privacy computing component library and a first dependency relationship between each first privacy computing component output by the large language model; the large language model is fine-tuned by privacy computing related training data.

[0083] Among them, inputting the first data into the large language model can be directly inputting the data content of the first data into the large language model, or inputting the data address saving the first data into the large language model, so that the large language model reads the first data from the data address.

[0084] The first dependency relationship between each first privacy computing component can be the order of calling each component, whether two or more components can be called in parallel, etc. If the output data of one component is the input of another component, or when one call is executed, another component is called, then the two components can be serially dependent on each other. If two components can be called at the same time, then the two components can be parallel dependent.

[0085] It should be noted that the plurality of first privacy computing components can include the same component, that is, a component can be used multiple times in different stages of privacy computing, and the parameters of each use of the component can be different.

[0086] The large language model can be any pre-trained large language model, which is fine-tuned by privacy computing related training data and knows the knowledge related to privacy computing platform and privacy computing.

[0087] Then, in step 206, at least based on the first dependency relationship between each first privacy computing component, a first DAG task flow is constructed.

[0088] Each node in the DAG task flow represents a first privacy computing component, and the directed edge between the nodes represents the calling order between the components. When there is a calling order between two components, for example, the calling of component 2 depends on the execution of component 1 being completed, then in the DAG task flow, there is a directed edge between the nodes of component 1 and component 2, which is pointed from the node of component 1 to the node of component 2. When there is no dependency between two components, there is no connecting edge between the two nodes in the DAG task flow.

[0089] For example, in an embodiment, the plurality of first privacy computing components selected by the large language model are: data reading, null value processing, virtual fusion, model training, model prediction, and exporting prediction results. These components are in the order of the preceding sequence, and the execution of the latter component depends on the completion of the former component. Then, the first DAG task flow can be constructed according to the first dependency relationship between each first privacy computing component as shown in Figure 3 . Figure 3 A schematic diagram of a DAG task flow according to an example is shown.

[0090] In another embodiment, the plurality of first privacy computing components selected by the large language model are: data reading, null value processing, virtual fusion, model training, model prediction, exporting prediction results, and exporting model parameters. Among them, the six components of data reading, null value processing, virtual fusion, model training, model prediction, and exporting prediction results are in sequence, and the execution of the latter component depends on the completion of the former component. While the exporting model parameters component depends on the execution of the model training component and can be executed in parallel with the model prediction component. Then, the first DAG task flow can be constructed according to the first dependency relationship between each first privacy computing component as shown in Figure 4 . Figure 4 A schematic diagram of a DAG task flow according to another example is shown.

[0091] In the foregoing embodiments, the large language model does not set specific parameter values for the parameter variables of each component. In these embodiments, the parameter variables in each component can take default values.

[0092] In some possible embodiments, the large language model also outputs the fifth parameter value of the parameter variable in each first privacy computing component. Step 206 specifically includes: constructing a first DAG task flow based on the first dependency relationship between each first privacy computing component, and setting the attribute value of the corresponding component node in the first DAG task flow based on the fifth parameter value.

[0093] By setting the parameter variable as the attribute of the node corresponding to the component in the DAG task flow, and setting the parameter value as the corresponding attribute value, the component parameter value output by the large language model can also be stored in the DAG task flow.

[0094] The above describes a process of using a fine-tuned large language model to assist in generating a DAG task flow.

[0095] In some possible implementations, after step 206, the method further includes:

[0096] Step 208, sending the first DAG task flow and the first data to an execution scheduler, so that the execution scheduler schedules each first privacy computing component to perform privacy computation on the first data according to the first dependency relationship, to obtain a first component product of each first privacy computing component and a first computation result of the first DAG task flow on the first data.

[0097] The execution scheduler is located in a privacy computing platform and can schedule each privacy computing component in the privacy computing platform to complete the task indicated by the DAG task flow. The output content of each first privacy computing component is its corresponding first component product. The first computation result is the final output result after the execution of the first DAG task flow.

[0098] Through the above steps 202 to 208, the user can use the fine-tuned large language model to assist in the design of a privacy computing task, which specifically embodies selecting appropriate privacy computing components and arranging the dependency relationship between the components to construct a corresponding DAG task flow. The execution scheduler schedules each privacy computing component according to the content in the DAG task flow to perform privacy computation on the data to be computed to complete the privacy computing task.

[0099] In some possible implementations, the parameter variables of the plurality of first privacy computing components each have a default first parameter value. When the user views each first component product and the first computation result, the user will also want to further improve the privacy computing task. These improvement requirements can be included in a second prompt text and form a multi-round dialogue with the aforementioned first prompt text and the answer of the large language model (the plurality of first privacy computing components and the first dependency relationship between each first privacy computing component). The large language model can help the user to improve the task flow based on the context information in the previous round of dialogue and the content in the second prompt text. At this time, the method further includes steps 210 to 214.

[0100] In step 210, a second prompt text is obtained, which contains improvement requirements proposed by the user based on each first component product and / or the first computation result.

[0101] The second prompt text is used to describe the improvement requirement for the privacy computing result, which can include an indication for the large language model to adjust the component parameter variable to improve the value of a specific evaluation index. Further, it can also include an indication for adjusting the parameter variable in a specific component.

[0102] Then, in step 212, the second prompt text is input into the large language model to obtain second parameter values of the large language model for parameter variable settings of the first privacy computing components.

[0103] The large language model can analyze the first component products and / or the first computing result according to the indication in the second prompt text, in combination with the first prompt text and the first data, and then output the improved second parameter values.

[0104] Next, in step 214, the attribute values of the corresponding component nodes in the first DAG task flow are set based on the second parameter values to update the first DAG task flow, obtaining a second DAG task flow.

[0105] After obtaining the second DAG task flow according to steps 210 to 214, in some possible implementations, after step 214, the method further includes:

[0106] In step 216, the second DAG task flow and the first data are sent to an execution scheduler, so that the execution scheduler updates the parameter variables of the first privacy computing components according to the second parameter values, and schedules the updated first privacy computing components to perform privacy computation on the first data according to the first dependency relationship, obtaining second component products of the first privacy computing components and a second computing result of the second DAG task flow on the first data.

[0107] Through the above steps 210 to 216, the user can use the fine-tuned large language model to adjust the parameter variables of each component in the first DAG task flow according to the computing result (each first component product and / or first computing result) in the first round of privacy computation, and construct a second DAG task flow based on the new parameters. Then, the execution scheduler schedules each privacy computing component according to the content in the second DAG task flow to perform privacy computation on the data to be computed, so as to optimize the computing result of the privacy computation task.

[0108] In some possible implementation manners, the parameter variable of each of the plurality of first privacy computing components has a default first parameter value. Meanwhile, the large language model also outputs an expected calculation result of the privacy calculation on the first data. After obtaining the first calculation result in step 208, the value of the parameter variable in the privacy computing component can be adjusted according to the difference between the first calculation result and the expected calculation result. At this time, after step 208, the method further includes steps 218 to 220.

[0109] In step 218, the first calculation result and the expected calculation result are input into the large language model to obtain third parameter values of parameter variable settings of the plurality of first privacy computing components output by the large language model according to the difference between the first calculation result and the expected calculation result.

[0110] The large language model can adjust the value of the parameter variable in each of the plurality of first privacy computing components according to the difference between the first calculation result and the expected calculation result, and then output improved third parameter values.

[0111] Then, in step 220, the attribute value of the corresponding component node in the first DAG task flow is set based on the third parameter values to update the first DAG task flow, and a third DAG task flow is obtained.

[0112] After obtaining the third DAG task flow, the calculation task in the third DAG task flow can also be scheduled and executed by referring to the method similar to step 216, which will not be described herein again.

[0113] In some possible implementation manners, the parameter variable of each of the plurality of first privacy computing components has a default first parameter value. Meanwhile, the user himself / herself can have a corresponding expected calculation result of the privacy calculation task. After obtaining the first calculation result in step 208, the value of the parameter variable in the privacy computing component can be adjusted according to the difference between the first calculation result and the expected calculation result. At this time, after step 208, the method further includes steps 222 to 224.

[0114] In step 222, the first calculation result and the expected calculation result input by the user are input into the large language model to obtain fourth parameter values of parameter variable settings of the plurality of first privacy computing components output by the large language model according to the difference between the first calculation result and the expected calculation result.

[0115] The large language model can adjust the value of the parameter variable in each of the plurality of first privacy computing components according to the difference between the first calculation result and the expected calculation result, and then output improved fourth parameter values.

[0116] At step 224, property values of corresponding component nodes in the first DAG task flow are set based on the fourth parameter value, so as to update the first DAG task flow and obtain a fourth DAG task flow.

[0117] After obtaining the fourth DAG task flow, the computing tasks in the fourth DAG task flow can also be scheduled for execution by referring to the method similar to step 216, which will not be described herein.

[0118] The above steps describe the process of designing and optimizing a privacy computing task using the fine-tuned large language model according to user requirements. In addition, the large language model can also be used to diagnose the rationality of the DAG task flow itself after the DAG task flow is constructed. Alternatively, the user can also design the DAG task flow and then submit it to the large language model for rationality diagnosis.

[0119] In some possible implementations, after step 206, the method further includes steps 226 and 228.

[0120] At step 226, target data to be computed, a target DAG task flow, and a target prompt text are obtained; the target DAG task flow includes a plurality of target privacy computing components and target dependency relationships between the target privacy computing components; and the target prompt text is used to enable the large language model to perform rationality diagnosis on the target DAG task flow.

[0121] The target data to be computed can be the first data in step 202, or can be different from the first data. The target DAG task flow can be the first DAG task flow in step 206, or can be a DAG task flow uploaded by the user and different from the first DAG task flow.

[0122] The target prompt text is used to describe the diagnosis requirement, which can include an instruction for the large language model to determine whether privacy computing on the target data according to the target DAG task flow can achieve an expected result.

[0123] Then, at step 228, the target data, the target DAG task flow, and the target prompt text are input into the large language model to obtain a first diagnosis result output by the large language model; the first diagnosis result is used to describe the rationality of privacy computing on the target data according to the target DAG task flow.

[0124] The first diagnosis result can include an evaluation of the target DAG task flow, and further include suggestions for improvement of the target DAG task flow. For example, specific privacy computing components can be added / removed based on the target DAG task flow, or the execution order between existing components can be adjusted, and the like.

[0125] In some possible implementations, the large language model can also diagnose the reasonableness of the calculation results obtained by executing the DAG task flow. After step 228, the method further includes steps 230 to 232.

[0126] In step 230, the target DAG task flow and the target data are sent to the execution scheduler, so that the execution scheduler schedules each target privacy computing component to perform privacy computation on the target data according to the target dependency relationship, obtains a target component product of each target privacy computing component, and obtains a target calculation result of the target DAG task flow on the target data.

[0127] Then, in step 232, at least the target component product and the target calculation result are sent to the large language model to obtain a second diagnosis result output by the large language model; the second diagnosis result is used to describe the reasonableness of the target component product and the target calculation result.

[0128] In one embodiment, the execution scheduler also outputs a target log in the privacy computation process. Step 232 specifically includes: sending the target component product, the target calculation result, and the target log to the large language model to obtain a second diagnosis result output by the large language model.

[0129] The DAG task flow may have design defects, and meanwhile, the execution scheduler may have various unexpected errors in the process of executing the task. Through steps 230 to 232, the reasonableness of the calculation result of the DAG task flow can be diagnosed.

[0130] According to the above steps, the large premonition model with privacy computation and privacy computation platform related knowledge after fine-tuning is used to assist users in specific operations such as privacy computation task design, privacy computation task execution, privacy computation task diagnosis, and privacy computation task optimization, so as to simplify the use difficulty of the users.

[0131] According to another aspect, an embodiment of a large language model fine-tuning method is also provided. Figure 5 A flowchart of a large language model fine-tuning method according to an embodiment is shown. The execution subject of the method can be any platform or server or device cluster with computing and processing capabilities, etc. As shown in the figure, Figure 5As shown, the method at least comprises: step 502, constructing a training set containing first type training samples and second type training samples; the input data of the first type training samples contains question texts for the functions of each privacy computing component in the privacy computing component library, and the label contains the function description texts of each privacy computing component; the input data of the second type training samples contains data to be calculated and texts describing user computing requirements, and the label contains multiple privacy computing components in the privacy computing component library and the dependency relationships between each privacy computing component; step 504, fine-tuning the large language model based on the training set.

[0132] The specific execution process of each of the above steps is described below.

[0133] In step 502, a training set containing first type training samples and second type training samples is constructed; the input data of the first type training samples contains question texts for the functions of each privacy computing component in the privacy computing component library, and the label contains the function description texts of each privacy computing component; the input data of the second type training samples contains data to be calculated and texts describing user computing requirements, and the label contains multiple privacy computing components in the privacy computing component library and the dependency relationships between each privacy computing component.

[0134] The first type training samples are used to fine-tune the large language model so that it learns the functions of each privacy computing component in the privacy computing component library. The second type training samples are used to fine-tune the large language model so that it learns how to design related privacy computing tasks according to the data to be calculated and user requirements.

[0135] Then, in step 504, the large language model is fine-tuned based on the training set.

[0136] Any fine-tuning method can be used to fine-tune the large language model, such as the P-Tuning method, the Prefix-Tuning method, the Adapter Tuning method, etc., which are not limited here.

[0137] In some possible implementations, the training set further includes third type training samples, and the input data of the third type training samples contains texts describing improvement result requirements, and the label contains parameter values of parameter variables of multiple privacy computing components.

[0138] The third type training samples are used to fine-tune the large language model so that it learns how to adjust the parameter values of parameter variables in the privacy computing component according to the user's improvement requirements.

[0139] In some possible implementation manners, the training set further includes fourth type training samples, the input data of the fourth type training samples includes the real calculation result and the expected calculation result, and the label includes the parameter values of the parameter variables of the plurality of privacy computing components.

[0140] The fourth type training samples are used to fine-tune the large language model to learn how to adjust the parameter values of the parameter variables in the privacy computing components according to the difference between the real calculation result and the expected calculation result.

[0141] In some possible implementation manners, the training set further includes fifth type training samples, the input data of the fifth type training samples includes the data to be calculated and the text describing the user calculation demand, and the label includes the plurality of privacy computing components in the privacy computing component library, the dependency relationship between the privacy computing components, and the parameter values of the parameter variables of the pluralityity of privacy computing components.

[0142] The fifth type training samples are used to fine-tune the large language model to learn how to design the related privacy computing task and give the parameter values of the parameter variables in the privacy computing components according to the data to be calculated and the user demand.

[0143] In some possible implementation manners, the training set further includes sixth type training samples, the input data of the sixth type training samples includes the data to be calculated and the text describing the user calculation demand, and the label includes the expected calculation result of the privacy calculation performed on the input data.

[0144] The sixth type training samples are used to fine-tune the large language model to learn how to predict the expected calculation result of the privacy computing task according to the data to be calculated and the user demand.

[0145] In some possible implementation manners, the training set further includes seventh type training samples, the input data of the seventh type training samples includes the data to be calculated and the DAG task flow, and the label includes the rationality diagnosis result of the privacy calculation performed on the target data according to the DAG task flow.

[0146] The seventh type training samples are used to fine-tune the large language model to learn how to diagnose the rationality of the DAG task flow according to the data to be calculated and the DAG task flow.

[0147] In some possible implementation manners, the training set further includes eighth type training samples, the input data of the eighth type training samples includes the data to be calculated, the DAG task flow, the privacy computing component product and the privacy calculation result, and the label includes the rationality diagnosis result of the privacy computing component product and the privacy calculation result.

[0148] The eighth type of training sample is used to fine-tune the large language model to learn how to diagnose the rationality of the privacy computing component product and the privacy computing result obtained by executing the DAG task flow according to the data to be calculated, the DAG task flow, the privacy computing component product, and the privacy computing result.

[0149] By the method shown, the large language model can be fine-tuned to be familiar with various knowledge related to privacy computing and privacy computing platforms. Figure 5

[0150] According to another aspect, embodiments also provide a task flow processing apparatus in a privacy computing scenario. Figure 6 A schematic block diagram of a task flow processing apparatus in a privacy computing scenario according to an embodiment is shown, which can be deployed in any device, platform or device cluster with computing and processing capabilities. As shown, Figure 6 The apparatus 600 includes:

[0151] The first obtaining unit 602 is configured to obtain first data to be calculated and first prompt text for describing user computing requirements.

[0152] The first dependency determining unit 604 is configured to input the first data and the first prompt text into a large language model to obtain a plurality of first privacy computing components selected from a privacy computing component library and a first dependency relationship between each first privacy computing component output by the large language model; the large language model is fine-tuned by privacy computing related training data.

[0153] The first constructing unit 606 is configured to construct a first DAG task flow based at least on each first privacy computing component and the first dependency relationship.

[0154] In some possible implementations, the apparatus 600 further includes:

[0155] The first computing unit 608 is configured to send the first DAG task flow and the first data to an execution scheduler, so that the execution scheduler schedules each first privacy computing component to perform privacy computing on the first data according to the first dependency relationship, to obtain a first component product of each first privacy computing component and a first computing result of the first DAG task flow on the first data.

[0156] In some possible implementations, each parameter variable of the plurality of first privacy computing components has a default first parameter value; the apparatus 600 further includes:

[0157] The second obtaining unit 610 is configured to obtain second prompt text containing improvement requirements raised by a user based on each first component product and / or the first computing result.​

[0158] The first parameter determination unit 612 is configured to input the second prompt text into the large language model to obtain second parameter values of parameter variables of the plurality of first privacy computing components set by the large language model.

[0159] The second construction unit 614 is configured to set attribute values of corresponding component nodes in the first DAG task flow based on the second parameter values to update the first DAG task flow to obtain a second DAG task flow.

[0160] In some possible implementation manners, the apparatus 600 further includes:

[0161] The second calculation unit 616 is configured to send the second DAG task flow and the first data to an execution scheduler, so that the execution scheduler updates parameter variables of each first privacy computing component according to the second parameter values, and schedules each first privacy computing component to perform privacy computation on the first data according to the first dependency relationship after being updated, to obtain a second component product of each first privacy computing component and a second calculation result of the second DAG task flow on the first data.

[0162] In some possible implementation manners, the parameter variables of the plurality of first privacy computing components each have a default first parameter value; the large language model further outputs an expected calculation result of performing privacy computation on the first data; and the apparatus 600 further includes:

[0163] The second parameter determination unit 618 is configured to input the first calculation result and the expected calculation result into the large language model to obtain third parameter values of parameter variables of the plurality of first privacy computing components set by the large language model according to a difference between the first calculation result and the expected calculation result.

[0164] The third construction unit 620 is configured to set attribute values of corresponding component nodes in the first DAG task flow based on the third parameter values to update the first DAG task flow to obtain a third DAG task flow.

[0165] In some possible implementation manners, the parameter variables of the plurality of first privacy computing components each have a default first parameter value; and the apparatus 600 further includes:

[0166] The third parameter determination unit 622 is configured to input the first calculation result and an expected calculation result input by a user into the large language model to obtain fourth parameter values of parameter variables of the plurality of first privacy computing components set by the large language model according to a difference between the first calculation result and the expected calculation result.

[0167] The fourth construction unit 624 is configured to set attribute values of corresponding component nodes in the first DAG task flow based on the fourth parameter value, to update the first DAG task flow, and obtain a fourth DAG task flow.

[0168] In some possible implementation manners, the apparatus 600 further includes:

[0169] The third obtaining unit 626 is configured to obtain target data to be calculated, a target DAG task flow, and target prompt text; the target DAG task flow includes a plurality of target privacy calculation components and target dependency relationships between the target privacy calculation components; and the target prompt text is used to enable the large language model to perform rationality diagnosis on the target DAG task flow.

[0170] The first rationality diagnosis unit 628 is configured to input the target data, the target DAG task flow, and the target prompt text into the large language model, to obtain a first diagnosis result output by the large language model; and the first diagnosis result is used to describe rationality of performing privacy calculation on the target data according to the target DAG task flow.

[0171] In some possible implementation manners, the apparatus 600 further includes:

[0172] The third calculation unit 630 is configured to send the target DAG task flow and the target data to an execution scheduler, to enable the execution scheduler to schedule the target privacy calculation components to perform privacy calculation on the target data according to the target dependency relationships, to obtain target component products of the target privacy calculation components and a target calculation result of the target DAG task flow on the target data.

[0173] The first rationality diagnosis unit 632 is configured to send at least the target component products and the target calculation result to the large language model, to obtain a second diagnosis result output by the large language model; and the second diagnosis result is used to describe rationality of the target component products and the target calculation result.

[0174] According to another aspect, an embodiment of a large language model fine-tuning apparatus is also provided. Figure 7 A schematic block diagram of a large language model fine-tuning apparatus according to an embodiment is shown, which can be deployed in any device, platform, or device cluster with computing and processing capabilities. As shown in Figure 7 The apparatus 700 includes:

[0175] The construction unit 702 is configured to construct a training set containing first-class training samples and second-class training samples. The input data of the first-class training samples contains question texts for the functions of each privacy computing component in the privacy computing component library, and the label contains the function description text of each privacy computing component. The input data of the second-class training samples contains data to be calculated and text describing user computing requirements, and the label contains multiple privacy computing components in the privacy computing component library and the dependency relationship between each privacy computing component.

[0176] The fine-tuning unit 704 is configured to fine-tune the large language model based on the training set.

[0177] According to another aspect, an embodiment also provides a computer readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method described in any of the above embodiments.

[0178] According to another aspect, an embodiment also provides a computing device comprising a memory and a processor, wherein the memory has stored executable code, and the processor executes the executable code to implement the method described in any of the above embodiments.

[0179] Each of the above embodiments in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

[0180] The above describes specific embodiments of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in an order different than the order in the embodiments and still achieve the desired result. In addition, the processes depicted in the 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 or possible.

[0181] It is to be noted that, in the present text, the relative terms such as first and second, and the like are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between such entities or operations. Moreover, the terms "comprising", "including", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0182] It is understood by those skilled in the art that all or part of the steps of the above-mentioned embodiments can be completed by hardware, or by program to instruct relevant hardware, and the program can be stored in a computer readable storage medium, such as read-only memory, magnetic disk or optical disk.

[0183] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A task flow processing method in a privacy computing scenario, comprising: Obtain the first data to be calculated, and the first prompt text to describe the user's calculation needs; The first data and the first prompt text are input into the large language model to obtain multiple first privacy computing components selected from the privacy computing component library, as well as the first dependency relationships between the various first privacy computing components. The large language model was fine-tuned using privacy-preserving computation-related training data. A first DAG task flow is constructed based at least on each of the first privacy computing components and the first dependency relationship.

2. The method according to claim 1, further comprising: The first DAG task flow and the first data are sent to the execution scheduler, so that the execution scheduler schedules each first privacy computing component to perform privacy computing on the first data according to the first dependency relationship, so as to obtain the first component product of each first privacy computing component and the first computing result of the first DAG task flow with respect to the first data.

3. The method according to claim 2, wherein, Each of the plurality of first privacy computing components has a default first parameter value for its parameter variable; the method further includes: Obtain a second prompt text, which contains the user's improvement requests based on the outputs of each first component and / or the first calculation results; The second prompt text is input into the large language model to obtain the second parameter value set by the large language model for the parameter variables of several first privacy computing components; Based on the second parameter value, the attribute value of the corresponding component node in the first DAG task flow is set to update the first DAG task flow and obtain the second DAG task flow.

4. The method according to claim 3, further comprising: The second DAG task flow and the first data are sent to the execution scheduler, which updates the parameter variables of each first privacy computing component according to the second parameter value, and schedules each updated first privacy computing component to perform privacy computing on the first data according to the first dependency relationship, so as to obtain the second component product of each first privacy computing component and the second calculation result of the second DAG task flow with respect to the first data.

5. The method according to claim 2, wherein, Each of the multiple first privacy computing components has a default first parameter value for its parameter variable; The large language model also outputs the expected computation result of privacy computation on the first data; the method further includes: The first calculation result and the expected calculation result are input into the large language model to obtain the third parameter value set by the large language model for the parameter variables of several first privacy computing components based on the difference between the first calculation result and the expected calculation result. Based on the third parameter value, the attribute values ​​of the corresponding component nodes in the first DAG task flow are set to update the first DAG task flow and obtain the third DAG task flow.

6. The method according to claim 2, wherein, Each of the plurality of first privacy computing components has a default first parameter value for its parameter variable; the method further includes: The first calculation result and the expected calculation result input by the user are input into the large language model to obtain the fourth parameter value set by the large language model for the parameter variables of several first privacy computing components based on the difference between the first calculation result and the expected calculation result. Based on the fourth parameter value, the attribute values ​​of the corresponding component nodes in the first DAG task flow are set to update the first DAG task flow and obtain the fourth DAG task flow.

7. The method according to claim 1, wherein, The large language model also outputs the fifth parameter value of the parameter variables in each of the first privacy computing components; Based at least on the aforementioned first privacy computing components and first dependencies, a first DAG task flow is constructed, including: Based on the first dependencies of each first privacy computing component, a first DAG task flow is constructed, and the attribute values ​​of the corresponding component nodes in the first DAG task flow are set based on the fifth parameter value.

8. The method according to claim 1, further comprising: Obtain the target data to be calculated, the target DAG task flow, and the target prompt text; The target DAG task flow contains multiple target privacy computing components, as well as target dependencies between these components; The target prompt text is used to enable the large language model to perform a rationality diagnosis on the target DAG task flow; The target data, target DAG task flow, and target prompt text are input into the large language model to obtain a first diagnostic result output by the large language model; the first diagnostic result is used to describe the rationality of performing privacy calculations on the target data based on the target DAG task flow.

9. The method according to claim 8, further comprising: The target DAG task flow and target data are sent to the execution scheduler, so that the execution scheduler schedules each target privacy computing component to perform privacy computing on the target data according to the target dependency relationship, and obtains the target component products of each target privacy computing component, as well as the target computing results of the target DAG task flow with respect to the target data; At least the target component product and the target calculation result are sent to the large language model to obtain the second diagnostic result output by the large language model; The second diagnostic result is used to describe the rationality of the target component product and the target calculation result.

10. The method according to claim 9, wherein, The execution scheduler also outputs a target log during the privacy computation process; at least the target component output and the target computation result are sent to the large language model to obtain a second diagnostic result output by the large language model, including: The target component output, the target calculation result, and the target log are sent to the large language model to obtain the second diagnostic result output by the large language model.

11. A method for fine-tuning a large language model, comprising: Construct a training set, which includes training samples of the first class and training samples of the second class; The input data of the first type of training samples contains question texts about the functions of each privacy computing component in the privacy computing component library, and the tags contain functional description texts of each privacy computing component. The input data of the second type of training samples includes data to be computed and text describing the user's computing needs. The labels contain multiple privacy computing components in the privacy computing component library, as well as the dependencies between the privacy computing components. Based on the training set, fine-tune the large language model.

12. The method according to claim 11, wherein, The training set also includes a third type of training samples, the input data of which contains text describing the requirements for improving the results, and the labels contain parameter values ​​of parameter variables of multiple privacy computing components.

13. The method according to claim 11, wherein, The training set also includes a fourth type of training samples, whose input data contains real computation results and expected computation results, and whose labels contain parameter values ​​of parameter variables of multiple privacy computing components.

14. The method according to claim 11, wherein, The training set also includes a fifth type of training samples. The input data of the fifth type of training samples includes data to be computed and text describing the user's computing needs. The labels include multiple privacy computing components in the privacy computing component library, the dependencies between the privacy computing components, and the parameter values ​​of the parameter variables of the multiple privacy computing components.

15. The method according to claim 11, wherein, The training set also includes a sixth type of training samples. The input data of the sixth type of training samples contains data to be calculated and text describing the user's calculation needs. The labels contain the expected calculation results for privacy calculations on the input data.

16. The method according to claim 11, wherein, The training set also includes a seventh type of training samples. The input data of the seventh type of training samples includes the data to be calculated and the DAG task flow. The labels include a reasonableness diagnosis result describing the privacy calculation of the data to be calculated based on the DAG task flow.

17. The method according to claim 11, wherein, The training set also includes an eighth type of training samples. The input data of the eighth type of training samples includes data to be computed, DAG task flow, privacy computing component products and privacy computing results. The labels include reasonableness diagnostic results describing the privacy computing component products and privacy computing results.

18. A task flow processing device for a privacy computing scenario, comprising: The first acquisition unit is configured to acquire the first data to be calculated and the first prompt text used to describe the user's calculation needs; The first dependency determination unit is configured to input the first data and the first prompt text into the large language model to obtain multiple first privacy computing components selected from the privacy computing component library output by the large language model, and the first dependency relationship between each first privacy computing component; The large language model was fine-tuned using privacy-preserving computation-related training data. The first building unit is configured to build a first DAG task flow based at least on the respective first privacy computing components and the first dependency relationship.

19. The apparatus of claim 18, further comprising: The first computing unit is configured to send the first DAG task flow and the first data to the execution scheduler, so that the execution scheduler schedules each first privacy computing component to perform privacy computing on the first data according to the first dependency relationship, and obtains the first component product of each first privacy computing component and the first computing result of the first DAG task flow with respect to the first data.

20. A large language model fine-tuning device, comprising: The construction unit is configured to construct a training set, which includes first-class training samples and second-class training samples. The input data of the first type of training samples contains question texts about the functions of each privacy computing component in the privacy computing component library, and the tags contain functional description texts of each privacy computing component. The input data of the second type of training samples includes data to be computed and text describing the user's computing needs. The labels contain multiple privacy computing components in the privacy computing component library, as well as the dependencies between the privacy computing components. The fine-tuning unit is configured to fine-tune the large language model based on the training set.

21. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-17.

22. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-17.

Images