Instruction data generation method and device, computer readable storage medium, electronic equipment and computer program product

Abstract

The embodiment of the invention discloses an instruction data generation method and device, a computer readable storage medium, electronic equipment and a computer program product. The instruction data generation method can comprise the steps that task description information and first meta-prompt information of a target task are obtained; inputting the task description information and the first meta-prompt information into a first large language model, so that the first large language model can perform iterative optimization on initial instruction data generated based on the task description information under the guidance of the first meta-prompt information; according to the method, the target instruction data for guiding the target large model to execute the target task is obtained according to the task example information of the target task, and the target instruction data for guiding the target large model to execute the target task can be obtained by performing fusion processing on the candidate instruction data and the task example information of the target task.

Description

Technical Field

[0001] This specification relates to the technical field of large models, and particularly to a method and device for generating instruction data, a computer-readable storage medium, an electronic device, and a computer program product. Background Art

[0002] Artificial Intelligence Large Models (referred to as large models for short) can generally refer to machine learning models trained with large-scale data and having large-scale parameters and high computing power. These large models usually have a high degree of generality and generalization ability. Therefore, the Instruction Tuning technology can be used to fine-tune the parameters of the pre-trained large model to enhance the large model's ability to understand and execute specific tasks, so that the large model can better complete various downstream tasks. Currently, when fine-tuning a large model using instruction tuning, constructing a high-quality instruction dataset for guiding the large model to execute various tasks is crucial for improving the performance of the large model. It can not only help the model better understand human intentions but also improve the model's performance on various tasks. Based on this, how to efficiently and conveniently generate high-quality instruction data for guiding the large model to execute various tasks has become a technical problem that urgently needs to be solved. Summary of the Invention

[0003] Embodiments of this specification provide a method and device for generating instruction data, a computer-readable storage medium, an electronic device, and a computer program product, which can efficiently and conveniently generate high-quality instruction data for guiding a target large model to execute a target task to meet the requirement of accurately processing the target task using the target large model.

[0004] A method for generating instruction data provided by an embodiment of this specification includes:

[0005] Obtain task description information of a target task and first meta-prompt information; wherein, the first meta-prompt information is used to guide a first large language model to iteratively optimize the initial instruction data generated based on the task description information to obtain candidate instruction data for guiding the target large model to execute the target task;

[0006] Input the task description information and the first meta-prompt information into the first large language model to obtain the candidate instruction data generated by the first large language model;

[0007] Perform a fusion process on the candidate instruction data and task example information of the target task to obtain target instruction data for guiding the target large model to execute the target task.

[0008] The embodiments of this specification also provide an instruction data generation device, including:

[0009] An acquisition module, configured to acquire task description information of a target task and first meta-prompt information; wherein, the first meta-prompt information is used to guide a first large language model to iteratively optimize the initial instruction data generated by it based on the task description information, so as to obtain candidate instruction data for guiding a target large model to execute the target task;

[0010] A candidate instruction generation module, configured to input the task description information and the first meta-prompt information into the first large language model to obtain the candidate instruction data generated by the first large language model;

[0011] A target instruction generation module, configured to perform a fusion process on the candidate instruction data and the task example information of the target task to obtain target instruction data for guiding the target large model to execute the target task.

[0012] The embodiments of this specification also provide a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0013] The embodiments of this specification also provide an electronic device, including: a processor and a memory; wherein, the memory stores a computer program, and the computer program is adapted to be loaded and executed by the processor to implement the steps of the above method.

[0014] The embodiments of this specification also provide a computer program product, on which at least one instruction is stored, and when the at least one instruction is executed by a processor, the steps of the above method are implemented.

[0015] In the embodiments of this specification, after acquiring the task description information of the target task and the first meta-prompt information, the task description information and the first meta-prompt information can be input into the first large language model, so that the first large language model can iteratively optimize the initial instruction data generated by it based on the task description information under the guidance of the first meta-prompt information, so as to efficiently and conveniently generate candidate instruction data with better quality and capable of guiding the target large model to execute the target task. By performing a fusion process on the candidate instruction data and the task example information of the target task, the quality of the obtained target instruction data for guiding the target large model to execute the target task can be further improved to meet the requirement of accurately processing the target task by using the target large model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of an application scenario of an instruction data generation solution provided by the embodiments of this specification;

[0017] Figure 2 A flowchart of an instruction data generation method provided by an embodiment of this specification;

[0018] Figure 3 A schematic diagram of the principle of an instruction data generation provided by an embodiment of this specification;

[0019] Figure 4 A schematic diagram of the structure of an instruction data generation device provided by an embodiment of this specification;

[0020] Figure 5 A schematic diagram of the structure of an electronic device provided by an embodiment of this specification. Detailed implementation manners

[0021] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this specification.

[0022] Before specifically describing the instruction data generation solution provided by the embodiments of this specification, the relevant technical background will be described first.

[0023] With the continuous progress of large model technology, the demand for using large models in different scenarios to improve production efficiency has become even more urgent. Currently, people usually need to rely on the knowledge and experience of domain experts to manually write instruction data applicable to the target task, or the domain experts pre-define an instruction data template, and obtain the instruction data applicable to the target task by adjusting the instruction data model according to different task requirements. Subsequently, by using the above instruction data applicable to the target task to fine-tune the large model, the large model can learn how to generate outputs applicable to the target task according to user instructions, thereby improving the performance of the large model in processing the target task to better complete various downstream tasks.

[0024] However, when manually writing the instruction data required for fine-tuning a large model based on experts, not only does it consume a large amount of human resources, but also, as the task requirements continue to change, the instruction data written by experts needs to be continuously updated. As a result, there is a situation where the speed of manually writing and updating the instruction data cannot meet the requirements of large model fine-tuning. When generating the instruction data required for large model fine-tuning based on predefined instruction data templates, due to the limited number of instruction data templates, various tasks are often not covered, thus affecting the applicability of the instruction data generated based on the instruction data templates to novel or complex tasks. In addition, since the instruction data templates also need to be extended and maintained as the task requirements change, this also consumes a large amount of manpower and time, making it difficult to meet the usage requirements of a large amount of high-quality instruction data during the large model fine-tuning process.

[0025] Based on this, how to efficiently and conveniently generate high-quality instruction data for guiding a large model to execute a target task has become a technical problem that urgently needs to be solved.

[0026] Please refer to Figure 1 , which is a schematic diagram of an application scenario of an instruction data generation scheme provided by an embodiment of this specification.

[0027] As Figure 1 shown, when it is necessary to use the electronic device 101 to generate instruction data for guiding a target large model to execute a target task, the file data 102 containing the task description information of the target task, the first meta-prompt information, and the task example information of the target task can be obtained by using the electronic device 101 first. The task description information and the first meta-prompt information are input into the first large language model 103, so that the first large language model 103 can iteratively optimize the initial instruction data generated based on the task description information under the guidance of the first meta-prompt information, thereby obtaining candidate instruction data for guiding the target large model to execute the target task. Subsequently, by performing a fusion process on the candidate instruction data and the task example information of the target task, high-quality target instruction data 104 that can be used to guide the target large model to execute the target task can be obtained efficiently and conveniently.

[0028] Please refer to Figure 2 , which is a schematic flowchart of an instruction data generation method provided by an embodiment of this specification. The execution subject of this process can be an application program for generating instruction data, or an electronic device equipped with the above application program. The following will elaborate in detail on the Figure 2 shown process. The instruction data generation method can specifically include the following steps:

[0029] Step 202: Obtain the task description information of the target task and the first meta-prompt information; wherein, the first meta-prompt information is used to guide the first large language model to iteratively optimize the initial instruction data generated based on the task description information, so as to obtain candidate instruction data for guiding the target large model to execute the target task.

[0030] In the embodiments of this specification, the target task may refer to Figure 2 the downstream task that the instruction data required by the method in [reference] is used to guide the target large model to execute. In practical applications, the types of the target tasks can be set according to actual needs. For example, in the intelligent customer scenario, the target tasks may include answering various user questions; in the auxiliary teaching scenario, the target tasks may include solving various exercises; in the translation scenario, the target tasks may include translating the information input by the user into the information in a specified language, etc., and no specific limitation is made thereto. Among them, there can also be multiple types of the target large models. For example, it may include large language models (LLMs), computer vision large models (CVLMs), multi-modal large models, etc., and no specific limitation is made thereto.

[0031] In the embodiments of this specification, the task description information of the target task can generally be information for introducing, describing the task content, task form, and task objective of the target task. In practical applications, the format of the task description information of the target task can be various. For example, the task description information of the target task can be text information constructed using natural language, or it can also be voice information, image information, video information, etc., and no specific limitation is made thereto.

[0032] For the sake of easy understanding, an example of the task description information is given here. Assume that when the target task includes solving various math exercises, the task description information can be used to describe the scope of the math exercises to be solved, the solution ideas, the teacher's lesson preparation content, etc.; and when the target task includes answering user questions at the payment application, the task description information may include the question-answering rules preset by the service provider of the payment application, etc.; and no specific limitation is made thereto.

[0033] In the embodiments of this specification, Meta Prompt can be a technology designed to help large language models more effectively process and solve various cognitive tasks. Using Meta Prompt can generate prompts at the large language model in a structured manner, so as to continuously construct and optimize the prompt information at the large language model on the basis of minimizing changes to the existing prompts, guiding the large language model to better understand the input data of the large language model according to preset rules, clarifying the main objectives and requirements of the task, and moreover, ensuring that a more suitable output format and clear and specific language expression are used to generate the output data of the large language model. This technology of Meta Prompt can not only reduce the dependence of the large language model on a large number of task examples during operation, but also enhance the autonomy and adaptability of the large language model, enabling it to play a role in a wider range of tasks and fields.

[0034] Based on this, when it is necessary to use the first large language model to generate instruction data for guiding the target large model to execute the target task, in order to enable the first large language model to better understand the task description information of the target task and be able to generate high-quality instruction data suitable for the target task, the first Meta Prompt information can be obtained, and the first Meta Prompt information can be made to guide the first large language model to iteratively optimize the initial instruction data generated based on the task description information of the target task, so as to obtain high-quality candidate instruction data that can be used to guide the target large model to execute the target task.

[0035] Among them, the first Meta Prompt information can generally provide the understanding logic for the task description information of the target task, as well as the generation logic and iterative optimization logic for the candidate instruction data used to guide the target large model to execute the target task. In practical applications, the first Meta Prompt information can be either text information constructed using natural language, or it can also be voice information, image information, video information, or intelligent algorithm models, etc., and no specific limitation is made thereto.

[0036] Step 204, input the task description information and the first Meta Prompt information into the first large language model to obtain the candidate instruction data generated by the first large language model.

[0037] In the embodiments of this specification, the task description information and the first Meta Prompt information can be input into the input layer of the first large language model, so that the first large language model can perform feature extraction and analysis and understanding processing on the received task description information and the first Meta Prompt information, and thus be able to generate initial instruction data based on the task description information of the target task under the guidance of the first Meta Prompt information, and iteratively optimize the initial instruction data, so as to obtain high-quality candidate instruction data that can be used to guide the target large model to execute the target task.

[0038] For ease of understanding, examples of the candidate instruction data are given here. Assume that when the target task includes solving various math problems, the candidate instruction data can be "I want to improve my math grades. Please solve the following problems step by step for me", "Please deeply understand the following math problems and provide multiple solution ideas", etc.; and when the target task includes answering user questions at the payment application, the candidate instruction data can include "Please evaluate the possible risks in the user payment process and list the high-risk matters", "Please collect the currently available payment preferential rules and screen and analyze the payment preferential rules with higher user interest", etc. No specific limitations are made on this either.

[0039] Step 206: Perform a fusion process on the candidate instruction data and the task example information of the target task to obtain target instruction data for guiding the target large model to execute the target task.

[0040] In the embodiments of this specification, since when fine-tuning the large model, if the instruction data input to the large model contains high-quality task example information, it can often significantly improve the large model's ability to understand the requirements, rules, and expected output content of the task. Therefore, the task example information of the target task can be obtained in advance, and by performing a fusion process on the task example information of the target task and the above candidate instruction data, target instruction data containing one or more pieces of the task example information can be obtained, which is beneficial to improving the effect when using the target instruction data to fine-tune the target large model, so as to improve the target large model's ability to process the target task.

[0041] In practical applications, the task example information of the target task usually includes at least the input data (Input Data) that the target large model needs to process and the specific data content (Output Data) that needs to be output; in addition, the task example information of the target task can also include information for informing the type or format (Output Indicator) of the data that the target large model needs to output, and no specific limitation is made thereto. For the convenience of understanding, the task example information of the target task is exemplified herein. Suppose that when the target task includes solving various mathematical problems, the task example information can be "Input information: What is the result of 1 + 1, Output information: 2", "Input information: Is it normal that the sum of 1 and 1 is 3, Output information: Wrong", etc.; and when the target task includes answering user questions at the payment application, the task example information can include "The user wants to query whether a voucher can be received now. According to the current voucher receiving policy, this user does not meet the voucher receiving conditions, so the voucher cannot be received currently", "Input information: How long can the payment be received after the payment is successful, Output information: The other party's payment account can usually receive your payment within 5 minutes", etc.; and no specific limitation is made thereto.

[0042] In practical applications, there are often various ways to perform the fusion processing on the candidate instruction data and the task example information of the target task. For example, the candidate instruction data and the task example information of the target task can be simply concatenated to obtain the target instruction data, which is beneficial to simplifying the generation process of the target instruction data; or, the large language model can also be used to analyze, fuse, and optimize the candidate instruction data and the task example information of the target task to obtain the target instruction data, so as to improve the quality of the target instruction data, and no specific limitation is made thereto.

[0043] Figure 2 In the method in [reference], by enabling the first large language model to iteratively optimize the initial instruction data generated based on the task description information of the target task under the guidance of the first meta-prompt information, it can more accurately understand the target task and the requirements of the instruction data to be generated, so as to efficiently and conveniently generate candidate instruction data with good quality and capable of guiding the target large model to execute the target task, which is beneficial to saving human resources; and, by performing the fusion processing on the candidate instruction data and the task example information of the target task, different task knowledge can be conveniently fused in the instruction data, which is beneficial to improving the content richness and task adaptability of the generated target instruction data, so as to better meet the needs of accurately processing various target tasks by using the target large model. And since Figure 2The solution in [reference] can adapt to diverse task requirements, has strong personalization capabilities, and is thus easy to expand to different fields and tasks, with good universality and practicality.

[0044] Based on Figure 2 the method in [reference], the embodiments of this specification also provide some specific implementation schemes of this method, which will be described below.

[0045] In the embodiments of this specification, Figure 2 the method described in [reference] may further include:

[0046] Obtain the preset task template information corresponding to the target task; wherein, the preset task template information is used to guide the first large language model to parse the task description information, and to guide the first large language model to generate instruction data that can guide the target large model to execute the target task according to the parsing result of the task description information;

[0047] Correspondingly, step 204, inputting the task description information and the first meta-prompt information into the first large language model to obtain the candidate instruction data generated by the first large language model may include:

[0048] Input the task description information, the first meta-prompt information, and the preset task template information into the first large language model to obtain the candidate instruction data generated by the first large language model.

[0049] In the embodiments of this specification, in order to further improve the capabilities of the first large language model to deeply understand the task description information of the target task and clarify the format and content of the required output instruction data, etc., the preset task template information corresponding to the target task may also be obtained in advance, and the preset task template information, the task description information, and the above-mentioned first meta-prompt information are input into the input end (for example, Prompt Encoder) of the first large language model, so that the first large language model can combine the preset task template information and the first meta-prompt information to more deeply parse the task description information of the target task, so as to clarify the format and content of the instruction data required to be generated for guiding the target large model to execute the target task according to the parsing result of the task description information, which is beneficial to improving the accuracy and task adaptability of the candidate instruction data output by the output end of the first large language model.

[0050] In the embodiments of this specification, there can be various types of the preset task template information. For example, the preset task template information may include at least one of a preset task template text constructed using natural language, a mind map of the target task, and a task flow chart of the target task.

[0051] Among them, the preset task template text constructed using natural language can be text information with good generality set in advance. The preset task template text can generally be used to guide the large language model to extract key information from the task description information, and clarify the task content, task purpose in combination with the task description information. Moreover, it can also be used to guide the format, content logic, etc. of the instruction data required by the large language model. There is no specific limitation on this.

[0052] A mind map is an effective graphical thinking tool for expressing divergent thinking. It can start from a central theme and radiate multiple branch themes in all directions. Keywords or short sentences can be marked on each branch theme to concisely express the core information. A mind map can not only improve the intuitiveness and vividness of information, but also help people better understand and remember relevant content. Based on this, a mind map of the target task can be drawn manually in advance according to the task description information of the target task and other relevant materials. After receiving the mind map of the target task, the first large language model can use image analysis and processing technology to obtain the target task-related information carried in the mind map to assist the first large language model in deeply and correctly understanding the target task.

[0053] A task flow chart (i.e., a work flow chart) can visually present the steps involved in the task flow, and describe relevant information such as the content, occurrence order, and execution conditions of these steps. The task flow chart can clearly and briefly explain the content of different stages of the task flow, as well as the interrelationships and flow relationships between each task stage from start to end, so as to facilitate people to master the overall task flow. Based on this, a task flow chart of the target task can be drawn manually in advance according to the task description information of the target task and other relevant materials. After receiving the task flow chart of the target task, the first large language model can use image analysis and processing technology to obtain the target task-related information carried in the task flow chart to assist the first large language model in deeply and correctly understanding the target task.

[0054] Combined with the above content, since the preset task template information can be either text data or image data, in order to enable the first large language model to generate candidate instruction data in combination with various formats of preset task template information, the first large language model can be implemented using a pre-trained multi-modal large language model capable of processing text and images. Of course, if the formats of the task description information, the first meta-prompt information, and the preset task template information are the same, the first large language model can also be implemented using a pre-trained single-modal large language model, which has good flexibility.

[0055] In the embodiments of this specification, the first meta-prompt information may include an iterative optimization strategy for the initial instruction data. Correspondingly, inputting the task description information, the first meta-prompt information, and the preset task template information into the first large language model to obtain the candidate instruction data generated by the first large language model may include:

[0056] Using the first large language model to analyze and process the received task description information, the first meta-prompt information, and the preset task template information to obtain initial instruction data for guiding the target large model to execute the target task.

[0057] Using the first large language model to perform iterative optimization processing on the initial instruction data according to the iterative optimization strategy when it is determined that the preset instruction iterative optimization condition is satisfied based on the iterative optimization strategy to obtain the candidate instruction data.

[0058] In the embodiments of this specification, after analyzing and understanding the task description information, the preset task template information, and the first meta-prompt information of the target task, the first large language model can initially generate initial instruction data for guiding the target large model to execute the target task. However, since the first large language model may not be able to capture sufficient detailed information about the target task during a single instruction generation process and also lacks a process for evaluating and optimizing the quality of the initial instruction data, it is easy for the initial instruction data to have problems with insufficient accuracy, and thus it cannot meet the high-quality requirements for instruction data in complex or variable tasks.

[0059] Based on this, the first meta-prompt information may be made to include an iterative optimization strategy for the initial instruction data. When the iterative optimization strategy is executed, it can generally be divided into two stages: strategy evaluation and strategy update. Among them, in the strategy evaluation stage, the quality and performance of the instruction data generated by the first large language model in this iteration process can be evaluated to determine whether the preset instruction iterative optimization condition is satisfied in combination with the quality and performance of the instruction data. If the preset instruction iterative optimization condition is satisfied, the strategy update stage can be entered. In the strategy update stage, the quality evaluation result of the optimized instruction data can be used as the update target, and self-querying can be continuously performed in accordance with the preset instruction update strategy by combining the negative feedback information generated in the previous iteration process to ensure that the first large language model deeply understands each input data it receives until the preset instruction iterative optimization condition is no longer satisfied, thereby finally generating optimized candidate instruction data, which is beneficial to improving the quality of the candidate instruction data.

[0060] In practical applications, there can be various types of the preset instruction iteration optimization conditions, which can be set according to actual requirements. For example, the quality score of the instruction data generated in this iteration process does not reach the threshold, the loss function value corresponding to the instruction data generated in this iteration process is greater than the threshold, the number of iteration rounds does not reach the threshold, the iteration duration does not reach the threshold, etc. No specific limitation is made in this regard.

[0061] In the embodiments of this specification, the fusion processing of the candidate instruction data and the task example information of the target task to obtain the target instruction data for guiding the target large model to execute the target task may include:

[0062] Using a second large language model to perform fusion processing on the candidate instruction data and the task example information of the target task to obtain the target instruction data carrying the task example information.

[0063] In the embodiments of this specification, since a large language model can generally have a certain ability to understand human natural language and can automatically generate coherent and logically reasonable text content according to input context information, such as articles, stories, conversations, etc., it can effectively improve the content generation efficiency. Based on this, a second large language model can be used to perform fusion processing on the candidate instruction data and the task example information of the target task to obtain target instruction data carrying the task example information with coherent word order, reasonable logic, and appropriate format, thereby improving the generation efficiency and quality of the target instruction data.

[0064] In practical applications, in order to effectively ensure the quality of the target instruction data generated by the second large language model, the second large language model can be pre-finely tuned using candidate instruction data samples, corresponding task example information samples, and high-quality instruction data samples after the fusion of the two, so that the second large language model learns the fusion processing knowledge of the candidate instruction data and the task example information of the target task, and learns the feature knowledge of the target specified data it needs to output.

[0065] Of course, it is also possible to not pre-finely tune the second large language model, but instead input the prompt information that can be used to guide the second large language model to generate the target instruction data into the second large language model, which is convenient, fast, and has good effectiveness. Based on this, Figure 2 The method in

[0066] obtain second meta-prompt information; wherein, the second meta-prompt information is used to guide the second large language model to perform fusion processing on the candidate instruction data and the task example information to obtain the target instruction data conforming to the first preset data structure.

[0067] Correspondingly, the step of using the second large language model to perform fusion processing on the candidate instruction data and the task example information of the target task to obtain the target instruction data carrying the task example information may include:

[0068] Input the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data generated by the second large language model and conforming to the first preset data structure.

[0069] In the embodiments of the present specification, the first meta-prompt information may include a fusion strategy for the candidate instruction data and the task example information, and feature information reflecting the characteristics of the first preset data structure that the target instruction data required to be output by the second large language model should conform to. Among them, the fusion strategy may reflect the example content that needs to be identified and extracted from the task example information, and the number of example contents that need to be added to the candidate instruction data, etc. The first preset data structure may be the word order feature and logical architecture of the instruction data, and may also include the data format of the target instruction data, etc., and no specific limitation is made thereto.

[0070] By inputting the candidate instruction data, the task example information, and the second meta-prompt information into the input end of the second large language model, the second large language model can parse and fuse the candidate instruction data and the task example information under the guidance of the second meta-prompt information, and output the target instruction data conforming to the first preset data structure from the output end, which is beneficial to ensuring the convenience and efficiency of generating the target instruction data, and is also beneficial to improving the quality of the target instruction data.

[0071] In the embodiments of the present specification, the second meta-prompt information is specifically used to guide the second large language model to screen out high-quality instruction data from the candidate instruction data according to the parsing result of the task description information, and to obtain the target instruction data conforming to the first preset data structure by performing fusion processing on the high-quality instruction data and the task example information.

[0072] Correspondingly, the step of inputting the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data generated by the second large language model and conforming to the first preset data structure may include:

[0073] Input the candidate instruction data, the task example information, the second meta-prompt information, and the parsing result of the task description information into the second large language model to obtain the target instruction data generated by the second large language model and conforming to the first preset data structure.

[0074] In the embodiments of this specification, since there are often multiple pieces of candidate instruction data output by the first large language model, and the quality of different candidate instruction data is often uneven, based on this, the second meta-prompt information can also be used to guide the second large language model to screen the candidate instruction data. Thus, by fusing and optimizing the screened high-quality instruction data with the task example information of the target task, it is beneficial to further improve the quality of the target instruction data output by it.

[0075] In practical applications, in order to enable the second large language model to have the ability to evaluate or optimize the quality of instruction data, the parsing result obtained by analyzing and understanding the task description information of the target task by combining the first large language model with the first meta-prompt information and the preset task template information can also be input into the second large language model. Thus, under the guidance of the second meta-prompt information, the second large language model can determine the relevance or semantic consistency between each candidate instruction data and the parsing result of the above task description information, so that the candidate instruction data with stronger relevance or more semantic consistency can be used as high-quality instruction data. Among them, the number of the high-quality instruction data is generally less than or equal to the candidate instruction data, and no specific limitation is made on this.

[0076] In practical applications, in order to avoid the excessive data volume of the target instruction data from affecting the efficiency when using it to fine-tune the target large model, and also to ensure the accuracy and effectiveness of the target instruction data, based on this, the second large language model can be made to integrate a preset number of task example information into a high-quality instruction data to obtain the target instruction data under the guidance of the second meta-prompt information; or, the second large language model can also be made to continue to optimize the high-quality instruction data integrated with a preset number of task example information to obtain the target instruction data under the guidance of the second meta-prompt information; or, the second large language model can also be made to optimize the task example information of the target task received by it under the guidance of the second meta-prompt information, so as to integrate the optimized task example information into the high-quality instruction data to generate the target instruction data, with good flexibility, and no specific limitation is made on this either.

[0077] In practical applications, the second large language model and the first large language model described above can be the same model, so as to reduce the resource occupancy at the electronic device by reusing the large language model. Or, the second large language model and the first large language model described above can also be different models. Moreover, the output end of the first large language model can be directly connected to the input end of the second large language model to input the candidate instruction data into the second large language model. In addition, the output end of the Prompt Encoder of the first large language model is also connected to the input end of the second large language model, so as to input the parsing result of the task description information into the second large language model, which is beneficial to improving the operation convenience of the generation process of the target instruction data, and no specific limitation is made thereto.

[0078] In the embodiments of this specification, Figure 2 the method described in may further include:

[0079] Perform key information extraction processing on the task example information of the target task to obtain the task input information and task output information of the target task.

[0080] Generate a task example that conforms to the second preset data structure according to the task input information and the task output information.

[0081] Correspondingly, the step of inputting the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data that conforms to the first preset data structure generated by the second large language model may include:

[0082] Input the candidate instruction data, the second meta-prompt information, and the task example that conforms to the second preset data structure into the second large language model to obtain the target instruction data that conforms to the first preset data structure generated by the second large language model.

[0083] In the embodiments of this specification, since the structured task example is not only beneficial to the second large language model to understand the task example, but also beneficial to improving the quality of the target instruction data carrying the task example finally generated by the second large language model. Therefore, the task example information of the target task can be structured first, so as to input the task example that conforms to the second preset data structure into the second large language model.

[0084] In practical applications, the second preset data structure can be set according to actual needs. For example, the second preset data structure can indicate the relative setting positions, data formats, etc. of the task input information and task output information included in the task example, and no specific limitation is imposed thereon. Based on this, it is necessary to first extract the task input information and task output information of the target task from the task example information of the target task, so as to generate a task example carrying the task input information and the task output information and conforming to the second preset data structure according to the requirements of the second preset data structure.

[0085] In the embodiments of this specification, Figure 2 the method described above may further include:

[0086] Using a third large language model to generate a task example of the target task conforming to a third preset data structure according to the task example information of the target task; wherein, the task example includes the task input information and task output information in the task example information.

[0087] Correspondingly, the step of inputting the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data generated by the second large language model and conforming to the first preset data structure may include:

[0088] Inputting the candidate instruction data, the second meta-prompt information, and the task example conforming to the third preset data structure into the second large language model to obtain the target instruction data generated by the second large language model and conforming to the first preset data structure.

[0089] In the embodiments of this specification, the content generation ability of the large language model can also be used to input the task example information of the target task and the prompt information / meta-prompt information for reflecting the third preset data structure into the third large language model, so as to efficiently and conveniently generate a task example of the target task conforming to the third preset data structure by using the third large language model, complete the structural conversion of the task example information of the target task, and is beneficial to improving the quality of the target instruction data generated by the second large language model based on the task example of the third preset data structure. Among them, the third preset data structure and the above-mentioned second preset data structure are usually consistent, and of course, there may also be differences, which will not be elaborated here.

[0090] For the sake of easy understanding, Figure 3 FIG. is a schematic diagram of the principle of generating instruction data provided by the embodiments of this specification. As Figure 3As shown, the task description information of the target task, the preset task template information, and the corresponding first meta-prompt information can be input into the first large language model. The first large language model can parse the task description information of the target task under the guidance of the preset task template information and the corresponding first meta-prompt information to obtain the parsing result of the task description information. In addition, the first large language model can also, under the guidance of the first meta-prompt information, continuously iterate and optimize the instruction data it generates to guide the target large model to execute the target task based on the parsing result of the task description information, so as to output candidate instruction data.

[0091] Subsequently, on the one hand, the candidate instruction data, the parsing result of the task description information, and the corresponding second meta-prompt information can be input into the second large language model; on the other hand, the structured task example information that conforms to the second / third preset data structure obtained by structuring the task example information for the target task can also be input into the second large language model, so that the second large language model can, under the guidance of the second meta-prompt information, combine the parsing result of the task description information to fuse and optimize the structured task example information and the candidate instruction data, and thus output the target instruction data that conforms to the first preset data structure.

[0092] In the embodiments of this specification, the fusion process of the candidate instruction data and the task example information of the target task to obtain the target instruction data for guiding the target large model to execute the target task may include:

[0093] According to the fourth preset data structure, add the task example information of the target task to the specified position in the candidate instruction data to obtain the target instruction data that conforms to the fourth preset data structure.

[0094] In the embodiments of this specification, it is also possible not to use the large language model to fuse the candidate instruction data and the task example information of the target task, but directly preset the fourth preset data structure that the target instruction data needs to conform to, and then, according to the preset strategy, add the task example information of the target task to the specified position in the candidate instruction data, so as to efficiently and conveniently obtain the target instruction data that conforms to the fourth preset data structure. Among them, the fourth preset data structure and the above-mentioned first preset data structure are usually consistent, and of course, there may also be differences, which will not be elaborated here.

[0095] Based on the same idea, the embodiments of this specification also provide a device corresponding to the above method. Please refer to Figure 4 , which is a schematic structural diagram of an instruction data generation device provided by the embodiments of this specification. As Figure 4As shown, the instruction data generation device 4 can be implemented as all or part of an electronic device through software, hardware, or a combination of both. According to some embodiments, the instruction data generation device 4 may include an acquisition module 41, a candidate instruction generation module 42, and a target instruction generation module 43, where:

[0096] The acquisition module 41 is configured to acquire task description information of a target task and first meta-prompt information; wherein, the first meta-prompt information is used to guide a first large language model to iteratively optimize initial instruction data generated based on the task description information, so as to obtain candidate instruction data for guiding a target large model to execute the target task.

[0097] The candidate instruction generation module 42 is configured to input the task description information and the first meta-prompt information into the first large language model to obtain the candidate instruction data generated by the first large language model.

[0098] The target instruction generation module 43 is configured to perform a fusion process on the candidate instruction data and task example information of the target task to obtain target instruction data for guiding the target large model to execute the target task.

[0099] Optionally, Figure 4 the device in may further include:

[0100] A task template acquisition module, configured to acquire preset task template information corresponding to the target task; wherein, the preset task template information is used to guide the first large language model to parse the task description information, and to guide the first large language model to generate instruction data capable of guiding the target large model to execute the target task according to the parsing result of the task description information.

[0101] The candidate instruction generation module 42 may specifically be configured to:

[0102] Input the task description information, the first meta-prompt information, and the preset task template information into the first large language model to obtain the candidate instruction data generated by the first large language model.

[0103] Optionally, the preset task template information may include at least one of a preset task template text constructed in natural language, a mind map of the target task, and a task flow chart of the target task.

[0104] Optionally, the first meta-prompt information may include an iterative optimization strategy for the initial instruction data.

[0105] The candidate instruction generation module 42 may include:

[0106] The first processing unit is configured to analyze and process the received task description information, the first meta-prompt information, and the preset task template information by using the first large language model, so as to obtain initial instruction data for guiding the target large model to execute the target task.

[0107] The second processing unit is configured to, when it is determined that the preset instruction iteration optimization condition is satisfied based on the iteration optimization strategy by using the first large language model, perform iterative optimization processing on the initial instruction data according to the iteration optimization strategy to obtain the candidate instruction data.

[0108] Optionally, the target instruction generation module 43 may be configured to:

[0109] Fuse the candidate instruction data with the task example information of the target task by using the second large language model to obtain the target instruction data carrying the task example information.

[0110] Optionally, Figure 4 the device in

[0111] The meta-prompt acquisition module is configured to acquire second meta-prompt information; wherein, the second meta-prompt information is used to guide the second large language model to fuse the candidate instruction data with the task example information to obtain the target instruction data conforming to the first preset data structure.

[0112] The target instruction generation module 43 may specifically be configured to:

[0113] Input the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data generated by the second large language model and conforming to the first preset data structure.

[0114] Optionally, the second meta-prompt information may specifically be used to guide the second large language model to screen out high-quality instruction data from the candidate instruction data according to the parsing result of the task description information, and, by fusing the high-quality instruction data with the task example information, to obtain the target instruction data conforming to the first preset data structure.

[0115] The target instruction generation module 43 may specifically be configured to:

[0116] Input the candidate instruction data, the task example information, the second meta-prompt information, and the parsing result of the task description information into the second large language model to obtain the target instruction data generated by the second large language model and conforming to the first preset data structure.

[0117] Optionally, Figure 4 the device in Figure 4 may further include:

[0118] a key information extraction module, configured to perform key information extraction processing on the task example information of the target task to obtain the task input information and task output information of the target task.

[0119] a first task example generation module, configured to generate a task example that conforms to a second preset data structure according to the task input information and the task output information.

[0120] The target instruction generation module 43 may specifically be configured to:

[0121] input the candidate instruction data, the second meta-prompt information, and the task example that conforms to the second preset data structure into the second large language model to obtain the target instruction data that conforms to the first preset data structure generated by the second large language model.

[0122] Optionally, Figure 4 the device in Figure 4 may further include:

[0123] a second task example generation module, configured to use a third large language model to generate a task example that conforms to a third preset data structure for the target task according to the task example information of the target task; wherein, the task example includes the task input information and task output information in the task example information.

[0124] The target instruction generation module 43 may specifically be configured to:

[0125] input the candidate instruction data, the second meta-prompt information, and the task example that conforms to the third preset data structure into the second large language model to obtain the target instruction data that conforms to the first preset data structure generated by the second large language model.

[0126] Optionally, the target instruction generation module 43 may specifically be configured to:

[0127] add the task example information of the target task to a specified position in the candidate instruction data according to a fourth preset data structure to obtain the target instruction data that conforms to the fourth preset data structure.

[0128] The above device embodiments correspond to the method embodiments. For specific descriptions, reference may be made to the descriptions in the method embodiment section, which will not be elaborated here. The device embodiments are obtained based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments. For specific descriptions, reference may be made to the corresponding method embodiments.

[0129] The embodiments of this specification also provide a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the instruction data generation method as described above. Figure 2 For the specific execution process, reference can be made to the detailed description of the relevant embodiments of the instruction data generation method, which will not be elaborated here.

[0130] This specification also provides a computer program product, which stores at least one instruction. The at least one instruction is loaded and executed by the processor to implement the instruction data generation method as described above. Figure 2 For the specific execution process, reference can be made to the detailed description of the relevant embodiments of the instruction data generation method, which will not be elaborated here.

[0131] The embodiments of this specification also provide Figure 5 a schematic structural diagram of the electronic device as shown. As Figure 5 shown, at the hardware level, the electronic device may include a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the instruction data generation method as described above. Figure 5 For the specific execution process, reference can be made to the detailed description of the relevant embodiments of the instruction data generation method, which will not be elaborated here.

[0132] Of course, in addition to the software implementation, this specification does not exclude other implementation manners, such as logic devices or a combination of software and hardware. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but may also be hardware or a logic device.

[0133] The embodiments in this specification are all described in a progressive manner. For the parts that are the same or similar among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the computer-readable storage medium, the computer program product, and Figure 5 the electronic device as shown, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the method embodiments.

[0134] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to circuit structures such as diodes, transistors, switches, etc.) or software improvements (improvements to method flows). However, with the development of technology, many method flow improvements today can be regarded as direct improvements to hardware circuit structures. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that an improvement to a method flow cannot be implemented using a hardware entity module. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is an integrated circuit whose logic function is determined by the user programming the device. Designers can program themselves to "integrate" a digital system onto a single PLD, without having to ask a chip manufacturer to design and fabricate a dedicated integrated circuit chip. Moreover, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly implemented using "logic compiler" software, which is similar to the software compilers used in program development and writing. The original code before compilation also has to be written in a specific programming language, which is called a Hardware Description Language (HDL). There is not just one type of HDL, but many, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones currently are VHDL (Very-High-Speed Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also be aware that by simply performing a little logical programming on the method flow using the above-mentioned several hardware description languages and programming it into an integrated circuit, it is easy to obtain the hardware circuit that implements the logical method flow.

[0135] The controller can be implemented in any suitable manner. For example, the controller can take the form of, for example, a microprocessor or a processor and a computer-readable medium storing computer-readable program code (such as software or firmware) executable by the (micro)processor, logic gates, switches, an application specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of the controller include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art also know that, in addition to implementing the controller in the form of pure computer-readable program code, it is entirely possible to logically program the method steps to enable the controller to be implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. to achieve the same functions. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be regarded as the structures within the hardware component. Or even, the devices for implementing various functions can be regarded as either software modules for implementing the method or structures within the hardware component.

[0136] The systems, devices, modules, or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0137] For the convenience of description, the above devices are described by dividing them into various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0138] Those skilled in the art should understand that the embodiments of this specification can be provided as a method, a system, or a computer program product. Therefore, this specification can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program code.

[0139] This specification is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the specification. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0140] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufacture including instruction means that implement the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0141] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0142] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0143] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.

[0144] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0145] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0146] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0147] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0148] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for the relevant parts, reference can be made to the partial description of the method embodiment.

[0149] The above description is only for the embodiments of this specification and is not intended to limit this specification. For those skilled in the art, various modifications and changes can be made to this specification. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this specification shall be included within the scope of the claims of this specification.

Claims

1. A method for generating instruction data, the method comprising: Obtaining task description information of a target task and first meta-prompt information; wherein, the first meta-prompt information is used to guide a first large language model to iteratively optimize initial instruction data generated based on the task description information, so as to obtain candidate instruction data for guiding a target large model to execute the target task; Inputting the task description information and the first meta-prompt information into the first large language model to obtain the candidate instruction data generated by the first large language model; Performing a fusion process on the candidate instruction data and task example information of the target task to obtain target instruction data for guiding the target large model to execute the target task.

2. The method according to claim 1, the method further comprising: Obtaining preset task template information corresponding to the target task; wherein, the preset task template information is used to guide the first large language model to parse the task description information, and to guide the first large language model to generate instruction data capable of guiding the target large model to execute the target task according to the parsing result of the task description information; The step of inputting the task description information and the first meta-prompt information into the first large language model to obtain the candidate instruction data generated by the first large language model includes: Inputting the task description information, the first meta-prompt information and the preset task template information into the first large language model to obtain the candidate instruction data generated by the first large language model.

3. The method according to claim 2, wherein the preset task template information includes: At least one of a preset task template text constructed using natural language, a mind map of the target task, and a task flow chart of the target task.

4. The method according to claim 2, wherein the first meta-prompt information includes an iterative optimization strategy for the initial instruction data; The step of inputting the task description information, the first meta-prompt information and the preset task template information into the first large language model to obtain the candidate instruction data generated by the first large language model includes: Using the first large language model to analyze and process the received task description information, the first meta-prompt information and the preset task template information to obtain initial instruction data for guiding the target large model to execute the target task; Using the first large language model to perform iterative optimization processing on the initial instruction data according to the iterative optimization strategy when it is determined that a preset instruction iterative optimization condition is satisfied based on the iterative optimization strategy, to obtain the candidate instruction data.

5. The method according to claim 2, the step of performing a fusion process on the candidate instruction data and task example information of the target task to obtain target instruction data for guiding the target large model to execute the target task includes: Using a second large language model to perform a fusion process on the candidate instruction data and task example information of the target task to obtain the target instruction data carrying the task example information.

6. The method according to claim 5, the method further comprising: Obtain the second meta-prompt information; wherein, the second meta-prompt information is used to guide the second large language model to perform a fusion process on the candidate instruction data and the task example information to obtain the target instruction data that conforms to the first preset data structure; The step of using the second large language model to perform a fusion process on the candidate instruction data and the task example information of the target task to obtain the target instruction data carrying the task example information includes: Input the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data generated by the second large language model that conforms to the first preset data structure.

7. According to the method described in claim 6, the second meta-prompt information is specifically used to guide the second large language model to screen out high-quality instruction data from the candidate instruction data according to the parsing result of the task description information, and, by performing a fusion process on the high-quality instruction data and the task example information, to obtain the target instruction data that conforms to the first preset data structure; The step of inputting the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data generated by the second large language model that conforms to the first preset data structure includes: Input the candidate instruction data, the task example information, the second meta-prompt information, and the parsing result of the task description information into the second large language model to obtain the target instruction data generated by the second large language model that conforms to the first preset data structure.

8. According to the method described in claim 6, the method further includes: Perform key information extraction processing on the task example information of the target task to obtain the task input information and task output information of the target task; Generate a task example that conforms to the second preset data structure according to the task input information and the task output information; The step of inputting the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data generated by the second large language model that conforms to the first preset data structure includes: Input the candidate instruction data, the second meta-prompt information, and the task example that conforms to the second preset data structure into the second large language model to obtain the target instruction data generated by the second large language model that conforms to the first preset data structure.

9. According to the method described in claim 6, the method further includes: Use a third large language model to generate a task example of the target task that conforms to the third preset data structure according to the task example information of the target task; wherein, the task example contains the task input information and task output information in the task example information; Inputting the candidate instruction data, the task example information, and the second meta-prompt information into the second large language model to obtain the target instruction data that conforms to the first preset data structure and is generated by the second large language model includes: Inputting the candidate instruction data, the second meta-prompt information, and the task example that conforms to the third preset data structure into the second large language model to obtain the target instruction data that conforms to the first preset data structure and is generated by the second large language model.

10. The method according to claim 1, wherein the fusion processing of the candidate instruction data and the task example information of the target task to obtain the target instruction data for guiding the target large model to execute the target task includes: Adding the task example information of the target task to a specified position in the candidate instruction data according to the fourth preset data structure to obtain the target instruction data that conforms to the fourth preset data structure.

11. An instruction data generation device, comprising: An acquisition module, configured to acquire the task description information of the target task and the first meta-prompt information; wherein, the first meta-prompt information is used to guide the first large language model to iteratively optimize the initial instruction data generated based on the task description information to obtain the candidate instruction data for guiding the target large model to execute the target task; A candidate instruction generation module, configured to input the task description information and the first meta-prompt information into the first large language model to obtain the candidate instruction data generated by the first large language model; A target instruction generation module, configured to perform fusion processing on the candidate instruction data and the task example information of the target task to obtain the target instruction data for guiding the target large model to execute the target task.

12. A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

13. An electronic device, comprising: A processor and a memory; wherein, the memory stores a computer program, and the computer program is adapted to be loaded and executed by the processor to perform the steps of the method according to any one of claims 1 to 10.

14. A computer program product, on which at least one instruction is stored, and when the at least one instruction is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images