Information processing method, device and equipment

Abstract

The embodiment of the invention discloses an information processing method, device and equipment, and the method comprises the steps: obtaining a training data set used for training a target large model, the training data set comprises a plurality of data groups, and each data group comprises input data, output data and expected output data; based on the input data, the output data and the expected output data in each data group, generating first prompt information corresponding to each data group through a first large model; based on the first prompt information corresponding to each data group, generating second prompt information by selecting one or more pairs of prompt information and performing information cross processing on each pair of selected prompt information; and determining prompt information corresponding to the target large model based on the first prompt information and the second prompt information corresponding to each data group.

Description

Technical Field

[0001] This document relates to the field of computer technology, and in particular to an information processing method, apparatus, and device. Background Technology

[0002] With the continuous development of deep learning technology, large language models (LLMs) have become powerful tools for processing and understanding natural language, performing well in various application scenarios such as text generation, translation, privacy protection, and question answering systems. One of the keys to their success lies in the design of prompts (i.e., prompting), which guides the large model to generate the expected output data. However, manually designing effective prompts not only consumes a lot of manpower and time for trial and error, but also faces challenges: customized prompts need to be designed for different tasks, lacking generality; moreover, with the diversification of user needs and the rapid iteration of models, the sustainability of manually designed prompts is becoming increasingly low. Therefore, a better optimization mechanism for large model prompts is needed to effectively improve the optimization quality of prompts for specific tasks and enhance the performance of large models in handling complex tasks. Summary of the Invention

[0003] The purpose of the embodiments in this specification is to provide a better optimization mechanism for large model prompts, thereby effectively improving the optimization quality of prompts for specific tasks and enhancing the performance of large models in handling complex tasks.

[0004] To achieve the above technical solution, the embodiments in this specification are implemented as follows: This specification provides an information processing method, comprising: acquiring a training dataset for training a target large model, the training dataset including multiple data groups, each data group including input data, output data, and expected output data; generating a first prompt message corresponding to each data group using a first large model based on the input data, output data, and expected output data in each data group; generating a second prompt message based on the first prompt message corresponding to each data group by selecting one or more pairs of prompt messages and performing information cross-processing on each selected pair of prompt messages; and determining the prompt message corresponding to the target large model based on the first prompt message and the second prompt message corresponding to each data group.

[0005] This specification provides an information processing apparatus, comprising: a dataset acquisition module for acquiring a training dataset for training a target large model, the training dataset including multiple data groups, each data group including input data, output data, and expected output data; an information processing module for generating a first prompt message corresponding to each data group using a first large model based on the input data, output data, and expected output data in each data group; a cross-processing module for generating a second prompt message based on the first prompt message corresponding to each data group by selecting one or more pairs of prompt messages and performing cross-processing on each selected pair of prompt messages; and a prompt message determination module for determining the prompt message corresponding to the target large model based on the first prompt message and the second prompt message corresponding to each data group.

[0006] This specification provides an information processing device comprising: a processor; and a memory arranged to store computer-executable instructions, wherein the executable instructions, when executed, cause the processor to: acquire a training dataset for training a target large model, the training dataset comprising multiple data groups, each data group comprising input data, output data, and expected output data; generate a first prompt message corresponding to each data group using a first large model based on the input data, output data, and expected output data in each data group; generate a second prompt message based on the first prompt message corresponding to each data group by selecting one or more pairs of prompt messages and performing information cross-processing on each selected pair of prompt messages; and determine the prompt message corresponding to the target large model based on the first prompt message and the second prompt message corresponding to each data group.

[0007] This specification also provides a storage medium for storing computer-executable instructions. When executed by a processor, the executable instructions implement the following process: acquiring a training dataset for training a target large model, the training dataset including multiple data groups, each data group including input data, output data, and expected output data; generating a first prompt message corresponding to each data group using a first large model based on the input data, output data, and expected output data in each data group; generating a second prompt message based on the first prompt message corresponding to each data group by selecting one or more pairs of prompt messages and performing information cross-processing on each selected pair of prompt messages; and determining the prompt message corresponding to the target large model based on the first prompt message and the second prompt message corresponding to each data group.

[0008] This specification also provides a computer program product, including a computer program that, when executed by a processor, performs the following process: acquiring a training dataset for training a target large model, the training dataset including multiple data groups, each data group including input data, output data, and expected output data; generating a first prompt message corresponding to each data group based on the input data, output data, and expected output data in each data group, using a first large model; generating a second prompt message based on the first prompt message corresponding to each data group by selecting one or more pairs of prompt messages and performing information cross-processing on each selected pair of prompt messages; and determining the prompt message corresponding to the target large model based on the first prompt message and the second prompt message corresponding to each data group. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the structure of an information processing system described in this specification; Figure 2 This is a schematic diagram of an information processing procedure described in this specification; Figure 3 This is a schematic diagram of another information processing procedure described in this specification; Figure 4 This is a schematic diagram of yet another information processing procedure described in this specification; Figure 5 This is a schematic diagram of yet another information processing procedure described in this specification; Figure 6 This is a schematic diagram of yet another information processing procedure described in this specification; Figure 7 This is a schematic diagram of yet another information processing procedure described in this specification; Figure 8 This is a schematic diagram of an information processing device described in this specification; Figure 9 This is a schematic diagram of an information processing device described in this specification. Detailed Implementation

[0010] This specification provides an information processing method, apparatus, and device through its embodiments.

[0011] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0012] This specification provides an embodiment of a black-box large-scale model prompt information optimization mechanism based on genetic algorithms. Currently, prompt information optimization is mainly divided into continuous prompt optimization and discrete prompt optimization. Continuous prompt optimization typically transforms the optimization of prompt information into the optimization of a large model through soft prompts, offering a certain degree of flexibility. However, this approach has limitations: it relies on additional neural network training, resulting in high computational costs; moreover, it requires controllable model parameters (white-box mode), thus it cannot be applied in scenarios where model parameters are unknown or cannot be fine-tuned; furthermore, its adaptability to complex tasks is limited because soft prompts cannot fully capture task-specific reasoning logic. Discrete prompt optimization, on the other hand, mainly generates multiple versions of prompt information and selects the better prompt information for optimization. Although it can explore multiple possibilities, this approach lacks a feedback mechanism, the exploration process is too random and inefficient, and it cannot effectively balance exploration and development, easily getting trapped in local optima.

[0013] In the process of prompt optimization, when the model parameters in a large model are known, prompt optimization methods such as PrefixTuning and P-Tuning can be used. These methods obtain context-sensitive prompts by fine-tuning the large model, thereby improving the performance of the large model in specific tasks. However, these methods require fine-tuning the large model, which involves a large amount of computing resources, especially when dealing with ultra-large-scale models, resulting in high computational costs. Moreover, since each task requires separate optimization learning, the adaptability of the large model in multi-task scenarios is limited, leading to strong task dependence. In addition, the prompts obtained by fine-tuning the large model often perform poorly on unknown tasks or domains, resulting in poor generalization ability of the large model. Furthermore, these methods are only applicable to transparent models (i.e., models whose parameters are accessible) and cannot be applied to large models whose internal architecture cannot be modified (such as closed-source black-box large language models).

[0014] For black-box large-scale models, multiple versions of prompts can be generated and the optimal prompt selected. However, the generation and selection process of prompts in this way usually lacks an effective feedback mechanism, resulting in a high degree of randomness in the exploration process. It is impossible to fully utilize the results of previous interactions to guide the generation of future prompts. Moreover, this method is difficult to balance exploring new prompts with developing the potential of existing prompts, and it is easy to get stuck in local optima, that is, to remain in the existing effect for a long time without being able to break out. It is also inefficient when traversing multiple combinations of prompts, resulting in serious consumption of time and resources.

[0015] To address this, this specification proposes an optimization mechanism for large-scale model prompts based on genetic algorithms. Leveraging the randomness of genetic algorithms, a more balanced approach is achieved between existing and new prompts. Simultaneously, a feedback mechanism is used to specifically optimize the generated prompts. Since fine-tuning of the large-scale model's parameters is unnecessary, computational costs are significantly reduced, resulting in a balance between flexibility and efficiency. This effectively improves the quality of prompt optimization for specific tasks and enhances the performance of large-scale models in handling complex tasks. Specific details can be found in the following embodiments.

[0016] The information processing methods provided in one or more embodiments of this specification are applicable to the information processing implementation environment. (Refer to...) Figure 1 The implementation environment includes at least: Client 100 and server 200. Furthermore, server 200 can include various algorithms, large models, and specified network models (such as convolutional neural network models, BERT models, etc.), among which: Client 100 can run on terminal devices, which can be mobile phones, personal computers, tablets, e-book readers, wearable devices, devices that interact with information based on AR (Augmented Reality) and VR (Virtual Reality), and laptop computers, etc. Client 100 can be installed on terminal devices. Client 100 can be an application, a browser, or a subroutine embedded in an application, etc.

[0017] Server 200 can run on a server, which can be one or more servers, a server cluster consisting of several servers, or a cloud server on a cloud computing platform. Server 200 can be installed on the server. Server 200 can be an application or a subroutine embedded in an application. Various different algorithms, large models, and specified network models can be integrated into server 200. Alternatively, server 200 can call any one or more of various different algorithms, large models, and specified network models to perform corresponding operations.

[0018] In addition, it may include a database 300, which may be set in the server on which the server 200 runs, or it may be set outside the server on which the server 200 runs. The database 300 may store data in the training dataset, various prompts and other related information.

[0019] In this implementation environment, server 200 can collect training datasets for training the target large model through clients 100 of multiple different users, or server 200 can collect training datasets for training the target large model from data stored in a local database. The training dataset includes multiple data groups, each containing input data, output data, and expected output data. Then, server 200 can generate first prompt information corresponding to each data group based on the input data, output data, and expected output data in each data group through the first large model. Then, based on the first prompt information corresponding to each data group, second prompt information is generated by selecting one or more pairs of prompt information and performing information cross-processing on each selected pair of prompt information. Finally, the prompt information corresponding to the target large model can be determined based on the first and second prompt information corresponding to each data group. In this way, there is no need to fine-tune the model parameters of the large model or rely too much on expert experience. By balancing exploration and development capabilities through genetic algorithms, human intervention is reduced, achieving low-cost and high-efficiency prompt information optimization. Optimized prompt information can be generated for specific tasks, thereby improving the performance of the large model in dealing with various practical tasks.

[0020] like Figure 2 As shown in the embodiments of this specification, an information processing method is provided. The execution subject of this method can be a terminal device or a server, etc. The terminal device can be a mobile terminal device such as a mobile phone or tablet computer, a computer device such as a laptop or desktop computer, or an IoT device (specifically, a smartwatch, in-vehicle device, etc.). The server can be a single server or a server cluster composed of multiple servers. The server can be a backend server for financial business or online shopping business, or a backend server for an application, etc. This embodiment uses a server as the execution subject for detailed description. For the case where the execution subject is a terminal device, please refer to the following server case processing, which will not be repeated here. The method may specifically include the following steps: In step S202, a training dataset for training the target large model is obtained. The training dataset includes multiple data groups, each of which includes input data, output data, and expected output data.

[0021] The target large model can be any large model. Large models are complex artificial intelligence models trained on massive datasets, typically employing deep learning architectures (especially the Transformer architecture widely used in Natural Language Processing (NLP)). Large models have extremely high parameter counts, often reaching tens of billions or more, giving them powerful data processing and generation capabilities. Large models can include various types, such as large language models, visual language models, and multimodal models. Large language models, by learning language rules and patterns from massive amounts of text data, can demonstrate good performance on various NLP tasks, including but not limited to question answering, text generation, translation, summarization, dialogue interaction, and sentiment analysis. They also exhibit high flexibility and creativity in these tasks. Due to their strong generalization ability and high adaptability to language structure, large language models have become an important pillar of modern NLP technology, exemplified by GPT-3, GPT-4, and Tongyi 1000 Questions. The target large model can be a black-box large model (i.e., a large model with unknown or undisclosed parameters) or a large model with known or publicly disclosed parameters. Input data can be the input data of the target large model, and output data can be the data output by the target large model after the input data is input into it. Expected output data can be the desired output data corresponding to the above input data. Specifically, expected output data can be the data of the accurate answer corresponding to the input data, etc. Expected output data can be manually constructed data, or data generated by a specified algorithm, network model, or large model, etc. Specifically, if the target large model is a large language model, the input data can be text data or voice data such as user input questions, and the output data can be text data. If the target large model is a visual-language large model, the input data can include image data (or video data) and text data (which can be text data converted from voice data), etc., and the output data can include image data and / or text data, etc. If the target large model is a multimodal large model, the input data can include image data, text data, voice data, video data, etc., and the output data can include one or more of image data, text data, voice data, video data, etc., which can be set according to the actual situation.

[0022] In implementation, training datasets for training the target large model can be obtained in various ways. For example, a data set consisting of input data, output data, and expected output data can be obtained from a specified database. Multiple such data sets can be obtained and combined to form a training dataset. Alternatively, a data set consisting of input data, output data, and expected output data can be obtained from the Internet through web crawlers. Multiple such data sets can be obtained and combined to form a training dataset. Or, user-provided input data and output data of the target large model can be obtained from the business system where the target large model is deployed, along with the corresponding expected output data, thus obtaining a data set consisting of input data, output data, and expected output data. Multiple such data sets can be obtained and combined to form a training dataset, etc. The specific method can be set according to the actual situation.

[0023] In step S204, based on the input data, output data, and expected output data in each data group, the first prompt information corresponding to each data group is generated through the first large model.

[0024] The first large model can be any large model, and can include multiple types, such as a large language model, a visual language model, a multimodal model, etc. Furthermore, the first large model can be the aforementioned target large model, or a large model different from the target large model. The first prompt can be a prompt applied to the large model. The first prompt is text input into the large model to guide it in generating output results that meet the requirements. The first prompt can be a question, an instruction, or a piece of context, helping the large model better understand and respond to various queries. The first prompt can include explicit instruction prompts, context-supplementary prompts, and example-guided prompts. By providing detailed background information and examples, the first prompt can help the large model generate more relevant and accurate output results.

[0025] In implementation, a "reverse reasoning" approach can be used to encourage the large model to generate possible prompts. Specifically, a prompt for the first large model can be constructed based on the input data, output data, and expected output data in each data set. For example, based on a set of labeled data sets (containing input data, output data, and expected output data), a prompt to be assigned to the user (or labeler) can be inferred and generated. This prompt can then be input into the first large model, guiding it to generate the first prompt corresponding to each data set. The first prompt for each data set can be one or more, depending on the specific circumstances.

[0026] It should be noted that the prompt information corresponding to the training dataset can also be collected in a variety of different ways. This prompt information can be a prompt information designed by technicians or users, or a prompt information used in actual business applications, etc. The prompt information collected above can be used as the first prompt information and subsequent related processing can be performed.

[0027] In step S206, based on the first prompt information corresponding to each data group, a second prompt information is generated by selecting one or more pairs of prompt information and performing information cross-processing on each selected pair of prompt information.

[0028] In implementation, the generated first prompt information can be subjected to pairwise crossover and mutation based on the idea of ​​genetic algorithms, introducing randomness and exploration. Specifically, the crossover rate can be predefined to control the trigger probability of pairwise crossover processing of the generated first prompt information, and the crossover processing logic can be pre-set to generate crossover processing strategies for new prompt information. In addition, multiple rounds of iteration can be traversed to simulate the genetic process. Selection rules can be pre-set, such as random selection, selection based on similarity (i.e., selecting two first prompt information with a similarity greater than a preset threshold), or selection based on semantics (i.e., selecting two first prompt information with the same or similar semantics), etc., and one or more pairs of first prompt information can be selected from the generated first prompt information through these selection rules. For any one of the selected pairs of first prompt messages, the crossover rate defined above can be used to control whether the pair of first prompt messages needs to be cross-processed. If the crossover rate determines that the pair of first prompt messages needs to be cross-processed, the crossover processing logic set above can be invoked to generate a new prompt message. The crossover processing strategy for generating a new prompt message can then be used to cross-process the two first prompt messages in the pair, resulting in a new prompt message. This new prompt message can be used as the second prompt message. Alternatively, the new prompt message obtained from the crossover processing can be mutated. Specifically, specified characters (such as numbers or specified words) or specified statements in the new prompt message can be mutated (e.g., replacing specified characters with preset characters, replacing specified statements with preset strings, or replacing them with other statements with the same semantics). This mutated new prompt message can be used as the second prompt message. Other methods can also be used to determine the second prompt message, which can be set according to the actual situation. One or more different second prompt messages can be obtained through the above methods.

[0029] In step S208, based on the first and second prompt information corresponding to each data group, the prompt information corresponding to the target large model is determined.

[0030] In implementation, any two of the generated first and second prompts can be selected for comparison each time. Through pairwise comparison, the superior prompt is selected. This superior prompt is then compared pairwise with other superior prompts. Finally, through the above comparison, the superior prompt can be selected as the prompt corresponding to the target large model. Alternatively, each first prompt and the input data in each data group can be used to construct a corresponding prompt, which can then be input into the target large model to obtain the corresponding output data. Similarly, each second prompt and the input data in each data group can be used to construct a corresponding prompt, which can also be input into the target large model to obtain the corresponding output data. The obtained output data can be compared with the corresponding expected output data to obtain the similarity between them. Through the above methods, the numerical value of the similarity between each first and second prompt can be obtained. The prompt corresponding to the highest similarity value (either the first or second prompt) can be used as the prompt corresponding to the target large model, etc. Specific settings can be made according to actual circumstances; this embodiment does not limit this.

[0031] This specification provides an information processing method. By acquiring a training dataset for training a target large model, the training dataset includes multiple data groups, each containing input data, output data, and expected output data. Then, based on the input data, output data, and expected output data of each data group, a first prompt message corresponding to each data group is generated using a first large model. Based on the first prompt message corresponding to each data group, a second prompt message is generated by selecting one or more pairs of prompt messages and performing cross-processing on each selected pair. Finally, based on the first and second prompt messages corresponding to each data group, the prompt message corresponding to the target large model is determined. This method eliminates the need for fine-tuning the model parameters of the large model and avoids reliance on excessive expert experience. By balancing exploration and development capabilities through a genetic algorithm, it reduces human intervention and achieves low-cost, high-efficiency prompt message optimization. Optimized prompt messages can be generated for specific tasks, thereby improving the performance of the large model in handling various practical tasks.

[0032] In practical applications, the specific processing method of step S206 can vary. The following provides another optional processing method, which may specifically include the processing of steps S2062 and S2064. Based on this, in the above... Figure 2 Based on this, the specific steps included in this method can be as follows: Figure 3 As shown.

[0033] In step S2062, based on each first prompt message and the input data in each data group, the first output data corresponding to the input data in each data group is generated through the target large model. Based on the first output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, each first prompt message is optimized to obtain the third prompt message.

[0034] In implementation, for any given first prompt, the input data in each data group can be combined with that first prompt to obtain the corresponding prompt for each data group. Then, the prompt for each data group can be input into the target large model, guiding it to output the first output data corresponding to the input data in each data group. For other first prompts, the same process can be used to obtain the first output data corresponding to the input data in each data group under each first prompt. Combining the first output data obtained above, the first output data corresponding to the input data in each data group can be obtained.

[0035] Each first output data point and its corresponding expected output data can be encoded or feature-extracted separately to obtain a data representation for each first output data point and its corresponding expected output data. This data representation can be presented in the form of embedding vectors or matrices. For any data set, a similarity algorithm can be pre-defined, such as cosine similarity or Euclidean distance similarity. The similarity algorithm can be used to calculate the similarity between the data representation of the first output data corresponding to the input data in the data set and the data representation of the expected output data corresponding to the input data in the data set. Based on the calculated similarity, the differences between the first output data corresponding to the input data in the data set and the expected output data corresponding to the input data in the data set can be determined, thereby marking the specific location of the difference between the first output data corresponding to the input data in the data set and the expected output data corresponding to the input data in the data set. Using the above method, the similarity between the first output data corresponding to the input data in each data group and the expected output data corresponding to the input data in the corresponding data group can be obtained. This allows us to determine the difference between the first output data corresponding to the input data in each data group and the expected output data corresponding to the input data in the corresponding data group, and to mark the specific locations where the differences exist. For any given data group, there can be an information group (First Prompt, First Output Data, Expected Output Data, Difference Information). The difference information can include the difference between the first output data and the expected output data, and the specific locations where the differences exist. Based on the content of the information group, the optimization direction for the first prompt information corresponding to each data group can be determined. Based on the difference information, the first prompt information for each data group can be optimized along this direction to obtain the optimized first prompt information, which can then be used as the third prompt information.

[0036] The corresponding pseudocode for the above processing is as follows: # Define the main process For prompt in B: # Iterate through the Prompt collection B initialize metrics # Initialize metrics used to calculate overall accuracy For (input, expected output) in training dataset: # Iterate through the training dataset New Output = DriverLLM(prompt, Input) # Generates a new output based on the current Prompt driving the LLM. Similarity metric = ComputeSimilarity(new output, expected output) # Calculates the similarity between the new output and the expected output Record the difference points (new output, expected output) # Mark the specific location of the deviation based on the difference points Provide suggestions for improvement (prompt, new output, expected output, points of difference) # Generate suggestions for optimizing the prompt. Summarize the overall accuracy of the prompt (record similarity metric) # Calculate the overall accuracy of the current prompt on the training dataset. If optimization is necessary (overall accuracy of the prompt): # Determine whether the prompt needs optimization. New prompt = Optimized prompt(prompt, suggestions) # Adjust or enhance the prompt using suggestions The `UpdatePromptSet(prompt, newprompt)` method generates a `PromptSetC`. This method updates the optimized `Prompt` to the `PromptSetC`. Summary of the overall effect of the output prompt (overall accuracy of the prompt, suggestions for improvement) Here, set B is the set consisting of the first hints received.

[0037] In step S2064, one or more pairs of third prompt information are selected from the obtained third prompt information, and information cross-processing is performed on each pair of third prompt information to generate second prompt information.

[0038] In implementation, a crossover rate can be predefined to control the trigger probability of pairwise crossover processing of the obtained third-party prompts. Crossover processing logic can also be pre-defined to generate new prompts. Furthermore, multiple iterations can be performed to simulate a genetic process. Selection rules can be pre-defined, such as random selection, selection based on similarity, or selection based on semantics. These selection rules can then be used to select one or more pairs of third-party prompts from the obtained prompts. For any one of the selected pairs of third prompt messages, the crossover rate defined above can be used to control whether the pair of third prompt messages should be cross-processed. If the crossover rate determines that the pair of third prompt messages should be cross-processed, the cross-processing logic set above can be invoked to generate a new prompt message. This new prompt message can then be used as the second prompt message. Alternatively, the new prompt message obtained from the cross-processing can be mutated. Specifically, specified characters or statements in the new prompt message can be mutated to obtain a mutated new prompt message, which can then be used as the second prompt message. Other methods can also be used to determine the second prompt message, which can be set according to the actual situation. One or more different second prompt messages can be obtained through the above methods.

[0039] In practical applications, the specific processing method for generating the first output data corresponding to the input data in each data group based on each first prompt information and the input data in each data group in step S2062 can vary. The following provides another optional processing method, which may specifically include the processing in steps S20622 and S20624. Based on this, in the above... Figure 3 Based on this, the specific steps included in this method can be as follows: Figure 4 As shown.

[0040] In step S20622, based on the second model, each first prompt message is optimized through question-and-answer processing to obtain the fourth prompt message.

[0041] The second major model can be any major model, and can include multiple types, such as a major language model, a visual language model, a multimodal model, etc. In addition, the second major model can be the target major model mentioned above, or it can be a major model that is different from the target major model.

[0042] In implementation, the initial prompt information can be preliminarily optimized. Specifically, for each prompt information, a second major model can be used to optimize it in a question-and-answer format. During this process, common or frequently asked questions can be posed to the prompt information, and the second major model outputs corresponding answer data. This answer data is then used to optimize the prompt information. These questions can include various types, such as, "Can this prompt information be refined and broken down to make it easier to understand?", "What is the core task of this prompt information? Is the description of this prompt information clear?", "What is the task theme of this prompt information? Is it too broad?", etc. These questions can be pre-written with corresponding question templates based on expert experience, or set by technical personnel based on relevant business information of the target major model, or randomly generated question templates using a specified major language model. Afterwards, corresponding questions can be generated based on the first prompt information and the aforementioned question templates, as can be set according to the actual situation.

[0043] The above question can be input into the second model. Based on this question, the second model analyzes the first prompt information to determine the corresponding answer data. In this way, the question and answer data corresponding to each first prompt information can be obtained. The answer data from the second model can then be used to optimize the first prompt information to obtain the fourth prompt information.

[0044] In step S20624, based on each fourth prompt message and the input data in each data group, the first output data corresponding to the input data in each data group is generated through the target large model.

[0045] In implementation, for any fourth prompt, the input data in each data group can be combined with that fourth prompt to obtain the prompt corresponding to each data group. Then, the prompt corresponding to each data group can be input into the target large model, guiding the target large model to output the first output data corresponding to the input data in each data group. For other fourth prompts, the same method can be used to process them, obtaining the first output data corresponding to the input data in each data group under each fourth prompt.

[0046] In practical applications, the specific processing method for selecting one or more pairs of third prompt information from the obtained third prompt information in step S2064 can vary. The following provides another optional processing method, which specifically includes the processing steps S206402 to S206406. Based on this, in the above... Figure 4 Based on this, the specific steps included in this method can be as follows: Figure 5 As shown.

[0047] In step S206402, the similarity between different third prompts is determined based on the obtained third prompt information.

[0048] In implementation, similarity algorithms can be used to calculate the similarity between different third-party prompts. These algorithms can include various types, such as cosine similarity and Euclidean distance similarity algorithms, and the specific algorithm can be chosen based on the actual situation. Each third-party prompt can be encoded or have its features extracted to obtain a data representation corresponding to each prompt. This data representation can be presented in the form of embedding vectors or matrices. Then, the similarity between any two different third-party prompts can be calculated using the aforementioned similarity algorithms.

[0049] In step S206404, one or more pairs of third prompt information are constructed by constructing two third prompt information with a similarity greater than a preset threshold into a pair of third prompt information.

[0050] The preset threshold can be set according to the actual situation, such as 80% or 95%.

[0051] In step S206406, based on a preset selection probability, one or more pairs of third prompt information are selected from the constructed one or more pairs of third prompt information.

[0052] The selection probability can be set according to the actual situation, such as 70% or 60%.

[0053] The pseudocode for the aforementioned cross-processing is as follows: # Global configuration: Define hyperparameters define crossover_rate # Define the crossover rate, controlling the probability of triggering a crossover event. define generate_offspring(prompt_1, prompt_2) # Generate a new Prompt using custom cross-logic # Main Process: Genetic Iterative Optimization For generation in range(maximum number of iterations): # Iterate through multiple rounds to simulate the genetic process. New collection = [] # Used to store newly generated Prompt For prompt_1, prompt_2 in SelectPairs(Prompt set C): # Select two prompts from set C based on similarity or other strategies. If random() < crossover_rate: # Crossover rate controls whether crossover processing is performed offspring_1 = generate_offspring(prompt_1, prompt_2) # Custom crossover generation logic to generate offspring Prompt new_set.append(offspring_1) Else: new_set.append(prompt_1) # If no crossover occurs, directly retain the original Prompt new_set.append(prompt_2) new_set = Deduplicate(new_set) # Deduplication process to ensure there are no duplicate Prompts in the generated Prompt set performance_metrics = evaluate_performance(new_set) # Evaluate the overall performance of the current Prompt set Prompt set D = choose_best_prompts(new_set, performance_metrics) # Based on the evaluation results, select the better Prompts from the new set as the new set return Prompt set D # Return the optimized Prompt set Among them, set C is the set composed of the obtained third prompt information.

[0054] In practical applications, the specific processing method of the above step S208 can be various. Here is another optional processing method, which can specifically include the processing of the following step S2082 and step S2084. Based on this, in the above Figure 2 On this basis, the specific steps included in the method can be as Figure 6 shown.

[0055] In step S2082, based on the first prompt information and the second prompt information corresponding to each data group, as well as the input data in each data group, use the target large model to generate the second output data corresponding to the input data in each data group. Based on the second output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, optimize the first prompt information and the second prompt information corresponding to each data group to obtain the fifth prompt information.

[0056] The first prompt information for each data group can be all the first prompt information (i.e., including the first prompt information that has undergone cross-processing and the first prompt information that has not undergone cross-processing), or it can be the first prompt information that has not undergone cross-processing. The specific setting can be determined according to the actual situation.

[0057] In implementation, for any given first prompt, the input data in each data group can be combined with that first prompt to obtain the corresponding prompt for each data group. Then, the prompt for each data group can be input into the target large model, guiding the target large model to output the second output data corresponding to the input data in each data group. For other first prompts, the same method can be used to obtain the second output data corresponding to the input data in each data group under each first prompt.

[0058] Similarly, for any second prompt, the input data in each data group can be combined with that second prompt to obtain the corresponding prompt for each data group. Then, the prompt for each data group can be input into the target large model, guiding the target large model to output the second output data corresponding to the input data in each data group. For other second prompts, the same method can be used to obtain the second output data corresponding to the input data in each data group under each second prompt. Combining the second output data obtained above, the second output data corresponding to the input data in each data group can be obtained.

[0059] Each second output data point and its corresponding expected output data can be encoded or feature-extracted separately to obtain a data representation for each second output data point and its corresponding expected output data. This data representation can be presented in the form of embedding vectors or matrices. For any data set, a similarity algorithm can be pre-defined. This algorithm can calculate the similarity between the data representation of the second output data corresponding to the input data in that data set and the data representation of the expected output data corresponding to the input data in that data set. Based on the calculated similarity, the differences between the second output data corresponding to the input data in that data set and the expected output data corresponding to the input data in that data set can be determined, thereby pinpointing the specific locations where differences exist between the second output data corresponding to the input data in that data set and the expected output data corresponding to the input data in that data set.

[0060] By using the above method, the similarity between the second output data corresponding to the input data in each data group and the expected output data corresponding to the input data in the corresponding data group can be obtained. This allows for the determination of the difference between the second output data corresponding to the input data in each data group and the expected output data corresponding to the input data in the corresponding data group, and the specific location of the difference between the second output data corresponding to the input data in each data group and the expected output data corresponding to the input data in the corresponding data group can be marked. For any data group, there can be an information group (first prompt, second output data, expected output data, difference information) or (second prompt, second output data, expected output data, difference information). The difference information can include the difference between the second output data and the expected output data and the specific location of the difference. Based on the content of the above information group, the optimization direction of the first or second prompt for each data group can be determined. Based on the difference information, the first or second prompt for each data group can be optimized along this direction to obtain the optimized first or second prompt. The optimized first and second prompts can then be used as the fifth prompt.

[0061] In step S2084, based on the obtained fifth prompt information, the prompt information corresponding to the target large model is determined.

[0062] In implementation, any two fifth prompts can be selected from the obtained prompts each time for comparison. Through pairwise comparison, the better fifth prompt is selected. Then, the better fifth prompt is compared with other better fifth prompts. Finally, through the above comparison, the better fifth prompt can be selected as the prompt corresponding to the target large model.

[0063] In practical applications, the specific processing method of the above step S2084 can be varied. The following is another optional processing method, which may include the following: Based on the obtained fifth prompt information and the input data in each data group, generate the third output data corresponding to the input data in each data group through the target large model respectively; based on the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, determine the prompt information corresponding to the target large model from the obtained fifth prompt information.

[0064] In implementation, for any fifth prompt, the input data in each data group can be combined with that fifth prompt to obtain the prompt corresponding to each data group. Then, the prompt corresponding to each data group can be input into the target large model, guiding the target large model to output the third output data corresponding to the input data in each data group. For other fifth prompts, the same method can be used to process them, obtaining the third output data corresponding to the input data in each data group under each fifth prompt. Combining the third output data obtained above, the third output data corresponding to the input data in each data group can be obtained.

[0065] Each third output data can be compared with the corresponding expected output data to determine the difference between the two. The fifth prompt information corresponding to the smallest difference can be used as the prompt information corresponding to the target large model. Alternatively, the fifth prompt information corresponding to the difference less than the preset difference threshold can be obtained. The fifth prompt information corresponding to the difference less than the preset difference threshold can be randomly selected as the prompt information corresponding to the target large model, etc. The specific settings can be set according to the actual situation.

[0066] In practical applications, the specific processing method for determining the prompt information corresponding to the target large model from the obtained fifth prompt information based on the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group can be varied. The following provides another optional processing method, which can specifically include the processing of steps A2 and A4.

[0067] In step A2, based on the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the similarity between the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group is determined.

[0068] In implementation, each third output data and its corresponding expected output data can be encoded or feature extracted separately to obtain a data representation for each third output data and its corresponding expected output data. This data representation can be presented in the form of embedding vectors or matrices. For any data set, a similarity algorithm can be pre-defined to calculate the similarity between the data representation of the third output data corresponding to the input data in that data set and the data representation of the expected output data corresponding to the input data in that data set. Using the above method, the similarity between the data representation of the third output data corresponding to the input data in other data sets and the data representation of the expected output data corresponding to the input data in that data set can be calculated. Finally, the similarity between the third output data corresponding to the input data in each data set and the expected output data corresponding to the input data in each data set can be obtained.

[0069] In step A4, based on the similarity between the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the prompt information corresponding to the target large model is determined from the obtained fifth prompt information.

[0070] In implementation, the overall accuracy of each fifth prompt on the training dataset can be calculated based on the similarity between the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group. The fifth prompt corresponding to the highest overall accuracy can be used as the prompt corresponding to the target large model. Alternatively, the fifth prompt with an overall accuracy greater than a preset accuracy threshold can be obtained. The prompt corresponding to the target large model can be determined from the obtained fifth prompts by specifying selection rules (such as random selection rules or selection based on whether it contains specified information). The specific settings can be configured according to the actual situation.

[0071] The corresponding pseudocode for the above processing is as follows: For prompt in E: # Iterate through the Prompt collection E initialize metrics # Initialize metrics used to calculate overall accuracy For (input, expected output) in training dataset: # Iterate through the training dataset New Output = DriverLLM(prompt, Input) # Generates a new output based on the current Prompt driving the LLM. Similarity metric = ComputeSimilarity(new output, expected output) # Calculates the similarity between the new output and the expected output Summarize the overall accuracy of the prompt (record similarity metric) # Calculate the overall accuracy of the current prompt on the training dataset. Return the prompt with the highest overall accuracy as the final prompt. Here, set E can be a set consisting of the fifth hint information obtained.

[0072] In practical applications, the specific processing method for selecting one or more pairs of third prompt information from the obtained third prompt information in step S2064 can vary. The following provides another optional processing method, which specifically includes the processing steps S206408 to S206412. Based on this, in the above... Figure 3 Based on this, the specific steps included in this method can be as follows: Figure 7 As shown.

[0073] In step S206408, based on the obtained third prompt information and the input data in each data group, the fourth output data corresponding to the input data in each data group is generated through the target large model.

[0074] In implementation, reinforcement learning can be introduced to guide the selection and mutation processing of the genetic algorithm. Specifically, for any third cue information, the input data in each data group can be synthesized with the third cue information to obtain the cue information corresponding to each data group. Then, the cue information corresponding to each data group can be input into the target large model, guiding the target large model to output the fourth output data corresponding to the input data in each data group. For other third cue information, the same method can be used to process them to obtain the fourth output data corresponding to the input data in each data group under each third cue information. Combining the fourth output data obtained above, the fourth output data corresponding to the input data in each data group can be obtained.

[0075] In step S206410, based on the fourth output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the reward information of the fourth output data corresponding to the input data in each data group is determined by the reward model.

[0076] The reward model can be a model that gives corresponding rewards based on the performance of the generated third cue information in the task. The reward model can be a model constructed through the reward function in reinforcement learning.

[0077] In implementation, a set of prompt information samples can be obtained. The prompt information samples in the set can be input into the target large model to obtain the output data sample corresponding to each prompt information sample. The output data sample corresponding to each prompt information sample and the expected output data sample corresponding to the corresponding prompt information sample are input into the reward model to obtain the reward information corresponding to each prompt information sample. Based on the reward information corresponding to each prompt information sample and the sample reward information corresponding to the corresponding prompt information sample, the reward model is trained through a preset loss function to obtain the trained reward model.

[0078] For any given dataset, the fourth output data corresponding to the input data in that dataset and the expected output data corresponding to the input data in that dataset can be input into the trained reward model to obtain the reward information for the fourth output data corresponding to the input data in that dataset. In this way, the reward information for the fourth output data corresponding to the input data in each dataset can be obtained.

[0079] In step S206412, based on the reward information of the fourth output data corresponding to the input data in each data group, one or more pairs of third prompt information are selected from the obtained third prompt information.

[0080] In practice, the performance of the corresponding third prompt information in the process of generating the corresponding output data of the target large model can be determined based on the reward information of the fourth output data corresponding to the input data in each data group. Then, the better-performing third prompt information can be selected to construct one or more pairs of third prompt information.

[0081] Furthermore, the mutation processing of the genetic algorithm can be guided by the aforementioned reinforcement learning. Specifically, based on the new prompt information obtained from the crossover process and the input data in each data group, the fifth output data corresponding to the input data in each data group is generated through the target large model; based on the fifth output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the reward information of the fifth output data corresponding to the input data in each data group is determined through the reward model; based on the reward information of the fifth output data corresponding to the input data in each data group, the prompt information that needs to be mutated is selected from the new prompt information obtained from the aforementioned crossover process.

[0082] The specific processing steps described above can be found in the aforementioned related content, and will not be repeated here.

[0083] In practical applications, a user feedback mechanism can also be introduced, enabling the server to make real-time adjustments and optimizations based on user feedback. For details, please refer to steps B2 and B4 below.

[0084] In step B2, feedback information from users regarding the prompts corresponding to the target large model is received through a preset user feedback interface.

[0085] In step B4, based on the user's feedback on the prompt information corresponding to the target large model, the prompt information corresponding to the target large model is adjusted to obtain the adjusted prompt information corresponding to the target large model.

[0086] This specification provides an information processing method. By acquiring a training dataset for training a target large model, the training dataset includes multiple data groups, each containing input data, output data, and expected output data. Then, based on the input data, output data, and expected output data of each data group, a first prompt message corresponding to each data group is generated using a first large model. Based on the first prompt message corresponding to each data group, a second prompt message is generated by selecting one or more pairs of prompt messages and performing cross-processing on each selected pair. Finally, based on the first and second prompt messages corresponding to each data group, the prompt message corresponding to the target large model is determined. This method eliminates the need for fine-tuning the model parameters of the large model and avoids reliance on excessive expert experience. By balancing exploration and development capabilities through a genetic algorithm, it reduces human intervention and achieves low-cost, high-efficiency prompt message optimization. Optimized prompt messages can be generated for specific tasks, thereby improving the performance of the large model in handling various practical tasks.

[0087] Furthermore, even without external feedback, the model can self-correct through a self-correction mechanism and multiple generated inference chains to improve the answer data. Fine-tuning, on the other hand, typically requires external feedback to adjust and optimize the model's output data, thus improving the accuracy, robustness, generalization ability, and decision transparency of large models. Simultaneously, based on genetic algorithms and feedback mechanisms, it alleviates the lack of feedback mechanisms and the balance between exploration and development inherent in black-box hint optimization, mitigating local optima to some extent. Moreover, it is more practical, easier to implement, and easier to promote in real-world applications.

[0088] The above describes the information processing method provided in the embodiments of this specification. Based on the same idea, the embodiments of this specification also provide an information processing device, such as... Figure 8 As shown.

[0089] The information processing device includes: a dataset acquisition module 801, an information processing module 802, a cross-processing module 803, and a prompt information determination module 804, wherein: The dataset acquisition module 801 acquires a training dataset for training the target large model. The training dataset includes multiple data groups, and each data group includes input data, output data, and expected output data. The information processing module 802 generates the first prompt information corresponding to each data group based on the input data, output data and expected output data in each data group through the first large model; The cross-processing module 803 generates second prompt information based on the first prompt information corresponding to each data group by selecting one or more pairs of prompt information and performing cross-processing on each selected pair of prompt information. The prompt information determination module 804 determines the prompt information corresponding to the target large model based on the first prompt information and the second prompt information corresponding to each data group.

[0090] In the embodiments of this specification, the cross-processing module 803 includes: The first optimization unit generates first output data corresponding to the input data in each data group through the target large model based on each first prompt information and the input data in each data group. Based on the first output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the first prompt information is optimized to obtain the third prompt information. The cross-processing unit selects one or more pairs of third prompt information from the obtained third prompt information, and performs cross-processing on each pair of third prompt information to generate the second prompt information.

[0091] In this embodiment of the specification, the first optimization unit optimizes each first prompt message based on the second large model through a question-and-answer method to obtain a fourth prompt message; based on each fourth prompt message and the input data in each data group, the first output data corresponding to the input data in each data group is generated through the target large model.

[0092] In the embodiments of this specification, the cross-processing unit determines the similarity between different third prompt information based on the obtained third prompt information; constructs one or more pairs of third prompt information by constructing two third prompt information with similarity greater than a preset threshold as a pair of third prompt information; and selects one or more pairs of third prompt information from the constructed pair of third prompt information based on a preset selection probability.

[0093] In this embodiment of the specification, the prompt information determination module 804 includes: The second optimization unit, based on the first prompt information and the second prompt information corresponding to each data group, and the input data in each data group, generates the second output data corresponding to the input data in each data group through the target large model. Based on the second output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the unit optimizes the first prompt information and the second prompt information corresponding to each data group to obtain the fifth prompt information. The prompt information determination unit determines the prompt information corresponding to the target large model based on the obtained fifth prompt information.

[0094] In this embodiment of the specification, the prompt information determination unit, based on the obtained fifth prompt information and the input data in each data group, generates third output data corresponding to the input data in each data group through the target large model, and determines the prompt information corresponding to the target large model from the obtained fifth prompt information based on the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group.

[0095] In this embodiment of the specification, the prompt information determination unit determines the similarity between the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, based on the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group; and determines the prompt information corresponding to the target large model from the obtained fifth prompt information based on the similarity between the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group.

[0096] In the embodiments of this specification, the cross-processing unit generates fourth output data corresponding to the input data in each data group based on the obtained third prompt information and the input data in each data group through the target large model; based on the fourth output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, it determines the reward information of the fourth output data corresponding to the input data in each data group through the reward model; and based on the reward information of the fourth output data corresponding to the input data in each data group, it selects one or more pairs of third prompt information from the obtained third prompt information.

[0097] In the embodiments described in this specification, the device further includes: The feedback receiving module receives user feedback information regarding the prompt information corresponding to the target large model through a preset user feedback interface; The prompt information adjustment module adjusts the prompt information corresponding to the target large model based on user feedback, thereby obtaining the adjusted prompt information corresponding to the target large model.

[0098] For ease of description, the above devices are described by dividing them into various modules or units based on their functions. Of course, when implementing one or more embodiments of this specification, the functions of each module or unit can be implemented in one or more software and / or hardware components, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are merely illustrative; the division of each module and unit is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or modules can be combined or integrated into another system, or some features can be ignored or not executed, etc.

[0099] This specification provides an information processing device that acquires a training dataset for training a target large model. The training dataset includes multiple data groups, each containing input data, output data, and expected output data. Based on the input data, output data, and expected output data of each data group, a first prompt message corresponding to each data group is generated using a first large model. Then, based on the first prompt message for each data group, a second prompt message is generated by selecting one or more pairs of prompt messages and performing cross-processing on each selected pair. Finally, based on the first and second prompt messages for each data group, the prompt message corresponding to the target large model is determined. This eliminates the need for fine-tuning the model parameters of the large model and avoids reliance on excessive expert experience. By balancing exploration and development capabilities through a genetic algorithm, manual intervention is reduced, achieving low-cost and high-efficiency prompt message optimization. Optimized prompt messages can be generated for specific tasks, thereby improving the performance of the large model in handling various practical tasks.

[0100] Furthermore, even without external feedback, the model can self-correct through a self-correction mechanism and multiple generated inference chains to improve the answer data. Fine-tuning, on the other hand, typically requires external feedback to adjust and optimize the model's output data, thus improving the accuracy, robustness, generalization ability, and decision transparency of large models. Simultaneously, based on genetic algorithms and feedback mechanisms, it alleviates the lack of feedback mechanisms and the balance between exploration and development inherent in black-box hint optimization, mitigating local optima to some extent. Moreover, it is more practical, easier to implement, and easier to promote in real-world applications.

[0101] The above are the information processing apparatuses provided in the embodiments of this specification. Based on the same idea, the embodiments of this specification also provide an information processing device, such as... Figure 9 As shown.

[0102] The information processing device can provide terminal equipment or servers, etc., for the above embodiments.

[0103] Information processing devices can vary considerably due to differences in configuration or performance. They may include a communication interface 902, a user interface 904, a processor 906, and data storage 908. These components are interconnected and communicate with each other via a system bus, network, or other connection mechanism 910. The communication interface 902 enables the information processing device 900 to communicate with other devices, access networks, and transmission networks via analog or digital modulation. For example, the communication interface 902 may include a chipset and antenna for wireless communication with a radio access network or access point. Furthermore, the communication interface 902 can be a wired interface such as Ethernet, Token Ring, or a USB port, or a wireless interface such as Wi-Fi, Bluetooth, Global Positioning System (GPS), or a wide-area wireless interface (e.g., WiMAX or LTE). Of course, the communication interface 902 may also support other forms of physical layer interfaces and standard or proprietary communication protocols. The communication interface 902 may also include multiple physical communication interfaces, such as Wi-Fi, Bluetooth, and wide-area wireless interfaces.

[0104] User interface 904 includes receiving user input and providing output to the user. Therefore, user interface 904 may include input components such as a keypad, keyboard, touch-sensitive or presence-sensitive panel, computer mouse, trackball, joystick, microphone, still camera, and video camera, and output components such as a display screen (which may be combined with a touch-sensitive panel), CRT, LCD, LED, display using DLP technology, printer, and other similar devices known or developed in the future. User interface 904 may also generate auditory output via speakers, speaker jacks, audio output ports, audio output devices, headphones, and other similar devices known or developed in the future. In some embodiments, user interface 904 may include software, circuitry, or other forms of logic capable of transmitting and receiving data from external user input / output devices. Additionally or alternatively, information processing device 900 may support remote access from other devices via communication interface 902 or another physical interface (not shown). User interface 904 may be configured to receive user input, the position and movement of which may be indicated by indicators or cursors described herein. User interface 904 may also be configured as a display device for rendering or displaying text fragments.

[0105] The processor 906 may contain one or more general-purpose processors and / or special-purpose processors.

[0106] Data storage 908 may include one or more volatile and / or non-volatile storage components and may be integrated wholly or partially with processor 906. Data storage 908 may include removable and non-removable components.

[0107] Processor 906 is capable of executing program instructions 918 (e.g., compiled or uncompiled program logic and / or machine code) stored in data storage 908 to perform the various functions described herein. Data storage 908 may contain a non-transitory computer-readable medium on which program instructions are stored, which, when executed by information processing device 900, enable information processing device 900 to perform any methods, processes, or functions disclosed in this specification and / or the accompanying drawings. Execution of program instructions 918 by processor 906 may result in processor 906 using data 912.

[0108] For example, program instructions 918 may include an operating system 922 (e.g., an operating system kernel, device drivers, and / or other modules) installed on the information processing device 900, and one or more application programs 920 (e.g., a browser, social application, or game application). Similarly, data 912 may include operating system data 916 and application data 914. Operating system data 916 is primarily accessible to the operating system 922, while application data 914 is primarily accessible to one or more application programs 920. Application data 914 may reside in a file system visible or hidden from the user of the information processing device 900.

[0109] Application 920 can communicate with operating system 912 through one or more application programming interfaces (APIs). These APIs help application 920 read and / or write application data 914, transmit or receive information via communication interface 902, receive or display information on user interface 904, etc.

[0110] In some terminology, application 920 may be simply referred to as "app". Furthermore, application 920 can be downloaded to information processing device 900 through one or more online app stores or app markets. However, applications can also be installed on information processing device 900 in other ways, such as through a web browser or a physical interface on information processing device 900 (e.g., a USB port).

[0111] Specifically, in this embodiment, the information processing device 900 includes a data storage 908 and one or more program instructions 918, wherein one or more program instructions 918 are stored in the data storage 908, and one or more program instructions 918 are configured to be executed by one or more processors. The one or more program instructions include computer-executable instructions for performing the following: Obtain a training dataset for training the target large model, wherein the training dataset includes multiple data groups, each data group including input data, output data and expected output data; Based on the input data, output data, and expected output data in each data group, the first prompt information corresponding to each data group is generated through the first major model; Based on the first prompt information corresponding to each data group, a second prompt information is generated by selecting one or more pairs of prompt information and performing cross-processing on each selected pair of prompt information. Based on the first and second prompts for each data group, the prompts for the target large model are determined.

[0112] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the information processing device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0113] This specification provides an information processing device that acquires a training dataset for training a target large model. The training dataset includes multiple data groups, each containing input data, output data, and expected output data. Based on the input data, output data, and expected output data of each data group, a first prompt message corresponding to each data group is generated using a first large model. Then, based on the first prompt message for each data group, a second prompt message is generated by selecting one or more pairs of prompt messages and performing cross-processing on each selected pair. Finally, based on the first and second prompt messages for each data group, the prompt message corresponding to the target large model is determined. This eliminates the need for fine-tuning the model parameters of the large model and avoids reliance on excessive expert experience. By balancing exploration and development capabilities through a genetic algorithm, manual intervention is reduced, achieving low-cost and high-efficiency prompt message optimization. Optimized prompt messages can be generated for specific tasks, thereby improving the performance of the large model in handling various practical tasks.

[0114] Furthermore, based on the above Figures 1 to 7 This specification also provides a storage medium for storing computer-executable instruction information in one or more embodiments. In one specific embodiment, the storage medium may be a USB flash drive, optical disc, hard disk, etc. When the computer-executable instruction information stored in the storage medium is executed by a processor, it can realize the following process: Obtain a training dataset for training the target large model, wherein the training dataset includes multiple data groups, each data group including input data, output data and expected output data; Based on the input data, output data, and expected output data in each data group, the first prompt information corresponding to each data group is generated through the first major model; Based on the first prompt information corresponding to each data group, a second prompt information is generated by selecting one or more pairs of prompt information and performing cross-processing on each selected pair of prompt information. Based on the first and second prompts for each data group, the prompts for the target large model are determined.

[0115] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the above-described storage medium embodiment is basically similar to the method embodiment, so the description is relatively simple; relevant parts can be referred to the description of the method embodiment.

[0116] This specification provides a storage medium that acquires a training dataset for training a target large model. The training dataset includes multiple data groups, each containing input data, output data, and expected output data. Based on the input data, output data, and expected output data of each data group, a first prompt message corresponding to each data group is generated by a first large model. Then, based on the first prompt message for each data group, a second prompt message is generated by selecting one or more pairs of prompt messages and performing cross-processing on each selected pair. Finally, based on the first and second prompt messages for each data group, the prompt message corresponding to the target large model is determined. This eliminates the need for fine-tuning the model parameters of the large model and avoids reliance on excessive expert experience. By balancing exploration and development capabilities through a genetic algorithm, manual intervention is reduced, achieving low-cost and high-efficiency prompt message optimization. Optimized prompt messages can be generated for specific tasks, thereby improving the performance of the large model in handling various practical tasks.

[0117] Furthermore, based on the above Figures 1 to 7 This specification also provides one or more embodiments of a computer program product, including a computer program, which, when executed by a processor, can perform the following processes: Obtain a training dataset for training the target large model, wherein the training dataset includes multiple data groups, each data group including input data, output data and expected output data; Based on the input data, output data, and expected output data in each data group, the first prompt information corresponding to each data group is generated through the first major model; Based on the first prompt information corresponding to each data group, a second prompt information is generated by selecting one or more pairs of prompt information and performing cross-processing on each selected pair of prompt information. Based on the first and second prompts for each data group, the prompts for the target large model are determined.

[0118] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the above-described embodiment of a computer program product is relatively simple in description because it is fundamentally similar to the method embodiment; relevant parts can be referred to the description of the method embodiment.

[0119] This specification provides a computer program product that acquires a training dataset for training a target large model. The training dataset includes multiple data groups, each containing input data, output data, and expected output data. Based on the input data, output data, and expected output data of each data group, a first prompt message corresponding to each data group is generated using a first large model. Then, based on the first prompt message for each data group, a second prompt message is generated by selecting one or more pairs of prompt messages and performing cross-processing on each selected pair. Finally, based on the first and second prompt messages for each data group, the prompt message corresponding to the target large model is determined. This eliminates the need for fine-tuning the model parameters of the large model and avoids reliance on excessive expert experience. By balancing exploration and development capabilities through a genetic algorithm, manual intervention is reduced, achieving low-cost and high-efficiency prompt message optimization. Optimized prompt messages can be generated for specific tasks, thereby improving the performance of the large model in handling various practical tasks.

[0120] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous. Moreover, although one or more embodiments of this specification provide method steps as described in the embodiments or flowcharts, it is understood that the order of steps listed in the embodiments or flowcharts is merely one possible execution order among many steps and does not represent the only execution order. Therefore, when method steps are involved in the claims, adjustments to the order of those steps, or parallelism between steps, are also within the scope of protection of the claims.

[0121] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, 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, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0122] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0123] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0124] For ease of description, the above apparatus is described by dividing it into various functional units. Of course, when implementing one or more embodiments of this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0125] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, one or more embodiments of this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of 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.

[0126] Embodiments in this specification are described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable parallel device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable parallel device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable fraud device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions can also be loaded onto a computer or other programmable device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0130] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0131] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, 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 technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0132] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical or equivalent elements in the process, method, article, or apparatus that includes said element. Furthermore, "a," "an," and "the" are not specifically singular and may include plural forms. Ordinal numbers such as "first," "second," etc., do not necessarily indicate order; they are often used to distinguish objects. For example, "first server" and "second server" usually refer to two servers, described as "first server" and "second server" to differentiate them; however, sometimes these two servers may be the same server. Moreover, in this specification, unless explicitly stated otherwise, "receiving and sending data" does not necessarily mean direct receiving and sending; it can be indirect receiving and sending (i.e., receiving and sending indirectly through one or more entities). Similarly, in this specification, unless otherwise stated, the relationships between structures can be direct or indirect.

[0133] Furthermore, the specific terms used in this specification to describe embodiments, such as "an embodiment," "one embodiment," or "some embodiments," refer to a particular feature, structure, or characteristic related to at least one embodiment of this specification. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different locations in this specification do not necessarily refer to the same embodiment. Moreover, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples, without contradiction.

[0134] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, one or more embodiments of this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of 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.

[0135] One or more embodiments of this specification can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. One or more embodiments of this specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0136] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0137] The above description is merely an embodiment of this specification and is not intended to limit this document. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims in this document.

Claims

1. An information processing method, the method comprising: Obtain a training dataset for training the target large model, wherein the training dataset includes multiple data groups, each data group including input data, output data and expected output data; Based on the input data, output data, and expected output data in each data group, the first prompt information corresponding to each data group is generated through the first major model; Based on the first prompt information corresponding to each data group, a second prompt information is generated by selecting one or more pairs of prompt information and performing cross-processing on each selected pair of prompt information. Based on the first and second prompts for each data group, the prompts for the target large model are determined.

2. The method according to claim 1, wherein generating second prompt information based on the first prompt information corresponding to each data group by selecting one or more pairs of prompt information and performing information cross-processing on each selected pair of prompt information includes: Based on each first prompt and the input data in each data group, the first output data corresponding to the input data in each data group is generated through the target large model. Based on the first output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, each first prompt is optimized to obtain the third prompt. Select one or more pairs of third prompt information from the obtained third prompt information, and perform information cross-processing on each pair of third prompt information to generate the second prompt information.

3. The method according to claim 2, wherein generating the first output data corresponding to the input data in each data group based on each first prompt information and the input data in each data group through the target large model comprises: Based on the second major model, each of the first prompt messages is optimized through a question-and-answer process to obtain the fourth prompt message; Based on each fourth prompt and the input data in each data group, the first output data corresponding to the input data in each data group is generated through the target large model.

4. The method according to claim 2, wherein selecting one or more pairs of third prompt information from the obtained third prompt information includes: Based on the obtained third-party prompt information, determine the similarity between different third-party prompt information; One or more pairs of third prompt information can be constructed by combining two third prompt information with a similarity greater than a preset threshold into a pair of third prompt information; Based on a preset selection probability, select one or more pairs of third prompt information from the constructed pair or more pairs of third prompt information.

5. The method according to claim 1, wherein determining the prompt information corresponding to the target large model based on the first prompt information and the second prompt information corresponding to each data group includes: Based on the first and second prompts for each data group, and the input data in each data group, the second output data corresponding to the input data in each data group is generated through the target large model. Based on the second output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the first and second prompts for each data group are optimized to obtain the fifth prompt. Based on the fifth prompt information obtained, the prompt information corresponding to the target large model is determined.

6. The method according to claim 5, wherein determining the prompt information corresponding to the target large model based on the obtained fifth prompt information includes: Based on the obtained fifth prompt information and the input data in each data group, the third output data corresponding to the input data in each data group is generated through the target large model. Based on the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the prompt information corresponding to the target large model is determined from the obtained fifth prompt information.

7. The method according to claim 6, wherein determining the prompt information corresponding to the target large model from the obtained fifth prompt information based on the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group includes: Based on the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, determine the similarity between the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group; Based on the similarity between the third output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the prompt information corresponding to the target large model is determined from the obtained fifth prompt information.

8. The method according to claim 2, wherein selecting one or more pairs of third prompt information from the obtained third prompt information includes: Based on the obtained third prompt information and the input data in each data group, the fourth output data corresponding to the input data in each data group is generated through the target large model; Based on the fourth output data corresponding to the input data in each data group and the expected output data corresponding to the input data in each data group, the reward information of the fourth output data corresponding to the input data in each data group is determined by the reward model; Based on the reward information of the fourth output data corresponding to the input data in each data group, select one or more pairs of third prompt information from the obtained third prompt information.

9. The method according to claim 1, further comprising: The system receives user feedback on the prompts corresponding to the target large model through a preset user feedback interface. Based on user feedback regarding the prompts corresponding to the target large model, the prompts corresponding to the target large model are adjusted to obtain the adjusted prompts corresponding to the target large model.

10. An information processing apparatus, the apparatus comprising: The dataset acquisition module acquires a training dataset for training a target large model. The training dataset includes multiple data groups, each of which includes input data, output data, and expected output data. The information processing module generates the first prompt information for each data group based on the input data, output data, and expected output data in each data group, using the first major model. The cross-processing module generates a second prompt message based on the first prompt message corresponding to each data group by selecting one or more pairs of prompt messages and performing cross-processing on each selected pair of prompt messages. The prompt information determination module determines the prompt information corresponding to the target large model based on the first prompt information and the second prompt information corresponding to each data group.

11. An information processing device, the information processing device comprising: processor; as well as A memory configured to store computer-executable instructions, which, when executed, cause the processor to: Obtain a training dataset for training the target large model, wherein the training dataset includes multiple data groups, each data group including input data, output data and expected output data; Based on the input data, output data, and expected output data in each data group, the first prompt information corresponding to each data group is generated through the first major model; Based on the first prompt information corresponding to each data group, a second prompt information is generated by selecting one or more pairs of prompt information and performing cross-processing on each selected pair of prompt information. Based on the first and second prompts for each data group, the prompts for the target large model are determined.

Images