Large language model optimization system and method based on low-rank adaptive technology

Abstract

The invention discloses a large language model optimization system and method based on a low-rank adaptive technology, and relates to the technical field of large language models, the optimization system comprises a low-rank adjustment module, an instruction preliminary analysis module, a data comparison module, an instruction optimization module and a model evaluation module; the method is technically characterized in that a large language model optimized by a low-rank self-adaptive technology is used as a data processing basis, so that subsequent continuous optimization is facilitated, and meanwhile, a data volume cardinal number needing to be processed can be effectively reduced; and according to the proportion of the preferred data set in all the preliminary retrieval data sets, a'question 'is sent to the'input end', so that the operation instruction is optimized, 'internal friction 'of the system is avoided, and through the'question' mode, the purpose of more accurately determining the operation instruction can be achieved, so that accurate feedback can be given, and the operation efficiency is improved. And the total amount of the data to be processed is reduced to a greater extent, so that the problem of excessive data to be processed is solved fundamentally.

Description

Technical Field

[0001] The present invention relates to the technical field of large language models, and specifically to a large language model optimization system and method based on low-rank adaptive technology. Background Art

[0002] In recent years, thanks to the support of big data and the enhancement of computing power, artificial intelligence technology has developed rapidly, and large language models based on generative artificial intelligence technology have emerged. It has performed well in simulating human thinking logic and language processing capabilities, laying the foundation for the development of intelligent auxiliary tools and intelligent assistants. All walks of life have also begun to explore the application of large language models (LLM) in their own fields. However, with the widespread application of large language models, reasoning efficiency and resource consumption have become the main bottlenecks restricting their expansion in practical applications, especially in complex tasks such as retrieval data generation and task identification.

[0003] At present, the Chinese patent with the existing patent application number "CN202410402548.2" discloses a large language model fine-tuning method based on low-rank matrix decomposition, including obtaining the pre-trained weight file of the large language model and the question-answer pairs required for fine-tuning the large language model; using the double quantization method to compress the model parameter accuracy in the pre-trained weight file; replacing the fully connected layer in the large language model structure with the LoRA layer; dequantizing the model parameters after precision compression twice in batches and calculating the output of the LoRA layer; in each batch, calculating the loss according to the output of the LoRA layer, and dynamically updating the parameters of the LoRA layer based on the loss back propagation until all batches are processed and the fine-tuned large language model is output. Although it is possible to master the knowledge of multiple fields by setting multiple parallel LoRA modules, and then use the routing module to direct the input questions to the LoRA module responsible for the field to obtain the output, it is mainly in terms of how to improve the data processing speed, such as the above, by dividing the required operation data into multiple independent branch operation data, and then multiple operation devices perform synchronous operations, etc. However, it ignores how to improve the optimization of the accuracy of the acquired instructions.

[0004] In addition, the Chinese patent with the existing patent application number "CN202411045654.6" discloses a large language model acceleration method and device, which has the following steps: S1, compress and decompose the pre-trained weight matrix W of the large language model; S2, perform QR decomposition on the U and VT matrices respectively to obtain QU, RU, QV, RV; S3, use the singular value decomposition algorithm to compress and decompose the matrix product; S4, merge the matrices respectively; S5, replace the weight matrix W with the obtained matrix for storage; S6, use the stored matrix for reasoning acceleration; S7, set the obtained rank r to the rank of the low-rank parameterized update matrix ΔW corresponding to the weight matrix W; S8, use the stored matrix to fine-tune the reasoning acceleration of the large language model. It also focuses on how to improve the speed of the operation itself, but whether it is through improving the operation speed of the equipment or adjusting the operation method, it will take a lot of time and energy for research and development, and it is necessary to find another way.

[0005] However, during the implementation of the above technical solution, it was found that there were at least the following technical problems: Unable to optimize the instructions to be executed: The existing large language model improvements are mainly divided into two categories (hardware and software). One is to improve the computing power of the computing device by hardware structure, for example, the processor Intel Core i9-14900K and the processor Intel Core i7-14700K, the i9 has 16 cores and 32 threads, while the i7 has 8 cores and 16 threads. It can be seen that the i9 can handle multiple tasks at the same time compared to the i7, reducing the CPU occupancy rate to improve computing power. This method requires a lot of time and energy to invest in research and development, and the research and development is difficult. At the same time, due to the inability to filter or optimize the data, the amount of data that needs to be processed remains unchanged, and this problem cannot be fundamentally solved; the other type is to reduce the amount of data that needs to be calculated by splitting and adjusting the retrieval instructions. However, no matter what calculation method is used, it is necessary to process and execute according to the received retrieval instructions. For example, when issuing a historical data retrieval instruction, since it is impossible to accurately determine which historical data is required, all data can only be retrieved. After system optimization, only the previous retrieval habits are added, so that the data is divided into before and after or important or not. However, as time goes by, the data continues to accumulate, resulting in the data that needs to be processed is still very large, and the data optimization has reached a bottleneck, so it is difficult to have a new breakthrough. To sum up, it can be seen that both hardware and software improvements are very difficult and have limited effects. Therefore, we need to open up a new way to optimize and adjust large language models. Summary of the invention

[0006] 1. Technical issues to be resolved In view of the deficiencies in the prior art, the present invention provides a large language model optimization system and method based on low-rank adaptive technology, which solves the problems raised in the background technology.

[0007] (II) Technical solution To achieve the above objectives, the present invention is implemented through the following technical solutions: A large language model optimization system based on low-rank adaptive technology, the optimization system includes: The low-rank adjustment module uses the cleaned and preprocessed input data as the data set for training the large language model, trains the large language model, and then uses the low-rank adaptive technology to perform low-rank decomposition and optimization on the initial model. The low-rank decomposition is to decompose the original high-dimensional parameter matrix into the product of two or more low-rank matrices. The instruction preliminary analysis module extracts the keywords in the received initial instruction as search terms, obtains the search data that can be called by the large language model, and divides it into corresponding search data sets according to the type. Then, according to the number of subsets in the search data set, it is divided into preliminary search data sets and discardable search data sets, and the discardable search data sets are discarded; The data comparison module, when receiving the preliminary search data set, sorts the data sets from large to small according to the number of subsets in the preliminary search data set, selects the first several preliminary search data sets as the preferred data sets, and compares the proportion of the preferred data sets in all preliminary search data sets with the preset screening ratio threshold. When the proportion is less than the screening ratio threshold, the preliminary search data sets are output as the search results in the sorted order; otherwise, the comparison data is output as the first search result, and the estimated impact index of the execution of the verification instruction is analyzed, and the verification instruction is issued; The instruction optimization module, when receiving the verification instruction, performs secondary classification on the preferred data set according to the classification standard stored in the database, analyzes the data classification evaluation value of the secondary classification, combines it with the historical search trend to select the corresponding classification standard as the optimization instruction output, and then records the instruction after the optimization instruction feedback as the secondary instruction, and screens and outputs the search data after the secondary classification; The model evaluation module obtains the proportion of the preferred data set in the preliminary search data set during the retrieval process, the estimated impact index, and the number of subsets of the preferred data set before and after instruction optimization, generates a model optimization efficiency evaluation coefficient, and calls the optimization solution correction strategy when the model optimization efficiency evaluation coefficient is lower than the preset optimization effect threshold.

[0008] Furthermore, the input data passes through all training layers, and the output of the training layer is combined with the input adjusted by the new parameters to generate a new adjusted output; and in the case of low-rank decomposition, importance-based pruning or structure-based pruning is used to remove unimportant parameters in the model.

[0009] Furthermore, the subset of the search data set represents the search data that can be directly obtained according to the search terms. When the search data set is classified, it is compared with the preset rejection threshold. When the number of subsets in the search data set is less than the preset rejection threshold, it is recorded as a discardable search data set; otherwise, it is recorded as a preliminary search data set. Among them, when searching according to the secondary instruction, if the number of subsets in the discardable search data set increases, and the ratio of the increase to the number of atomic sets is greater than 20%, a reuse instruction is issued to record the corresponding discardable search data set as a preliminary search data set; otherwise, the corresponding discardable search data set is recorded as a non-search data set under the search term, and is not displayed in the callable search data retrieved by the same search term.

[0010] Furthermore, the process of adjusting the discard threshold is as follows: Set the adjustment ratio, denoted as ; Get the number of discardable search datasets under the current search term and the number of preliminary search datasets, calculate the ratio of the discardable search dataset to the number of preliminary search datasets, denoted as , and The discard ratio threshold is compared with the preset discard ratio threshold, wherein the discard ratio threshold includes the discard upper threshold and the discard lower threshold; when >Discard the upper threshold or When the lower threshold is discarded, an adjustment instruction is issued and the adjustment strategy is executed; otherwise, no response is made; When receiving the adjustment instruction, the adjustment strategy is executed to calculate the average number of subsets of all retrieved data sets, which is recorded as , when the average number of subsets >Discard upper threshold× , an upward command is issued to × As the adjusted discard upper threshold; When the average number of subsets <Discard lower threshold× , a downward adjustment command is issued to ×The discard lower threshold is used as the adjusted discard lower threshold; otherwise, no response is made.

[0011] Furthermore, the analysis process of the preferred data set ratio is as follows: Retrieve the average reading volume, current reading volume and current total reading time under the previous n search instructions from the database; The average reading volume is used as the number of data selected from the preliminary search data set, and the selected preliminary search data set is recorded as the preferred data set. When the total number of preliminary search data sets is less than or equal to the average reading volume, an early output instruction is issued, and all preliminary search data sets are output as search results; otherwise, the proportion of the preferred data set in all preliminary search data sets is compared with a preset screening ratio threshold; When an advance output instruction is received, if the current reading volume ≥ the average reading volume and the current total reading time ≤ the preset reading time, a call instruction is issued to retrieve the discardable retrieval data set as an unconventional retrieval data set output; otherwise, a residence time distribution map of each retrieval data set under the current reading volume is generated, and the retrieval data set that does not reach the preset standard time is used as a screening criterion, and the retrieval data sets with the same classification as the screening criterion are removed from the unread retrieval data sets.

[0012] Furthermore, the analysis process of the estimated impact index is as follows: The number of discarded search data sets and the number of discarded search data sets converted into preliminary search data sets in the process of retrieving the first L searches are recorded as , ; By analyzing the formula Get the average conversion rate , where Indicates the number of retrieval data sets that can be discarded during the mth retrieval. Indicates the number of retrieval data sets that can be discarded and converted into preliminary retrieval data sets during the mth retrieval; The number of scheduled preliminary retrieval data sets and discardable retrieval data sets and their subsets are respectively denoted as , , , ; By analyzing the formula Get the estimated impact index , where Indicates the time required for data verification. Indicates the total number of preliminary retrieval data sets, Indicates the total number of datasets that can be discarded for retrieval. represents the i-th preliminary retrieval dataset, represents the number of subsets in the i-th preliminary retrieval data set, represents the jth discardable retrieval data set, represents the number of subsets in the jth discardable retrieval data set, Indicates the speed of recognition of the number of unit data sets, Represents the correction factor of the preset estimated impact index, Indicates the total amount of data in the dataset. Indicates the actual time required to execute the same check instruction last time. Indicates the estimated impact index of the last execution of the same check instruction. Indicates the time difference between the actual time required to execute the same verification instruction last time and the time required for data verification. They represent the weights of the preset data verification time, the total amount of data in the data set, and the actual time required for the last verification instruction, respectively. , e represents a natural constant; When the impact index is estimated >When the preset impact risk threshold is reached, a data output instruction is issued to output the preliminary search data set as the search result in the sorted order; otherwise, a qualified instruction is issued to start executing the verification instruction.

[0013] Furthermore, the historical search trend indicates the number of times various classification standards are selected under the same search term. The analysis process of the data classification evaluation value of the secondary classification is as follows: Retrieve the database storage classification standard to perform secondary classification on the preferred data set and obtain the classified classification set, where the preferred data set is a subset of the classification set, and obtain the two classification sets with the largest and smallest number of subsets. The number of subsets in the two classification sets is recorded as , ; when If the uniformity is greater than or equal to the preset uniformity threshold, a classification failure instruction is issued and a secondary classification is performed again; otherwise, a classification qualification instruction is issued and the classification standard corresponding to the secondary classification is recorded; When receiving the qualified classification instruction, the number of all classification sets and corresponding subsets after secondary classification is obtained, which are recorded as , , and retrieve the two classification sets with the largest and smallest number of subsets based on the classification standard, where the number of subsets is recorded as , ; By analyzing the formula: Get the data classification evaluation value of the secondary classification , where represents the total number of classification sets after secondary classification, represents the mth classification set, represents the number of subsets corresponding to the mth classification set, represents the number of subsets in the mth classification set, They represent the time required for the preset secondary classification and the weight of the uniformity of the classification set. ; By analyzing the formula Get the scoring coefficient of the classification standard , where represents the number of times the y-th classification standard is selected, represents the data classification evaluation value of the y-th classification standard, Indicates the preset number of selections and the weight of the data classification evaluation value, , according to the scoring coefficient of the classification standard The corresponding classification criteria are selected in order from large to small and issued as optimization instructions.

[0014] Furthermore, the formula for the model optimization efficiency evaluation coefficient is as follows: In the formula, represents the model optimization efficiency evaluation coefficient, Y represents the total number of preferred data sets before the optimization instruction, represents the kth preferred data set before the optimization instruction, represents the number of subsets corresponding to the kth preferred data set before the optimization instruction, and X represents the total number of preferred data sets after the optimization instruction. Indicates the xth preferred data set after the optimization instruction. Indicates the number of subsets corresponding to the xth preferred data set before the optimization instruction. They respectively represent the proportion of the preset preferred data set in the preliminary search data set, the estimated impact index, and the weights of the ratio of the number of subsets of the preferred data set before and after instruction optimization. , and when the total number of preliminary search data sets is ≤ the average number of views, Recorded as 1.

[0015] Furthermore, the correction strategy is activated when the model optimization efficiency evaluation coefficient is lower than the optimization effect threshold. The corresponding classification standards are selected in order from large to small according to the scoring coefficients, and the corresponding model optimization efficiency evaluation coefficients are calculated respectively. Then, they are compared with the optimization effect threshold. The classification standard corresponding to the model optimization efficiency evaluation coefficient being lower than the preset optimization effect threshold is selected as the optimization solution output.

[0016] Furthermore, a large language model optimization method based on low-rank adaptive technology includes the following steps: The cleaned and preprocessed input data is used as the data set for training the large language model. The large language model is trained, and then the initial model is low-rank decomposed and optimized by low-rank adaptive technology. The low-rank decomposition is to decompose the original high-dimensional parameter matrix into the product of two or more low-rank matrices. Extract keywords in the received initial instruction as search terms, obtain search data that can be called by the large language model, and divide them into corresponding search data sets according to the type. Then, divide them into preliminary search data sets and discardable search data sets according to the number of subsets in the search data sets, and discard the discardable search data sets; When receiving the preliminary search data set, the subsets in the preliminary search data set are sorted from large to small according to the number of subsets, the first several preliminary search data sets are selected as preferred data sets, and the proportion of the preferred data sets in all preliminary search data sets is compared with the preset screening ratio threshold. When the proportion is less than the screening ratio threshold, the preliminary search data sets are output as search results in the sorted order; otherwise, the comparison data is output as the first search result, and the estimated impact index of the execution of the verification instruction is analyzed, and the verification instruction is issued; When receiving the verification instruction, the preferred data set is secondary classified according to the classification standard stored in the database, and the data classification evaluation value of the secondary classification is analyzed, and the corresponding classification standard is selected as the optimization instruction output in combination with the historical search trend. After that, the instruction after the feedback of the optimization instruction is recorded as the secondary instruction, and the search data after the secondary classification is screened and output; Obtain the proportion of the preferred data set in the preliminary search data set during the retrieval process, the estimated impact index, and the number of subsets of the preferred data set before and after instruction optimization, generate a model optimization efficiency evaluation coefficient, and when the model optimization efficiency evaluation coefficient is lower than the preset optimization effect threshold, call the optimization solution correction strategy.

[0017] (III) Beneficial effects The present invention provides a large language model optimization system and method based on low-rank adaptive technology, which has the following beneficial effects: The large language model optimized by low-rank adaptive technology is used as the basis for data processing, which facilitates subsequent optimization and can effectively reduce the base amount of data to be processed. When executing operation instructions, some data is discarded according to the proportion of the search content in the directly obtained search data set, thereby reducing the basic amount of subsequent search data and improving the efficiency of data retrieval. In addition, according to the proportion of the preferred data set in all preliminary search data sets, "questions" are sent to the "input end" to optimize the operation instructions and avoid "internal friction" of the system. Moreover, through this "questioning" method, the purpose of the operation instruction can be determined more accurately, so as to give accurate feedback and minimize the total amount of data to be processed, so as to fundamentally solve the problem of too much data to be processed.

[0018] Before asking a "question", perform the following two steps: first, analyze the verification instructions to be performed to obtain an estimated impact index that can predict the difficulty of executing the verification instructions, so as to avoid those tedious and complicated verification instructions in advance and improve the efficiency of data analysis; second, use the classification standards stored in the database as the secondary classification standards for the preferred data set, divide the preferred data set into different categories, and then calculate and generate classification evaluation values ​​based on the classification results to determine whether the classification standards are appropriate and to what extent, so as to objectively evaluate the classification standards, combine them with historical retrieval trends, and select classification standards that meet the requirements as optimized instruction outputs, providing options for instruction feedback to reduce feedback time.

[0019] By using the proportion of the preferred data set in the preliminary retrieval data set in the process of asking "questions" to the "input end", the estimated impact index, and the number of subsets of the preferred data set before and after instruction optimization, we can obtain the model optimization efficiency evaluation coefficient for intuitively evaluating the efficiency improvement of the retrieval optimization method, so as to have a clear understanding of the entire optimization plan, thereby providing a basis for orderly improvement and adjustment, and then obtaining an optimization plan that meets the optimization needs through correction of the correction strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is the overall structure flow chart of the present invention; Figure 2 It is a schematic diagram of the model evaluation process of the present invention; Figure 3 This is a distribution diagram of the residence time of each search data set under the current reading volume in the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Low-Rank Adaptation (LoRA) is a technique for fine-tuning large language models (LLMs) that aims to solve the problem of huge computational resources and time consumption when fine-tuning such models. LoRA technology freezes the weights of the pre-trained model and injects a trainable rank decomposition matrix into each Transformer block, significantly reducing the number of parameters that need to be fine-tuned, thereby reducing computational costs and improving training efficiency.

[0023] With the widespread application of large language models (LLMs), reasoning efficiency and resource consumption have become the main bottlenecks limiting their expansion in practical applications, especially in complex tasks such as text and image prompt word generation, where optimizing memory usage and reasoning speed becomes particularly important. However, research has found that existing large language models spend a lot of time on reasoning and analysis, and the results obtained vary greatly from the needs of users. For ease of understanding, the time and proportion obtained during the reasoning ability test are expressed in integers (rounded off).

[0024] Reasoning ability test: Randomly issue search instructions to the large language model, and the interval between two adjacent instructions needs to be controlled within 10-15s. Then record the time required for the response and the total time required to complete the search, and record the proportion of the required content in the feedback results of the large language model. The result is that the response time is generally between 3-8s, and there are obvious differences in the same content due to the different number of keywords or content in the search instructions. For the convenience of understanding, the following uses conventional text retrieval as an example. For example, Guangzhou is on the 18th. If the query instruction "Get the historical temperature in Guangzhou" is used, the result obtained is the temperature change value from the 1st to the 17th in Guangzhou, and the response time is 3s (including the response and recognition time of the device); if the query instruction "Get the temperature in Guangzhou from the 13th to the 17th", the response time is only 1s.

[0025] After analysis, it is known that this is because when the large language model is performing reasoning and analysis, due to inaccurate instructions received, the time required for autonomous reasoning is different. In the process of autonomous reasoning, in order to ensure that the information is what the user needs, it is necessary to combine the user's search situation, coordinates, etc. to obtain comprehensive information. This increases the data base of reasoning, seriously affecting its response efficiency, and there is a large amount of unnecessary content in the obtained data. The reason for the existence of these unnecessary contents is that the model cannot analyze and reason whether these contents are what the user needs. This also reflects the inaccuracy of the input instructions from the side, resulting in complicated data and easy to affect the user's search.

[0026] The existing optimization methods still focus on how to improve the data processing speed, such as dividing the required operation data into multiple independent branch operation data, and then using multiple operation devices to perform synchronous operations. However, they ignore the optimization of how to improve the accuracy of the acquired instructions.

[0027] Summary: The large language model optimized by low-rank adaptive technology is used as the basis for data processing. Then, according to the proportion of the preferred data set in all preliminary search data sets, "questions" are asked to the "input end" to optimize the operation instructions and avoid "internal friction" of the system. Moreover, through this "questioning" method, the purpose of the operation instruction can be determined more accurately, so as to give accurate feedback and minimize the total amount of data to be processed, so as to fundamentally solve the problem of too much data to be processed.

[0028] Research and development concept: Since the initial instructions for reasoning and analysis by the large language model are issued by the user, they are practical and uncontrollable, so it is necessary to obtain the user's accurate needs. For this reason, we use the form of questions and answers to refine the instructions and analyze the preliminary instructions in the document process. For example, when it is necessary to retrieve the "working principle of air conditioning", the data obtained by the large language model is mainly divided into two types: "air conditioning refrigeration principle" and "air conditioning heating principle". Before the content is generated, a verification instruction can be issued and issued in the form of an inquiry. For example, whether you need "air conditioning refrigeration principle" or "air conditioning heating principle", the feedback instruction obtained will be executed as a secondary instruction. In this way, no matter whether you choose "air conditioning refrigeration principle" or "air conditioning heating principle", the first instruction issued can be optimized. In addition, the two materials can be prepared at the same time during the interval between inquiries to facilitate data extraction.

[0029] At the same time, in order to improve efficiency, when the ratio of the search "air conditioning refrigeration principle" and "air conditioning heating principle" exceeds the set threshold, the result with a larger proportion will be output; otherwise, a check will be performed.

[0030] Embodiment 1:

[0031] Basic data optimization The data is cleaned and preprocessed, and used as a data set for training large data models to train large language models. The initial model is then decomposed and optimized using low-rank adaptive technology. Low-rank decomposition is to decompose the original high-dimensional parameter matrix into the product of two or more low-rank matrices, that is, to decompose a large matrix into the product of two or more smaller, simpler matrices, and these small matrices usually have a lower rank. This decomposition method helps reduce the number of parameters, computational complexity and memory usage, and is often used for model compression and acceleration.

[0032] Among them, in the traditional Adapter, the number of trainable parameters of the model is reduced by inserting relatively small trainable parameters between each layer. However, this strategy has a significant flaw, that is, there is a lack of direct gradient path during the back propagation process, and the gradient needs to pass through all training layers. Therefore, this structure also faces certain challenges in parallel computing. Similarly, Prompt Tuning also faces a similar problem, that is, the lack of direct gradient path, resulting in the gradient passing through all training layers during the back propagation process. In addition, the task setting of Prompt Tuning is relatively idealized, attempting to achieve model optimization by only adjusting a small number of parameters at the input end. This method has limited impact on the deep part of the model, thereby limiting the effect of fine-tuning.

[0033] To address the above issues, the improved method of low-rank coding optimizes the Transformer structure. This method first passes the input data through all training layers, then combines the output of the training layer with the input adjusted by the new parameters, and finally generates a new adjusted output. In this way, the Transformer structure optimized by low-rank coding only needs to be tuned for the new parameters without loading the original training parameters. The improved low-rank coding transfers the step parameter memory based on LoRA (low-rank adaptation), which can effectively learn and integrate the reduced rank with the target data more fully.

[0034] Execution of preliminary instructions "Preliminary instruction" means the instruction issued for the first time. After receiving the initial instruction, the keywords are extracted as search terms, and the search is performed in the search data that can be called by the large language model. Then, the search data is divided into corresponding search data sets according to the type. The classification type is the secondary classification type obtained by the previous search for the same search term (the secondary classification will be introduced later). Then, according to the number of subsets in the search data set, it is divided into preliminary search data sets and discardable search data sets, and the discardable search data sets are discarded. Among them, the subset of the search data set refers to the search data that can be directly obtained according to the search term.

[0035] When classifying the retrieval data set, it is obtained by comparing the number of subsets in the retrieval data set with the preset discard threshold. When the number of subsets in the retrieval data set is less than the preset discard threshold, it is recorded as a discardable retrieval data set; otherwise, it is recorded as a preliminary retrieval data set. The reason for adopting this method is that for some very small amounts of data, the recorded content is limited and not universal. Suppose the content to be retrieved is: "Power Station Address", and there are 10 pieces of data related to Power Station 1, 15 pieces of data related to Power Station 2, and 2 pieces of data related to Power Station 3. It can be seen that for Power Station 3, there is very little relevant information (indicating that it is less concerned or less studied), and it is not concerned. Therefore, this type of data can be discarded to reduce the overall data retrieval volume.

[0036] Although these data have been discarded as unimportant data, they still need to be recycled in order to avoid special needs. For example, when searching according to the secondary instruction, if the number of subsets in the discardable search data set increases, it means that the secondary instruction is more inclined to the discarded part in the first classification. When the ratio of the increase in number to the number of atomic sets is greater than 20%, a reuse instruction is issued to record the corresponding discardable search data set as a preliminary search data set, and recall the content discarded in the first classification to meet the needs of unconventional data; on the contrary, when the ratio of the increase in number to the number of atomic sets is ≤20%, the corresponding discardable search data set is recorded as a non-search data set under the search term, and is not displayed in the callable search data retrieved by the same search term, thereby reducing the amount of basic data calculated.

[0037] However, it is found that different retrieval instructions obtain different amounts of data, so if a specified size of the discard threshold is used as the screening basis, errors are very likely to occur. For this reason, we propose an adjustment plan for the discard threshold to show that it is applicable to changing data. The steps for adjusting the discard threshold are as follows: Step 1: Set a standard adjustment ratio , to limit the adjustment range of the discard threshold; The second step is to obtain preliminary instructions and obtain the number of discardable search data sets based on the search terms in the instructions. and the number of preliminary search datasets, calculate the ratio of the discardable search dataset to the number of preliminary search datasets, denoted as , assuming that the total number of data that can be retrieved (that is, all the data that can be retrieved with the preliminary instructions) is 100 data, and after comparing with the discard threshold, the number of discarded retrieval data sets is 20, and the number of preliminary retrieval data sets is 100-20=80. Then the ratio of the discarded retrieval data sets to the number of preliminary retrieval data sets is ; The third step is to The discard ratio threshold is compared with the preset discard ratio threshold, and the discard ratio threshold is divided into the discard upper threshold and the discard lower threshold; when > When the upper threshold is discarded, it means that the discarded data accounts for a very large proportion of all the retrieved data. If all are discarded, the basic data that can be provided will be very small. On the contrary, when When the discarded data is less than the lower threshold, it means that the discarded data accounts for a small proportion, and there is still a lot of basic data to be analyzed. > Abandon the upper threshold or When the lower threshold is less than the discarding threshold, adjustments are required. At this time, adjustment instructions are issued and adjustment strategies are executed. When the value is less than or equal to the upper threshold value, no adjustment is required; Step 4: Adjust the strategy: Calculate the average number of subsets of all retrieved data sets, denoted as ; Upward adjustment: When the average number of subsets >Discard upper threshold× , an upward command is issued to ×The upper discard threshold is used as the adjusted upper discard threshold; Downward adjustment: When the average number of subsets <Discard lower threshold× , a downward adjustment command is issued to ×The discard lower limit threshold is used as the adjusted discard lower limit threshold; On the contrary, when discarding the upper threshold × ≥Average number of subsets ≥Discard lower threshold × , no response is given.

[0038] Data volume comparison The preliminary search data set obtained after the preliminary instruction analysis is used as the basic data for analysis. The preliminary search data sets are first sorted in descending order according to the number of internal subsets, and the first several preliminary search data sets are selected as the preferred data sets. The selection of the preferred data set is based on the average reading volume under the previous n same search instructions. Assuming that only 10 data items can be read on average each time a search is performed, 10 items are the screening criteria. There are 30 preliminary search data sets in total. They are sorted according to the number of internal data items, and the first 10 preliminary search data sets are selected. These 10 preliminary search data sets are the preferred data sets.

[0039] However, since the number of preliminary search data sets is not fixed and may vary, when the total number of preliminary search data sets is less than or equal to the average number of views, an early output instruction is issued to output all preliminary search data sets as search results; otherwise, the proportion of the preferred data set in all preliminary search data sets is compared with the preset screening ratio threshold. The comparison results are as follows: When the proportion is less than the screening ratio threshold, it means that the preliminary search data set after the preliminary instruction analysis is not enough to be selected, so the preliminary search data set can be directly output as the search result in the sorted order; on the contrary, when the proportion is greater than or equal to the screening ratio threshold, it means that the preliminary search data set after the preliminary instruction analysis is very large and needs to be refined, but in order to avoid data generation delays, the comparison data is first output as the first search result, so that there is no need to wait until the verification instruction is executed before the result is obtained.

[0040] When receiving the advance output instruction, if the current reading volume ≥ the average reading volume and the current total reading time ≤ the preset reading time, a call instruction is issued to retrieve the discardable search data set as the unconventional search data set output; otherwise, a residence time distribution diagram of each search data set under the current reading volume is generated, and the search data set that does not reach the preset standard time is used as the screening criterion, and the search data sets with the same classification as the screening criterion are removed from the unread search data sets. Assume that Figure 3 is a distribution graph, in which the Y axis is time (stay time) and the X axis is number (number of each search data set). Figure 3 It can be seen that the preview time mainly stays at number one and number four, while the time staying at number three is relatively short, about 15-20 seconds. At this time, if the standard duration is 20 seconds, and the time of numbers three and four in the figure is less than 20 seconds, then it is necessary to use the data of the same type as numbers three and four as the screening criteria and delete the remaining data that has not been previewed.

[0041] The above average reading volume, current reading volume and current total reading time are all obtained by retrieving the data obtained under the same search instructions for the previous n times from the database. The average reading volume is obtained by adding the reading volumes obtained for the previous n times and dividing it by n; the current reading volume indicates the number of preferred data sets that have been read; the current total reading time indicates the total time spent on reading the preferred data sets. These data are all retrieved from the database.

[0042] When a check instruction is issued, the execution of the check instruction is analyzed to obtain an estimated impact index that can predict the difficulty of executing the check instruction, so as to facilitate avoiding those cumbersome and complex check instructions in advance and improve the efficiency of data analysis. The steps for choosing whether to execute the check instruction according to needs are as follows: First retrieve the previous L identical terms from the database, and discard the number of search data sets during the search process. And the number of discarded search data sets converted into preliminary search data sets .

[0043] (1.1) In formula (1.1), It represents the average conversion ratio of the discardable retrieval dataset to the preliminary retrieval dataset. Indicates the number of retrieval data sets that can be discarded during the mth retrieval. It indicates the number of retrieval data sets that can be discarded and converted into preliminary retrieval data sets during the mth retrieval.

[0044] (1.2) (1.3) In formulas (1.2) and (1.3), Indicates the estimated impact index of executing the check instruction. Indicates the time required for data verification. Indicates the total number of preliminary retrieval data sets, Indicates the total number of datasets that can be discarded for retrieval. Indicates the number of preliminary retrieval data sets, represents the number of subsets in the initial retrieval dataset, represents the i-th preliminary retrieval dataset, represents the number of subsets in the i-th preliminary retrieval data set, Indicates the number of datasets that can be discarded for retrieval. Indicates the number of subsets in the data set that can be discarded for retrieval. represents the jth discardable retrieval data set, represents the number of subsets in the jth discardable retrieval data set, Indicates the speed of recognition of the number of unit data sets, Represents the correction factor of the preset estimated impact index, Indicates the total amount of data in the dataset. Indicates the actual time required to execute the same check instruction last time. Indicates the estimated impact index of the last execution of the same check instruction. Indicates the time difference between the actual time required to execute the same verification instruction last time and the time required for data verification. They represent the weights of the preset data verification time, the total amount of data in the data set, and the actual time required for the last verification instruction, respectively. , e represents a natural constant.

[0045] When the impact index is estimated >When the preset impact risk threshold is exceeded, it means that under the current classification standard, the data and time required for calculation are very large. If executed, it will take a lot of time and energy, which will seriously affect the efficiency of system execution. Therefore, a data output instruction is issued at this time, and the preliminary search data set is output as the search result in the sorted order, and the result is issued directly.

[0046] When the impact index is estimated ≤ the preset impact risk threshold, a qualified instruction is issued and the verification instruction is executed. The current classification standard is combined with the historical retrieval trend to select the corresponding classification standard as the optimization instruction output. After that, the instruction after the optimization instruction feedback is recorded as the secondary instruction, and the retrieval data after the secondary classification is screened and output.

[0047] Instruction Optimization This is also the focus of this application, which is mainly to issue "questions" to the "input end" according to the proportion of the preferred data set in all preliminary search data sets, so as to optimize the operation instructions and avoid the "internal friction" of the system. Moreover, through this "questioning" method, the purpose of the operation instruction can be determined more accurately, so as to give accurate feedback and minimize the total amount of data to be processed, so as to fundamentally solve the problem of too much data to be processed. In this way, some unclear instructions are refined to reduce the time for data analysis and data derivation.

[0048] The execution instruction optimization is turned on when the verification instruction is received. After it is turned on, the preferred data set is secondary classified according to the classification standards stored in the database (such as alphabetical order of the name, type, upload time, data volume, etc.), and the data classification evaluation value of the secondary classification is analyzed. The data classification evaluation value is calculated based on the classification results obtained after the secondary classification to generate a classification evaluation value that can determine whether the classification standard is appropriate and the degree of appropriateness, so as to objectively evaluate the classification standard. By combining it with the historical retrieval trend, the classification standard that meets the requirements can be selected as the optimization instruction output, providing options for instruction feedback to reduce the feedback time. The formula based on the classification evaluation value is as follows: (2.1) In formula (2.1), Indicates the data classification evaluation value for secondary classification under the current classification standard. represents the total number of classification sets after secondary classification, represents the number of classification sets, Indicates the number of subsets corresponding to the classification set, represents the mth classification set, represents the number of subsets corresponding to the mth classification set, represents the number of subsets in the mth classification set, Indicates the maximum number of subsets in all classification sets under the current classification standard. Indicates the minimum number of subsets in all classification sets under the current classification standard. They represent the time required for the preset secondary classification and the weight of the uniformity of the classification concentration. .

[0049] The database storage classification standard is retrieved to perform secondary classification on the preferred data set, and a classified classification set is obtained, wherein the preferred data set is a subset of the classification set, and two classification sets with the largest and smallest subset numbers are obtained. The numbers of subsets in the two classification sets are recorded as , ;when If the uniformity is greater than or equal to the preset uniformity threshold, a classification failure instruction is issued and a secondary classification is performed again; otherwise, a classification qualification instruction is issued and the classification standard corresponding to the secondary classification is recorded.

[0050] The scoring coefficients for the classification criteria are based on the following formula: (2.2) In formula (2.2), Indicates the scoring coefficient under the current classification standard, represents the number of times the y-th classification standard is selected, represents the data classification evaluation value of the y-th classification standard, Indicates the preset number of selections and the weight of the data classification evaluation value, , according to the scoring coefficient of the classification standard The corresponding classification criteria are selected in order from large to small and issued as optimization instructions.

[0051] Analysis of instruction optimization effects By using the proportion of the preferred data set in the preliminary retrieval data set in the process of asking "questions" to the "input end", the estimated impact index, and the number of subsets of the preferred data set before and after instruction optimization, we can obtain the model optimization efficiency evaluation coefficient for intuitively evaluating the efficiency improvement of the retrieval optimization method, so as to have a clear understanding of the entire optimization plan, thereby providing a basis for orderly improvement and adjustment, and then obtaining an optimization plan that meets the optimization needs through correction of the correction strategy.

[0052] In the formula, represents the model optimization efficiency evaluation coefficient, Y represents the total number of preferred data sets before the optimization instruction, represents the kth preferred data set before the optimization instruction, represents the number of subsets corresponding to the kth preferred data set before the optimization instruction, and X represents the total number of preferred data sets after the optimization instruction. Indicates the xth preferred data set after the optimization instruction. Indicates the number of subsets corresponding to the xth preferred data set before the optimization instruction. They respectively represent the proportion of the preset preferred data set in the preliminary search data set, the estimated impact index, and the weights of the ratio of the number of subsets of the preferred data set before and after instruction optimization. , and when the total number of preliminary search data sets is ≤ the average number of views, Recorded as 1. The historical search trend indicates the number of times various classification standards are selected under the same search term.

[0053] Correction strategy: It is turned on when the model optimization efficiency evaluation coefficient is lower than the optimization effect threshold. The corresponding classification standards are selected in order from large to small according to the scoring coefficients, and the corresponding model optimization efficiency evaluation coefficients are calculated respectively. Then, they are compared with the optimization effect threshold. The classification standard corresponding to the model optimization efficiency evaluation coefficient when it is lower than the preset optimization effect threshold is selected as the optimization solution output.

[0054] The weight coefficient is determined by the coefficient of variation method, which is a method of weighting each indicator according to the degree of variation between the current value of each evaluation indicator and the target value. If the numerical difference of an indicator is large, it can clearly distinguish the evaluated objects, indicating that the indicator has rich discrimination information, and thus the indicator should be given a larger weight. On the contrary, if the numerical difference of each evaluated object on a certain indicator is small, then the ability of this indicator to distinguish the evaluation objects is weak, and thus the indicator should be given a smaller weight. This method directly uses the information contained in each indicator to obtain the weight of the indicator through calculation, so it is objective.

[0055] Embodiment 2: Based on Embodiment 1, this embodiment further provides a large language model optimization method based on low-rank adaptive technology, and the optimization method includes the following steps: The cleaned and preprocessed input data is used as the data set for training the large language model. The large language model is trained, and then the initial model is low-rank decomposed and optimized by low-rank adaptive technology. The low-rank decomposition is to decompose the original high-dimensional parameter matrix into the product of two or more low-rank matrices. Extract keywords in the received initial instruction as search terms, obtain search data that can be called by the large language model, and divide them into corresponding search data sets according to the type. Then, divide them into preliminary search data sets and discardable search data sets according to the number of subsets in the search data sets, and discard the discardable search data sets; When receiving the preliminary search data set, the subsets in the preliminary search data set are sorted from large to small according to the number of subsets, the first several preliminary search data sets are selected as preferred data sets, and the proportion of the preferred data sets in all preliminary search data sets is compared with the preset screening ratio threshold. When the proportion is less than the screening ratio threshold, the preliminary search data sets are output as search results in the sorted order; otherwise, the comparison data is output as the first search result, and the estimated impact index of the execution of the verification instruction is analyzed, and the verification instruction is issued; When receiving the verification instruction, the preferred data set is secondary classified according to the classification standard stored in the database, and the data classification evaluation value of the secondary classification is analyzed, and the corresponding classification standard is selected as the optimization instruction output in combination with the historical search trend. After that, the instruction after the feedback of the optimization instruction is recorded as the secondary instruction, and the search data after the secondary classification is screened and output; Obtain the proportion of the preferred data set in the preliminary search data set during the retrieval process, the estimated impact index, and the number of subsets of the preferred data set before and after instruction optimization, generate a model optimization efficiency evaluation coefficient, and when the model optimization efficiency evaluation coefficient is lower than the preset optimization effect threshold, call the optimization solution correction strategy.

[0056] In the application, the several formulas involved are all calculated by taking their numerical values ​​after removing the dimensions, and the formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent actual situation. Some coefficients or weights in the formulas are set by technical personnel in this field according to actual conditions, so they will not be elaborated here.

[0057] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0058] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0059] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A large language model optimization system based on low-rank adaptive technology, characterized in that: The optimization system includes: Instruction preliminary analysis module: extract the initial instruction keywords as search terms, obtain the search data that can be called by the large model, classify them into search data sets, and divide them into preliminary and discardable search data sets according to the number of subsets, and discard the latter; Data comparison module: Receive the preliminary search data set, sort it in descending order by the number of subsets, take the first few as the preferred data set, calculate their proportion in the preliminary search data set and compare it with the screening ratio threshold; and according to the proportion result, take the preliminary search data set or the comparison data as the first search result, analyze and estimate the impact index and issue a verification instruction; Instruction optimization module: after receiving the verification instruction, it will reclassify the preferred data set according to the database classification standard, analyze the evaluation value, and output the optimization instruction in combination with the historical search trend. The instruction after feedback is the secondary instruction, which will filter and output the secondary classified search data; Model evaluation module: obtains the proportion of preferred data sets, estimated impact index, and the number of subsets before and after instruction optimization, generates a model optimization efficiency evaluation coefficient, and calls the optimization solution correction strategy when the coefficient is lower than the optimization effect threshold.

2. The large language model optimization system based on low-rank adaptive technology according to claim 1, characterized in that: Before the execution of the instruction preliminary analysis module, the large language model is trained with the cleaned and preprocessed input data, and the initial model is low-rank decomposed and optimized using low-rank adaptive technology; the input data passes through all training layers, and then the output of the training layer is combined with the input adjusted by the new parameters to generate a new adjusted output.

3. The large language model optimization system based on low-rank adaptive technology according to claim 1, characterized in that: The subset of the search data set means: the search data directly obtained according to the search terms; when classifying the search data set, it is compared with the preset rejection threshold; If the number of subsets in the retrieval data set is less than the preset discarding threshold, it is recorded as a discardable retrieval data set; Otherwise, it is recorded as the preliminary search data set; Among them, when searching according to the secondary instruction, if the number of subsets in the discardable search data set increases, and the ratio of the increase to the number of atomic sets is greater than 20%, a reuse instruction is issued to record the corresponding discardable search data set as the preliminary search data set; otherwise, the corresponding discardable search data set is recorded as the non-search data set under the search term.

4. The large language model optimization system based on low-rank adaptive technology according to claim 3, characterized in that: The discard threshold is adjusted according to the following process: Set the adjustment ratio ; Get the number of discardable search datasets under the current search term and the number of preliminary search datasets, and calculate the ratio of the discardable search datasets to the number of preliminary search datasets , and Compare with the preset discard ratio threshold interval; The discard ratio threshold includes: a discard upper limit threshold and a discard lower limit threshold; when >Discard the upper threshold or When the lower threshold is lower than the discard threshold, an adjustment instruction is issued and the adjustment strategy is executed to calculate the average number of subsets of all retrieved data sets. ; When the average number of subsets >Discard upper threshold× , an upward command is issued to ×The upper discard threshold is used as the adjusted upper discard threshold; when the average number of subsets <Discard lower threshold× , a downward adjustment command is issued to ×The discard lower limit threshold is used as the adjusted discard lower limit threshold; Otherwise, no response will be given.

5. The large language model optimization system based on low-rank adaptive technology according to claim 4, characterized in that: The analysis process of the preferred data set ratio is as follows: Retrieve the average reading volume, current reading volume and current total reading time under the previous n search instructions from the database; The average reading volume is used as the number of data selected from the preliminary search data set, and the selected preliminary search data set is recorded as the preferred data set. When the total number of preliminary search data sets is less than or equal to the average reading volume, an early output instruction is issued, and all preliminary search data sets are output as search results; otherwise, the proportion of the preferred data set in all preliminary search data sets is compared with a preset screening ratio threshold; When an advance output instruction is received, if the current reading volume ≥ the average reading volume and the current total reading time ≤ the preset reading time, a call instruction is issued to retrieve the discardable retrieval data set as an unconventional retrieval data set output; otherwise, a residence time distribution map of each retrieval data set under the current reading volume is generated, and the retrieval data set that does not reach the preset standard time is used as a screening criterion, and the retrieval data sets with the same classification as the screening criterion are removed from the unread retrieval data sets.

6. The large language model optimization system based on low-rank adaptive technology according to claim 5, characterized in that: The analysis process for estimating the impact index is as follows: In the process of retrieving the first L retrievals, the number of discarded retrieval data sets and the number of discarded retrieval data sets converted into preliminary retrieval data sets are recorded as , ; The average conversion ratio is calculated by ; The number of scheduled preliminary retrieval data sets and discardable retrieval data sets and their subsets are respectively denoted as , , , ; Through analysis and calculation, the estimated impact index is obtained ; When the impact index is estimated >When the preset impact risk threshold is reached, a data output instruction is issued to output the preliminary search data set as the search result in the sorted order; otherwise, a qualified instruction is issued to start executing the verification instruction.

7. The large language model optimization system based on low-rank adaptive technology according to claim 6, characterized in that: Historical search trends indicate: the number of times various classification standards are selected under the same search term; The analysis process of the data classification evaluation value of the secondary classification is as follows: Retrieve the database storage classification standard to perform secondary classification on the preferred data set and obtain the classified classification set, where the preferred data set is a subset of the classification set, and obtain the two classification sets with the largest and smallest number of subsets. The number of subsets in the two classification sets is recorded as , ; when If the uniformity is greater than or equal to the preset uniformity threshold, a classification failure instruction is issued and a secondary classification is performed again; otherwise, a classification qualification instruction is issued and the classification standard corresponding to the secondary classification is recorded; When receiving the qualified classification instruction, the number of all classification sets and corresponding subsets after secondary classification is obtained, which are recorded as , , and retrieve the two classification sets with the largest and smallest number of subsets based on the classification standard, where the number of subsets is recorded as , ; By analyzing the formula: Get the data classification evaluation value of the secondary classification , where represents the total number of classification sets after secondary classification, represents the mth classification set, represents the number of subsets corresponding to the mth classification set, represents the number of subsets in the mth classification set, All are weights; The scoring coefficient of the classification standard is obtained by calculation ; Scoring coefficients based on classification criteria The corresponding classification criteria are selected in order from large to small and issued as optimization instructions.

8. The large language model optimization system based on low-rank adaptive technology according to claim 7, characterized in that: The formula for evaluating the model optimization efficiency coefficient is as follows: In the formula, represents the model optimization efficiency evaluation coefficient, Y represents the total number of preferred data sets before the optimization instruction, represents the kth preferred data set before the optimization instruction, represents the number of subsets corresponding to the kth preferred data set before the optimization instruction, and X represents the total number of preferred data sets after the optimization instruction. Indicates the xth preferred data set after the optimization instruction. Indicates the number of subsets corresponding to the xth preferred data set before the optimization instruction. Respectively represent weights; When the total number of preliminary search data sets is ≤ the average number of views, Recorded as 1.

9. The large language model optimization system based on low-rank adaptive technology according to claim 1, characterized in that: Correction strategy: It is turned on when the model optimization efficiency evaluation coefficient is lower than the optimization effect threshold. The corresponding classification standards are selected in order from large to small according to the scoring coefficients, and the corresponding model optimization efficiency evaluation coefficients are calculated respectively. Then, they are compared with the optimization effect threshold. The classification standard corresponding to the model optimization efficiency evaluation coefficient when it is lower than the preset optimization effect threshold is selected as the optimization solution output.

10. A large language model optimization method based on low-rank adaptive technology, using any system described in claims 1 to 9, characterized in that: The optimization method comprises the following steps: The cleaned and preprocessed input data is used as the data set for training the large language model. The large language model is trained, and then the initial model is low-rank decomposed and optimized by low-rank adaptive technology. The low-rank decomposition is to decompose the original high-dimensional parameter matrix into the product of two or more low-rank matrices. Extract keywords in the received initial instruction as search terms, obtain search data that can be called by the large language model, and divide them into corresponding search data sets according to the type. Then, divide them into preliminary search data sets and discardable search data sets according to the number of subsets in the search data sets, and discard the discardable search data sets; When receiving the preliminary search data set, the subsets in the preliminary search data set are sorted from large to small according to the number of subsets, the first several preliminary search data sets are selected as preferred data sets, and the proportion of the preferred data sets in all preliminary search data sets is compared with the preset screening ratio threshold. When the proportion is less than the screening ratio threshold, the preliminary search data sets are output as search results in the sorted order; otherwise, the comparison data is output as the first search result, and the estimated impact index of the execution of the verification instruction is analyzed, and the verification instruction is issued; When receiving the verification instruction, the preferred data set is secondary classified according to the classification standard stored in the database, and the data classification evaluation value of the secondary classification is analyzed, and the corresponding classification standard is selected as the optimization instruction output in combination with the historical search trend. After that, the instruction after the feedback of the optimization instruction is recorded as the secondary instruction, and the search data after the secondary classification is screened and output; Obtain the proportion of the preferred data set in the preliminary search data set during the retrieval process, the estimated impact index, and the number of subsets of the preferred data set before and after instruction optimization, generate a model optimization efficiency evaluation coefficient, and when the model optimization efficiency evaluation coefficient is lower than the preset optimization effect threshold, call the optimization solution correction strategy.

Images