Large model calibration method

Abstract

The present application provides a large model calibration method. The method comprises: obtaining a preset original learning data set for a large model; and on the basis of at least one of a preconstructed geometric logic learning module, a preconstructed mathematical logic learning module, and a preconstructed agent instruction learning module, calibrating an algorithm and the original learning data set. Learning data is subjected to multi-dimensional calibration via one or more learning logics, meeting current requirements for output accuracy of large models in the field of artificial intelligence and enabling the trained large models to satisfy iterative learning demands, thus enhancing applicability. In addition, through learning and calibration of the agent instruction learning module and other promote modules, in each large model iteration, the processing logic or input requirements of external agents can be met, and a plurality of external agents can be interfaced to form designated internal and external loops and sequential logical inferences, leveraging the advantages of various agents and external software to support diverse outputs.

Description

A large model correction method

[0001] The present application claims priority to the Chinese patent application No. CN 202410756133.5, filed on June 12, 2024, and titled "A large model correction method", the disclosure of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the field of artificial intelligence, in particular to a large model correction method. BACKGROUND

[0003] With the rapid development of artificial intelligence (AI), large models are widely used in various industries. However, in practical applications, especially for logic fitting or calculation implemented by Agent, Markov chain, etc., large models have the problem of low output accuracy.

[0004] Taking Agent as an example, as a software and hardware system with autonomous ability, social ability, reaction ability and pre-movement ability, Agent can generally be used only in closed artificial intelligence application scenarios (such as playing chess) in specific business fields, that is, Agent is usually used as a closed model.

[0005] However, when using closed models such as Agent to implement iterative optimization of large models, the connected logic may fail due to drift and other problems. Here, as users use it, the large model will incorporate new data and perform feedback iteration, causing model drift or subtle changes in convolution methods to change factor values, which makes it easy for logic fitting in the logic module of the large model, such as basic addition, subtraction, multiplication and division, to make errors.

[0006] Therefore, there is an urgent need for a correction scheme to improve the output accuracy of large models. SUMMARY

[0007] The embodiments of the present application at least provide a large model correction method to improve the accuracy of the output of the large model by learning the correction of the data set and its algorithm.

[0008] In a first aspect, the embodiments of the present application provide a large model correction method, comprising:

[0009] obtaining a preset original learning data set for a large model;

[0010] Based on at least one of the pre-constructed geometric logic learning module, mathematical logic learning module, and Agent instruction learning module, the algorithm of the large model and the original learning data set are corrected respectively to obtain a corrected large model algorithm and a corrected learning data set.

[0011] The corrected large model algorithm and the corrected learning data set are used to conform to the processing logic of the corresponding learning module, and can make the trained large model meet the demand of iterative operation.

[0012] Optionally, the geometric logic learning module is constructed according to the following steps:

[0013] A first factor relationship for representing a geometric logic sequential relationship is obtained; the geometric logic sequential relationship is extracted by the Agent model after verifying the screened and designed learning data set;

[0014] Based on the first factor relationship, a compatible algorithm is extracted in the Agent model algorithm and corresponding learning data is extracted in the data set corresponding to the Agent model.

[0015] Optionally, the algorithm of the large model and the original learning data set are corrected respectively, including:

[0016] The learning data extracted from the data set corresponding to the Agent model is input into the large model for iterative operation to determine whether the output result conforms to the geometric logic sequential relationship in the first factor relationship;

[0017] The learning data extracted from the data set corresponding to the Agent model is mixed with the original learning data set preset by the large model, and is input into the large model for iterative operation to determine whether the geometric logic sequential relationship in the first factor relationship appears;

[0018] If the geometric logic sequential relationship in the first factor relationship is met, a geometric logic Promote relationship label is added to the corresponding target learning data;

[0019] If the geometric logic sequential relationship in the first factor relationship is not met, the learning data in the original learning data set of the large model that does not conform to the first factor relationship is deleted, and / or the addition of the algorithm of the large model is performed.

[0020] Optionally, the mathematical logic learning module is constructed according to the following steps:

[0021] A second factor relationship for representing a mathematical logic sequential relationship is obtained; the mathematical logic sequential relationship is extracted by the Agent after verifying the screened and designed learning data set;

[0022] Extract a compatible algorithm in the Agent model algorithm based on the second factor relationship and extract corresponding learning data in a data set corresponding to the Agent model.

[0023] Optionally, the algorithm of the large model and the original learning data set are respectively corrected, including:

[0024] The learning data extracted from the data set corresponding to the Agent model is input into the large model for iterative calculation to determine whether the output result conforms to the mathematical logic order relationship in the second factor relationship.

[0025] The learning data extracted from the data set corresponding to the Agent model is mixed with the original learning data set preset by the large model, and is input into the large model for iterative calculation to determine whether the mathematical logic order relationship in the first factor relationship appears.

[0026] If the mathematical logic order relationship in the second factor relationship is met, a mathematical logic Promote relationship label is added to the corresponding target learning data.

[0027] If the second factor relationship is not met, the learning data in the original learning data set of the large model that does not conform to the second factor relationship is deleted, and / or the addition of the algorithm of the large model is performed.

[0028] Optionally, the Agent instruction learning module is constructed according to the following steps:

[0029] A third factor relationship for representing an Agent instruction order relationship is obtained; the Agent instruction order relationship is extracted by a large model connected with an external Agent model after the learning data set that meets the requirements of the connected external Agent model is verified after being screened and designed;

[0030] Based on the third factor relationship, a compatible algorithm is extracted in the large model connected with the external Agent model and corresponding learning data is extracted in the data set of the connected large model.

[0031] Optionally, the algorithm of the large model and the original learning data set are respectively corrected, including:

[0032] The learning data extracted from the data set of the large model connected with the external Agent model is input into the large model for iterative calculation to determine whether the output result conforms to the third factor relationship of the Agent instruction learning module.

[0033] If the third factor relationship of the Agent instruction learning module is met, an Agent instruction Promote relationship label is added to the corresponding target learning data.

[0034] If the third factor relationship is not met, the learning data in the original learning data set that does not meet the third factor relationship is deleted, and / or the big model algorithm is added.

[0035] Optionally, when the correction is based on the factor relationship, the method further comprises:

[0036] The compatible algorithm or the specialized algorithm required for adding the factor relationship;

[0037] The order relationship is determined in one or more of the following ways:

[0038] The learning data in the original learning data set is grouped to obtain a plurality of grouped learning data groups; statistical tests are performed on the plurality of learning data groups, and the order relationship is determined based on whether the factor relationship or the factor corresponding value of each group is significantly different and whether the judgment result meets the corresponding geometric logic, mathematical logic, or Agent instruction logic.

[0039] The order relationship is determined based on a target algorithm having a causal inference or a time or order data fitting and testing characteristic, and the target algorithm includes one of a Bayesian algorithm, a Markov chain algorithm, a decision tree regression, a random forest regression, a breakpoint test, and a T-test.

[0040] Optionally, the method further comprises:

[0041] A Promote learning data set and its factor relationship are obtained, and the Promote learning data set is added with a Promote relationship label; the Promote relationship label includes at least one of a geometric logic Promote relationship label, a mathematical logic Promote relationship label, and an Agent instruction Promote relationship label; the Promote relationship label includes a logic Promote that emerges after correction or other logic Promote that is confirmed by an external Agent model;

[0042] It is determined whether the factor relationship of the output data set obtained by performing an algorithm on any Promote learning data in the Promote learning data set coincides with or coexists with the input factor relationship of the Promote learning data itself or other Promote learning data.

[0043] If yes, it is determined that the Promote learning data forms a circular or sequential logical deduction.

[0044] Optionally, the method further comprises:

[0045] The new learning data set and the factor relationship are combined in the determined circular or sequential logical deduction direction to form a new circular or sequential logical deduction direction; or

[0046] According to the new cycle or sequential logical deduction direction, the input sequence of different Agent models is matched to form a large model, and the large model is connected with various Agent models and corresponding cycle or sequential logical deduction is carried out; or,

[0047] By modifying the Promote learning data in whole or in part, the output factor relationship obtained by the calculation of any Promote learning data does not form overlap or coexistence with the input factor relationship of itself or other Promote learning data, so as to break or stop the cycle or sequential logical deduction; or,

[0048] By combining the learning data set, a method of forming or breaking the cycle or sequential logical deduction direction of the large model is formed.

[0049] Using the above large model correction method, when the original learning data set preset for the large model is obtained, the algorithm of the large model and the original learning data set can be corrected based on at least one of the geometric logic learning module, the mathematical logic learning module and the Agent instruction learning module, to obtain the corrected large model algorithm and the corrected learning data set. Since the learning data is corrected in more dimensions by one or more learning logics, the corrected learning data set not only conforms to the processing logic of the learning module, but also makes the trained large model meet the demand of iterative operation, which meets the requirement of the current artificial intelligence field for the accuracy of the large model output, and is more applicable.

[0050] Other advantages of the present application will be described in more detail in conjunction with the following description and drawings.

[0051] It should be understood that the above description is only a summary of the technical solutions of the present application, so as to enable a general understanding of the technical means of the present application, and then the content of the specification is implemented. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced. The drawings herein are incorporated into the specification and form part of the specification, which shows the embodiments consistent with the present application, and is used to illustrate the technical solutions of the present application together with the specification. It should be understood that the drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope of protection. For those skilled in the art, other related drawings can also be obtained from these drawings without creative labor. Moreover, the same reference numerals are used to represent the same parts throughout the drawings. In the drawings:

[0053] FIG. 1 shows a flowchart of a large model correction method according to an embodiment of the present application;

[0054] FIG. 2 shows a schematic diagram of a large model correction device according to an embodiment of the present application;

[0055] FIG. 3 shows a schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0056] Exemplary embodiments of the present application will be described more fully hereinafter with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it is to be understood that the present application can be embodied in various forms without being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0057] In the description of the embodiments of the present application, it should be understood that terms such as "include" or "have" are intended to indicate that there exist the disclosed features, numbers, steps, actions, components, parts or combinations thereof in the specification, and do not exclude the presence or addition of one or more other features, numbers, steps, actions, components, parts or combinations thereof.

[0058] Unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" herein is only a description of the relationship between the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone.

[0059] The terms "first", "second", and the like are used only to facilitate the description of the same or similar technical features, and should not be understood to indicate or imply the relative importance or quantity of the technical features. Therefore, the features defined by "first", "second", and the like can explicitly or implicitly include one or more such features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of the term "a plurality of" is two or more than two.

[0060] It is found through research that the prior art is mainly based on logical Agent, but after each iteration of the large model, because of the change of convolutional layer or model drift, when interfacing with external Agent, problems are likely to occur, so a systematic correction or alignment method is needed.

[0061] Currently, when using a single logical Agent alone, only historical problems and answers can be fitted, and it is impossible to form a logical cycle within or even between different logics. For example, it can only answer specific questions such as the number of digits of pi or the upper limit of addition operation, and mathematical logic operation will be limited to the calculation ability of the Agent or the preset data set. In addition, it is also impossible to share the advantages of large model iteration and each Agent, and it is impossible to form a cycle of problem recognition to various logics or output tools, continuous use, and cross use.

[0062] To at least partially solve one or more of the above problems and other potential problems, the present application provides a large model logic correction method. In this scheme, the method for correcting the large model can be based on the geometric and mathematical logic learning module of other Agent logic or other Agent verification. The method can also be a correction method for interfacing with external Agents after large model iteration. At the same time, the Promote design and correction of the cycle or sequential logic deduction of multiple Agent logics and external Agents are more applicable.

[0063] Through a more systematic correction method, the present application can enable the large model to learn the logic of multiple Agents at the same time, interface with multiple external Agents, form a cycle or sequential deduction, and perform correction again after each large model iteration to ensure the interface with external Agents and the relationship logic between internal learning data sets.

[0064] The external Agent here can be extended to external macros, packaged software, or other software and hardware; the mathematical logic and geometric logic here can be extended to other mapping logics (such as related mapping logics required by external Agent input).

[0065] To facilitate the understanding of the present embodiment, first, a large model correction method disclosed by the present embodiment is introduced in detail. The execution subject of the large model correction method provided by the present embodiment is generally an electronic device with certain calculation ability, which includes, for example, a terminal device or a server or other processing device. The terminal device can be a user equipment (User Equipment, UE), a mobile device, a user terminal, a personal digital assistant (Personal Digital Assistant, PDA), a handheld device, etc. In some possible implementation manners, the large model correction method can be realized by calling computer readable instructions stored in the memory by the processor.

[0066] Referring to FIG. 1, a flowchart of a large model correction method provided by the present embodiment is shown, and the method includes the following steps S101-S102:

[0067] S101: Obtain a preset original learning data set for a large model;

[0068] S102: Correct the algorithm of the large model and the original learning data set based on at least one of the pre-constructed geometric logic learning module, mathematical logic learning module, and Agent instruction learning module, to obtain a corrected large model algorithm and a corrected learning data set; wherein the corrected large model algorithm and the corrected learning data set are used to meet the processing logic of the corresponding learning module, and can make the trained large model meet the demand of iterative operation.

[0069] In order to facilitate understanding of the large model correction method provided by the embodiments of the present application, the application scenarios of the method will be briefly introduced first. The large model correction method provided by the embodiments of the present application can be mainly used in the field of artificial intelligence, for example, it can be applied to various artificial intelligence subfields such as image recognition and face detection, without specific limitation here. In particular, in the large model learning stage and the correction stage after going online, the corrected large model has better model performance, so as to adapt to more complex and variable downstream application scenarios.

[0070] In the case of obtaining a preset original learning data set for a large model, at least one of the pre-constructed geometric logic learning module, mathematical logic learning module, and Agent instruction learning module can be used to correct the algorithm of the large model and the original learning data set. Here, one or more learning logics are used to correct the learning data in more dimensions, which meets the current requirement of the artificial intelligence field for the accuracy of the large model output, and also makes the trained large model meet the demand of iterative learning, which is more applicable. At the same time, through the learning and correction of the Agent instruction learning module and other Promote modules, the processing logic or input demand of the external Agent can be met at each iteration of the large model, multiple external Agents can be interfaced, a specified internal and external loop can be formed, a sequential logic deduction can be performed, the advantages of various Agents and external software can be utilized, and more diversified outputs can be applied.

[0071] The embodiments of the present application not only correct the algorithm of the large model, but also correct the original learning data set itself, so that the corrected large model algorithm and the corrected learning data set meet the processing logic of the corresponding learning module, and can make the trained large model meet the demand of iterative operation.

[0072] The geometric logic learning module is mainly used to emphasize that the corrected result meets the geometric logic sequential relationship, the mathematical logic learning module is mainly used to emphasize that the corrected result meets the mathematical logic sequential relationship, and the Agent instruction learning module is mainly used to emphasize that the corrected result meets the Agent instruction sequential relationship.

[0073] The algorithms and / or learning data sets corrected based on the various learning modules described above can better meet the needs of iterative operations. This is mainly because in the prior art, when training a large model, the output accuracy is often high due to the inability to meet the iterative operation caused by convolution migration and the like. Therefore, the correction scheme based on geometric logic learning, mathematical logic learning, and Agent instruction learning is provided to improve the accuracy of the subsequent large model output.

[0074] Considering the different effects of the various learning modules on the correction process, the large model correction method provided by the embodiments of the present application will be described in detail in combination with each learning module.

[0075] First aspect: in the actual construction process of the geometric logic learning module, the following steps are included:

[0076] Step one, obtain a first factor relationship for representing a geometric logic sequential relationship; the geometric logic sequential relationship is extracted by the Agent model after verifying the learning data set that has been screened and designed;

[0077] Step two, extract a compatible algorithm in the Agent model algorithm based on the first factor relationship and extract the corresponding learning data in the data set corresponding to the Agent model.

[0078] The first factor relationship here can be a factor relationship that conforms to the geometric logic sequential relationship. For example, the learning data set is composed of various data "a right triangle extends three squares, two squares connected to the right angle are painted the same color, and the third square is painted a different color". After extraction by the Agent model, if the learning data set appears the specified factor relationship "right triangle", "the area of the two connected squares (blue) = the area of the third connected square (red)", the common variation (corresponding to the first factor relationship described above), at this time, the Biotheorem operation logic that conforms to the geometric equivalent logic learning is embedded in the factor relationship table of the large model, and the theorem will also be reflected in the factor annotation of the data set.

[0079] In addition to the Biotheorem operation logic, the area fitting can be implemented using the Bayesian algorithm random points, the same number of red and blue points, and other operation logics.

[0080] For example, the learning data set is composed of various drawings with different shapes of triangles of the same base and height. At this time, if the factor relationship has "triangle" (convex point detection of 3), "same area" (Bayesian algorithm fitting, all triangles of the same base and height, the same number of random points to get the same area), and "height" (linear regression fitting of the third point on a line, and the line vector eigenvalue fitting is a parallel line to the third side). If the learning data is a continuous data learning data set (A B) of isosceles triangles, A is a triangle of the same base and height with different shapes, and B is a right triangle of the same base and height, it can be used as a Promote instruction learning data set, and can be used to convert a triangle into a right triangle of the same base and height, and the logic Promote instruction learning.

[0081] For example, an equilateral triangle has three equal angles, and the factor relationship has "triangle" (convex point detection of 3), triangle three sides equal (Bayesian detection), 3 equal angles (angle fitting such as vector sin+icos fitting, matrix eigenvalue fitting), etc.

[0082] The geometric logical sequence relationship in the first factor relationship includes but is not limited to the determination based on Bayesian area detection, sin+icos fitting, Fourier transform product volume, matrix equivalent transformation product volume fitting, convex point detection, etc., which are described as follows:

[0083] Convex point detection (turning point), generally using differential algorithm, when a significant differential difference appears and then stabilizes, it is considered to have a turning point (which can be a right angle, obtuse angle or acute angle), or further differential (arc angle); Bayesian area detection can also be length detection. That is, randomly select points on the picture, if the number of points selected on two pictures is the same, it means that the two areas are approximately the same. If the two lines on the picture are widened into a rectangle, it can also be used to detect length approximation; sin+icos fitting, sin+icos can refer to any point in a plane space, and is particularly suitable for fitting some natural patterns such as spiral lines and heart lines, and can also be used to define the same angle; Fourier transform is extended from sin+icos fitting, which can transform sound waves and light waves into combinations of different wavelengths, and can also be used to convert point graphs into continuous functions and convert continuous functions into characteristic frequency values (points, columns), which is mainly a approximation algorithm; in addition, there are irrational number fitting of pi, e, vector conversion, Bayesian similar module, etc., which can be flexibly used in equivalent transformation logic, and will not be described here.

[0084] Based on the above-mentioned geometric logic learning module, the algorithm of the large model and the original learning data set can be corrected according to the following steps:

[0085] Step one, input the learning data extracted from the Agent model corresponding data set to the large model for iterative calculation, determine whether the output result conforms to the geometric logical sequence relationship in the first factor relationship;

[0086] Step two, mix the learning data extracted from the Agent model corresponding data set with the large model preset original learning data set, and input it to the large model for iterative calculation, to determine whether the geometric logical sequence relationship in the first factor relationship appears;

[0087] Step three, if the geometric logical sequence relationship in the first factor relationship is met, add the geometric logical Promote relationship label to the corresponding target learning data; if the geometric logical sequence relationship in the first factor relationship is not met, delete the learning data in the large model original learning data set that does not conform to the first factor relationship, and / or add the large model algorithm.

[0088] Here, the geometric logic is mostly equivalent transformation logic, such as the same area of isosceles triangle, Bise's theorem, the proportional length of similar triangle, topology, etc. If there is a corresponding algorithm module, provide suitable "learning data set", the specified factor relationship of these theorems (i.e. the first factor relationship) will appear covariation phenomenon (i.e. "covariation factor relationship"). Based on this, the learning data extracted from the Agent model corresponding data set can be mixed with the large model preset original learning data set and input to the large model for iterative calculation to determine whether the geometric logical sequence relationship in the first factor relationship appears, and the data set that does not conform to the geometric theorem is excluded. At this time, the large model completes the learning and confirmation (correction) process of the geometric equivalent model.

[0089] Since the large model correction method provided by the embodiment of the present application excludes the data set that does not conform to the geometric logic, the data set of the entire model conforms to the fitted geometric theorem at this time. For example, if a heart line pattern is fitted, the picture data in the data set, such as a cup that reflects sunlight, will have a heart line pattern shadow at the cup opening; at the same time, the large model data set will be labeled with the corresponding equivalent factor relationship label (i.e. geometric logical Promote relationship label), so that the large model can quickly and effectively output the logic consistent output.

[0090] The equivalent change of geometry does not necessarily have a time sequence relationship, but a factor covariation or coexistence. If it is equivalent transformation logic, sometimes the output can be selected through the label (but it may be limited by the data set). That is, through the correction of the geometric equivalent factor relationship, and then correcting the data set of the large model, the logic of the output of the large model can be greatly affected (through increasing the algorithm of the large model, increasing the richness of the large model, and learning the learning data set as an auxiliary).

[0091] Here, System Promote instruction design learning can also be carried out by the embodiments of the present application, some commonly used equivalent transformation directions are established, and the common factor of output and input is used to achieve the purpose of logical continuous derivation. Further optimize and expand the logical deduction ability and range of large model. The embedded geometric System Promote can also emerge more logical deduction applications iteratively, and has the opportunity to fit the extensible compatible factor relationship, while having high model iteration stability.

[0092] The large model learning correction of geometry and mathematical logic is the mapping from equivalent logic learning correction to logical System Promote learning correction. Through logical correct data set correction and continuous deduction logical System Promote instruction learning, the features of large model learning correction of geometry and mathematical logic are extracted. In geometry learning correction, different equivalent transformation factor relationships appear at the same time, and are used as the learning basis for subsequent commonly used equivalent deduction logical System Promote.

[0093] The second aspect: in the actual construction process of the mathematical logic learning module, the following steps are included:

[0094] Step one, obtain the second factor relationship for representing the mathematical logic sequence relationship; the mathematical logic sequence relationship is obtained by Agent after verifying the learning data set screened and designed;

[0095] Step two, extract compatible algorithm in Agent model algorithm based on the second factor relationship and extract corresponding learning data in the data set corresponding to the Agent model.

[0096] The second factor relationship here can be a factor relationship consistent with the mathematical logic sequence relationship. Taking the student learning mathematics question bank as a learning data set as an example. The data bank includes time sequence picture data, and each picture has equivalent data and corresponding formula (omission =) attached. For example, the addition data bank before evolution includes the following groups of data: non-carrying single-digit addition (data set 1), carrying to ten to ten-digit addition without carrying (data set 2), wherein data set 1 is 1, 2, 3, 4, etc. Single-digit picture, fitting to emerge carrying logic and convolution layer, just need limited data set, but have extension, enough to add more than this data set digit.

[0097] For example, (sample A first, sample B second) are: (123+89, 0123+0089), (0123+0089, 0122+10+0080), (0122+10+0080, 0132+0080), (0132+0080, 0112+100+0000), (0112+100+0000, 0212+0000)..... Here, the data is analyzed first, and the lowest sample number is determined by the second factor relationship, and then the sample type is found by the clustering relationship to estimate the lowest sample number, and the logical factor type is determined.

[0098] As there are three types of 123+89, 0132+0080, and 0122+10+0080 in the factor type here, and the fruit type has 0132+0080, 0122+10+0080, and 0212+0000. The type data of the factor needs to meet the minimum sample number, and the data of the fruit also meets the minimum sample number, so that the mathematical logic sequence relationship can be identified, and then the Markov chain model of the three types of factors is fitted. If there is a type of data factor that does not exist, the logical algorithm ends, and the factor difference from the fruit to the factor appears to be concentrated in a specific type and stable to 100% or a specific proportion, at which point the logical extension appears.

[0099] The algorithm here is the fitting result of the above mapping relationship algorithm, which can effectively and stably fit the specified factor relationship that can pass through convolution extension.

[0100] Among them, the addition learning data set and the algorithm are matched to form or constitute different compatible factor relationships, as long as the logicality on the closedness and the extension combined with convolution. For example, because the addition carry of the unit number will not exceed 1 bit, but after carry, it encounters continuous 9999, so a non-9 carry after carry and two sets of data sets of continuous 9 after carry can be established. For example: {(8+7->15+0), (18+7->25+0), (38+7->45+0), (18+17->25+10), (18+37->25+30), (25+10->35+0), (25+30->55+0)...(188+37->195+30), (195+30->225+0)} and the data set matched with the previous 9 continuous carry.

[0101] Wherein, {(9998+37->09998+37), (09998+37->10005+30)…} can be taken as a data set, and the corresponding two-class logical operation is fitted by cluster analysis. The algorithm is realized by fitting the mapping relationship operation. When the fitting is successful and the order relationship of the two groups of logic appears, the specified factor relationship that can pass through convolution extension is effectively and stably fitted. The corresponding learning data set is extracted based on this logic, and the two-class mathematical mapping logic and convolution combination are fitted, which can pass through the closed logical relationship to produce the extension of circular reasoning.

[0102] It should be noted that there is a specialized algorithm corresponding to the learning data set in the compatible algorithm here, thereby forming the combination of mathematical mapping logic and convolution. However, the learning data set that has not reached the aforementioned combination can still have the mathematical logic ability limited by the general data set of external Agent (such as only addition or multiplication of several digits). At this time, by embedding the mathematical logic into the large model and fitting the compatible factor relationship with extension, the following standards need to be met:

[0103] Firstly, the learning data should be composed of the data in the large model, such as 1+1 should be composed of the existing pictures of the data set in the large model; secondly, the quality of the data should reach a certain standard in order to correct the required. Thirdly, for the A-》B order data of the System Promote learning data set, A and B should form a logically closed, here the logically closed means that when B is input as the System Promote instruction, there should be corresponding data in the A data set, and the data not contained in A is the logical deduction of the final answer. Through convolution extension (such as the logical relationship of carry in addition and breaking ten in subtraction), the extension can be formed through convolution. Finally, if the fitting here is a non-continuous mapping relationship, the influence of model iteration and drift can be avoided as much as possible. At the same time, the algorithm fitting here can effectively and stably fit the mathematical logic order relationship that can pass through convolution extension.

[0104] Based on the above-mentioned mathematical logic learning module, the algorithm of the large model and the original learning data set can be corrected according to the following steps:

[0105] Step one, input the learning data extracted from the data set corresponding to the Agent model into the large model for iterative operation, and determine whether the output result meets the mathematical logic order relationship in the second factor relationship;

[0106] Step two, mix the learning data extracted from the data set corresponding to the Agent model with the original learning data set preset by the large model, and input them into the large model for iterative operation, and determine whether the mathematical logic order relationship in the first factor relationship appears;

[0107] Step three, if the mathematical logic sequence relationship in the second factor relationship is met, add a mathematical logic Promote relationship label to the corresponding target learning data; if the second factor relationship is not met, delete the learning data in the large model original learning data set that does not meet the second factor relationship, and / or add the large model algorithm.

[0108] Similar to the process of realizing correction based on geometric logic sequence relationship, since the large model correction method provided by the embodiment of the application excludes data sets that do not meet mathematical logic, the entire model data set meets the fitted mathematical logic at this time; at the same time, the large model data set will be labeled with the corresponding equivalent factor relationship label (i.e. mathematical logic Promote relationship label), so that the large model can quickly and effectively make logical output.

[0109] Third aspect: in the actual construction process of the Agent instruction learning module, specifically includes the following steps:

[0110] Step one, obtain a third factor relationship for representing the Agent instruction sequence relationship; the Agent instruction sequence relationship is obtained by extracting the large model connected with the external Agent model after verifying the learning data set that meets the requirements of connecting the external Agent model after being screened and designed;

[0111] Step two, based on the third factor relationship, extract the compatible algorithm in the large model connected with the external Agent model and extract the corresponding learning data in the data set of the connected large model.

[0112] The Agent instruction sequence relationship here needs to realize the mapping relationship between the external Agent and the input data conversion Promote data set (A->B). The external Agent and the input data conversion Promote data set (C->D), A is the internal data set of the large model, B is the input corresponding to the external Agent, the data format meets the requirements of the external Agent, C is the input corresponding to the external Agent->D is the internal data set format or Promote instruction format of the large model.

[0113] Based on the Agent instruction learning module constructed as described above, the algorithm and the original learning data set of the large model can be corrected according to the following steps:

[0114] Step one, input the learning data extracted from the data set of the large model connected with the external Agent model to the large model for iterative calculation, and determine whether the output result meets the third factor relationship of the Agent instruction learning module;

[0115] Step two, if the third factor relationship of the Agent instruction learning module is met, the corresponding target learning data is added with the Agent instruction Promote relationship label; if the third factor relationship is not met, the learning data in the original learning data set that does not meet the third factor relationship is deleted, and / or the addition of the large model algorithm is performed.

[0116] Similar to the process of implementing correction based on geometric logical sequential relationship, i.e., mathematical logical sequential relationship, since the large model correction method provided by the embodiment of the application excludes the data set that does not meet the Agent instruction sequential relationship, the data set of the entire model meets the fitted Agent instruction sequence; at the same time, the large model data set is labeled with the corresponding equivalent factor relationship label (i.e., the Agent instruction Promote relationship label), so that the large model can quickly and effectively make a logical output.

[0117] Regardless of which correction method is used, in the actual correction process, the large model correction method provided by the embodiment of the application can also add the compatible algorithm or specialized algorithm required by the factor relationship, and determine the sequential relationship according to one or more of the following methods:

[0118] First, group each learning data in the original learning data set to obtain a plurality of grouped learning data groups; statistical tests are performed on the plurality of learning data groups, and the sequential relationship is determined based on whether the factor relationship or the factor corresponding value of each group is significantly different and whether the corresponding geometric logic, mathematical logic, and Agent instruction logic are met.

[0119] Second, the sequential relationship is determined based on a target algorithm with causal inference or time or sequential data fitting and testing characteristics, and the target algorithm includes one of a Bayesian algorithm, a Markov chain algorithm, a decision tree regression, a random forest regression, a breakpoint test, and a T-test.

[0120] Here, an example of the "instruction learning sequential data set" composed of "digits 1-9" corresponding to the picture of Arabic numerals is given.

[0121] Among them, all the sets of Arabic numeral pictures are specified, and the corresponding test method of cluster fitting is specified, such as T-test. From the large model factor relationship table, find the covariant factor relationship, and exclude 0 and constant or random factors.

[0122] Group the data (1, 2, 3), (4, 5, 6), (1, 2, 3, 12, 23), (45, 54, 64), and exclude similar relationship factors and 0 (no related factors). Each group needs to meet the minimum sample size, at which time if there is a one-to-one 100% mapping relationship formed by 0, 1-9 Arabic numerals, it can be determined through clustered data.

[0123] If a one-to-one 100% mapping relationship (factor relationship mapping characteristics) cannot be formed, then look at which numbers do not work, and increase the learning data set. For example, 9 and 8, 7 and 1 are similar in shape, and through the discovery of errors, multiple practice, and then learning to identify, the fitting here is that the eigenvalue is not obvious, or it has not been convolved out, and increasing the number of samples and then performing large model fitting.

[0124] Among them, the "A-》B Markov factor relationship" for correction can be divided into A's factor relationship characteristics, B's factor relationship characteristics, and then A-》B mapping relationship. After A and B are grouped, the A-》B mapping relationship can fit the corresponding mapping logic relationship (such as Markov chain fitting, which can determine the mapping relationship characteristics of A and B).

[0125] The "simple factor relationship" at this time refers to the need for algorithm fitting based on the System Promote instruction, and passing the test. The "logical sequence factor relationship (designated factor relationship here)" has specified logical characteristics after fitting and passing the test, such as a 100% mapping relationship of corresponding numbers.

[0126] The algorithm here can be a method of fitting and testing time or sequential data, such as Markov (Bayes) fitting, decision tree regression, random forest regression, and corresponding tests (such as breakpoint test, T test), etc. Many large models have corresponding modules, and the main emphasis in the embodiments of the present application is to correct or add compatible algorithms required by learning data or correction procedures.

[0127] For the same (or compatible) algorithm, after the aforementioned algorithm is iteratively operated when the System Promote instruction learning data set is used as the learning data set, a compatible "logical sequence factor relationship (designated factor relationship here)" will be run out to complete learning and correction.

[0128] For example, the factor change relationship between the pre-data and post-data of the Markov fitting time series can be a functional relationship or a linkage change relationship of factors, and after grouping, the probabilities of different changes, continuous functions, or mapping changes are fitted. If it is a logical mapping change, then the "logical sequence factor relationship (designated factor relationship here)" feature corresponds to a 100% mapping relationship or all possible changes.

[0129] If there is no order relationship, then increase the learning data set and / or modify the specific algorithm.

[0130] Regarding the learning and confirmation of geometric logic, the geometric logic specifies the factor relationship as "covariance factor relationship" at this time; regarding the learning and confirmation of mathematical logic, the mathematical logic specifies the factor relationship as "time series relationship", and the time series relationship can be formed by the composition of the convolution layer and other statistical algorithms to form a compatible specified factor relationship. The iterative learning of the above-mentioned logics can increase the effectiveness and stability of the logics if the following technical features are added:

[0131] One is to form a logical mapping relationship. The "specified factor relationship" has the specified logical characteristics after fitting and passing the test, such as a 100% mapping relationship corresponding to the number.

[0132] Second, the specified simple factor relationship algorithm and correction. The "simple factor relationship" and the "specified factor relationship" are as follows: The "A-》B Markov factor relationship" for correction can be divided into the factor relationship characteristics of A and the factor relationship characteristics of B, and then the A-》B mapping relationship. After the A cluster and the B cluster, the A-》B mapping relationship can fit the corresponding mapping logical relationship.

[0133] At this time, the "simple factor relationship" refers to the algorithm fitting and passing the test based on the Promote instruction. The specified simple factor relationship is designed to correct the compatible algorithm (extensible convolution, logical combination) to make the original design "logical sequence factor relationship" convolution, mapping logical combination easy to fit.

[0134] Third, the specialized algorithm such as convolution layer combination. Determine whether the factor and the logical factor have extensibility and whether they have a specific module of compatibility by comparing the specified cluster. The specified simple factor relationship can only be related, but the specialized algorithm refers to a specific convolution and regression relationship combination in the compatible relationship.

[0135] The large model correction method provided by the embodiment of the application can also realize cyclic or sequential logical deduction in combination with an externally connected Agent model, and specifically includes the following steps:

[0136] Step one, obtaining a Promote learning data set added with a Promote relationship label and its factor relationship; the Promote relationship label includes at least one of a geometric logic Promote relationship label, a mathematical logic Promote relationship label, and an Agent instruction Promote relationship label; the Promote relationship label includes a logical Promote emerged after correction or other logical Promote confirmed by an external Agent;

[0137] Step two, determine whether the factor relationship of the output data set obtained by the calculation of any Promote learning data in the Promote learning data set is coincident or coexists with the input factor relationship of itself or other Promote learning data;

[0138] Step three, if yes, then determine the logical deduction of the cycle or sequence formed between each Promote learning data.

[0139] After the relevant fitting module is inserted, according to the definition of the large model, there are appropriate modules and sufficient data, and the equivalent logic will emerge. However, in actual application, some factors may cause the logic not to emerge, or other system, non-system problems may cause the relationship to emerge, in addition, there may be some data format and other systematic problems, so it is necessary to confirm again.

[0140] For example, the equal-area factor relationship, after the entire large model data set learning training is performed by the algorithm, if other covariant factor relationships emerge, a graph set can be output for confirmation, and after confirmation, the learning data set is output for use in future model correction or stabilization, or can be used as a design System Promote learning data set.

[0141] At this time, other logical sequence relationships can be calculated through the sequential data of the large model data set, which can provide manual confirmation of whether new System Promote instructions are emerging.

[0142] Regarding the confirmation of the emerging underlying relationships, because the underlying large model is quantitative data and mathematical logic algorithms, a large number of significant continuous and non-continuous changes can easily occur. For example, using Markov chain (Bayesian algorithm) fitting, various continuous data logic can occur at this time. For example, A1->A2 (10%), A3 (90%), at this time (A1->A2, A1->A3) is composed of quantities in the ratio of 1:9.

[0143] The mapping logic here is not easily affected by model iteration, and a large number of "overfitting" logical relationships can generally be excluded. Therefore, other relationships fitted by the mapping logic related fitting method can be manually confirmed for use in subsequent model stabilization and providing other model correction.

[0144] Here, the emerging relationships and Promote design can be integrated, and the continuity of geometric logic is the A factor relationship of another Promote or geometric logic data set A->B, so the factor relationship forms coexistence, but the data set can be the same or not necessarily the same.

[0145] However, the continuous part of the mathematical logic is formed by overlapping and convolution layers with the output learning data set and the input learning data set, such as a single-digit addition and carry; and the Agent interface data set is a logical mapping relationship, mainly overlapping, if cross-reasoning is required, or if these factor relationships are not designed well or are the same learning data set, if these relationships emerge after iteration, they will also form a cycle or sequential reasoning in large models.

[0146] Mathematical logic can form extensible cycle reasoning through algorithm convolution and specific factor relationships, such as fitting two sets of logical sequential relationships through addition, which can effectively and stably fit the specified factor relationships that can be extended through convolution.

[0147] In addition, most of the geometric logic in it is equivalent transformation logic, and according to the current situation of large models, with appropriate fitting modules and sufficient data, geometric equivalent logic will emerge. That is, through various equivalent tests and fittings, equivalent logic can be determined, which is naturally embedded in the logic of large models, so it can be integrated into the logic of large models, such as the area of an isosceles triangle being the same.

[0148] In the embodiments of the present application, the emergent relationship can use the same learning correction method as the various logics described above, such as proposing a factor relationship data set with a Bayesian probability of 95%, and confirming this logical relationship, then removing the large model learning data that cannot form a 100% mapping relationship, which completes the correction.

[0149] When actually performing logical reasoning, one can combine new learning data sets and factor relationships in a determined cycle or sequential logical reasoning direction to form new cycle or sequential logical reasoning directions; the second is to form a large model according to the input order of different Agent models, while interfacing with multiple Agents and performing corresponding cycle or sequential logical reasoning; the third is to modify the Promote learning data in whole or in part, so that the output factor relationship obtained by any Promote learning data through calculation does not form overlapping or coexistence with the input factor relationship of itself or other Promote learning data, in order to break or stop the cycle or sequential logical reasoning; the fourth is to form or break the cycle or sequential logical reasoning direction of the large model by combining learning data sets.

[0150] In the embodiments of the present application, geometric logic System Promote learning data sets can be established according to common geometric logic reasoning needs, that is, geometric logic System Promote is established. There is the same covariant "logical sequential factor relationship" (specified factor relationship here) between the input and output of different geometric logic System Promote learning data sets.

[0151] At this time, two geometric logic System Promote can form a continuous logical deduction (geometric equivalent logical deduction). Establish geometric logic System Promote implementation, for example, as follows:

[0152] Establish System Promote A-》B, System Promote C-》D, A, B are equivalent data, C, D are equivalent data, but B is based on the equivalence of A with the covariant characteristics of C, D, so that through System Promote instruction learning and correction, continuous geometric equivalent logical deduction can be achieved. For example, System Promote learns data A-》B, A, B are isosceles triangles, B is a right triangle. System Promote learns data C-》D, C is a right triangle, D is a half-high isosceles rectangle (equal area). If the large model learning and correction of the above-mentioned learning data set is completed, at this time, the input of "a triangle and the specified last System Promote (equal area rectangle)" can be realized, thereby realizing the logical deduction of converting a triangle into an equal area rectangle.

[0153] The mathematical logic System Promote algorithm here includes convolution layer and time series relationship algorithm, where the time series algorithm can be Bayesian algorithm or Markov chain algorithm (both have functions). The time series factor relationship of convolution layer and time series algorithm forms a compatible specified factor relationship, and learning and correction are completed.

[0154] Example 1: If the learning data set A--》B, the Bayesian algorithm fits, A data set is an isosceles triangle, and B data set is a corresponding isosceles right triangle, the Bayesian algorithm fits, and the data set is converted from A covariant factor relationship to "A covariant factor relationship + right angle" time series Bayesian probability of 100%.

[0155] Example 2: If the learning data set A--》B, A is composed of a convolution layer of step 1-digit and a one-digit addition (such as 1+1, 9+8), and B is composed of a convolution layer of step 2-digit and a one-digit addition (2, 18). After the Bayesian algorithm fits, the data set is converted from "A covariant factor relationship" to the Bayesian probability of 100% of the time series of the answer, that is, 1+1-》2, the Bayesian probability of the time series is 100%.

[0156] For the logical deduction of the external Agent Promote, the data sets A and B are logical Promote, the logical D output of the logical D of the text C, and the continuous logical automatic fitting of the data in the input of the embedded logic and the external Agent is established by the data set data format conversion or the built-in sequence and data set data format conversion through the specified order of the embedded logic fitting, the external Agent and the input data conversion Promote.

[0157] In addition, for the establishment of the external Agent and the input data conversion System Promote and the external Agent and the output data conversion System Promote, such as the external Agent and the input data conversion System Promote data set (A- B1), the external Agent and the input data output conversion System Promote data set (C1- D1), A is the internal data set of the large model, B1 is the external Agent required for input, the data format meets the requirements of the external Agent, and the calculation of the external Agent can be started.

[0158] Among them, C1 is the output corresponding to the external Agent, D1 is the format of the internal data set of the large model or the System Promote instruction format, so D1 will form a circular deduction when inputting the large model.

[0159] For the logical deduction of the external multiple Agent System Promote, such as the logical System Promote logical deduction sequence data set (A- B2). B2 belongs to the compatible data set A, which will form a circular deduction, B2 meets the input of another logical System Promote (text C2- logical D2), which will form the output of logical D2. Therefore, the continuous logical automatic fitting of the data in the input of the embedded logic and the external Agent can be established by the data set data format conversion or the built-in sequence and data set data format conversion through the specified order of the embedded logic fitting, the external Agent and the input data conversion System Promote.

[0160] Among them, the built-in logical deduction sequence will be calculated first, and the format of the data meets the input of the external Agent or the embedded logic. For example, B2 (Roman numerals such as X, V), B1 (Arabic numerals such as 2, 15), if the built-in learning data set is converted into Roman numerals in advance, the external Agent will not start even if it is input first, and the Roman numerals will form a logical deduction after being input into the large model.

[0161] In the embodiments of the present application, the external Agent can be converted in advance by the built-in logical deduction sequence factor table, and then output to various external Agents. The completed or incomplete conversion is converted back through the external Agent and the output data conversion System Promote, and then fitted with the built-in logic.

[0162] As A-》may be B1 or B2, the external Agent is first input in sequence, such as A-》B1 input Agent, which starts the external Agent, and then outputs the converted data to the large model. When A-》B2 input Agent is input, the Agent is not started, and therefore the large model is input again with B2, which forms the logical System Promote logical deduction sequence.

[0163] At this time, B2 may include B1, such as A-》10+10 or 10000+10000. Assuming that the Agent can only perform 100+100 operation, the System Promote loop logic has no limit, at this time, the input System Promote loop logic of 10+10, 10000+10000 can be input, but because the Agent is input first, the Agent outputs 20, 10000+10000 to the large model, and the large model outputs 20, 20000.

[0164] In the embodiments of the present application, different logical System Promote instruction learning data sets can also be mixed. For example, in order to be compatible with various external and embedded modules, multiple layers or "logical deduction learning sequence data set" may be required. For example, in order to be compatible with the input form of various embedded and external logical modules, the output and re-input of System Promote instruction are used. For example, "digital and code conversion learning sequence data set" (see equivalent logic), when the code input external module data is not consistent, it is converted into digital format through "digital and code conversion learning module" and then input. "Digital and code conversion learning data set" is also a part of System Promote "System Promote instruction learning data set", which uses the same algorithm and correction method to complete the correction. As can be seen, the entire large model System Promote instruction can automatically perform logical deduction.

[0165] The "instruction learning sequence data set" here includes A--》B sequence data (i.e. causal or time series data), A is the characteristic learning data of System Promote, and B is the corresponding System Promote instruction and logical element instruction thereafter. B needs to be compatible with or correspond to the input A of "equivalent logic, sequence logic learning data set" or the learning data format requirement of logical Agent input A.

[0166] Wherein A, B can be designed in multiple layers, the first layer is the System Promote instruction learning data set. B contains the use of human, time, place and object design as much as possible, which can maximize the subsequent logical automatic deduction application, and needs to be compatible with the Agent required instructions, or compatible with the large model logical meta-instruction. The quality of A data set needs to be able to identify specific data sets, but does not need to cover all data categories. Here you can cooperate with "zero sample learning" or expand the identification type with large model iteration evolution.

[0167] In addition, in the first layer of System Promote learning data set deduction of large model, it is suggested that B has elements such as "people, events, time, place, and objects" so that subsequent System Promote can be used for more generalized logical automatic deduction of various texts.

[0168] In the embodiments of the present application, System Promote "instruction learning sequence data set" of different logics can also be mixed. For example, an example is given in the A-》B mapping relationship. The A part of the System Promote "instruction learning sequence data set" about addition can be a Chinese text problem bank about addition for primary school students. The artificial mark time continuous result B needs to be compatible with "Agent required instructions" or "large model logical meta-instruction".

[0169] For example, A ("There are 3 apples and 2 oranges on the table at home in the morning. Eat 2 apples. How many apples are left?" )-》B (morning (time), home (place), 3-2 apples (event & object)).

[0170] The "number and code conversion learning sequence data set" therein. If the Agent needs a certain format of data, such as originally "there are..." 》5-3 (picture), the data format needs to be compatible with the Agent required instruction "5-3" (text), or compatible with the large model logical meta-instruction, such as text as the input element of Agent, if necessary, a layer of System Promote learning data set can also be added, 5-3 (picture) 》5-3 (text).

[0171] The "sequence logic learning data set" therein. For example: A ("1+3*5-8 / 2")-》B ("1+3*5-8 / 2", *) or A ("1+3*5-8 / 2")-》B ("3*5", "1+3*5-8 / 2"), which needs to meet the needs of continuous input System Promote instruction.

[0172] Then, through "zero data learning" and "few sample learning", the logic of large model iteration is expanded. Input A ("Based on the photo and conversation in the morning, how many apples are there in the house in the morning?", Conversation "There are 3 apples and 2 oranges on the table, eat 2 apples" and there is a photo of an apple tree outside the door with three apples on it) -> the problem drives the identification of B (morning (time), home (place), 3-2+1+1+1 apples (things and objects)"). At this time, the logical program of System Promote is entered, the logical sequence is identified first, and the sequence is entered into the mathematical logic result fitting. Finally, 4 apples are output.

[0173] The embodiment of the present application can also set "logical deduction sequence learning data set" for learning such inputs as "1+13*2". At this time, the multiplication logical System Promote instruction can be executed first, or the 13*2 input Agent can be intercepted, and if necessary, the "number and code conversion learning module" of the Agent output can be embedded back into the System Promote instruction, so as to form more generalized logical automatic deduction; or it can be directly output to Excel-Pyron plotting; or the product volume layer transformation fitting logical deduction result is transformed... and other operations.

[0174] The logical System Promote (text -> logic), logical System Promote (logical deduction sequence), geometric and mathematical logic (System Promote), external Agent and input data conversion System Promote, and external Agent and output data conversion System Promote in the embodiment of the present application are only to form System Promote cyclic deduction, that is, each data set is (A -> B), B is compatible with the input end data set of A or other System Promote, so as to form cyclic deduction when input again.

[0175] Among them, for logical System Promote (text -> logic), logical System Promote (logical deduction sequence), logical System Promote (logical deduction sequence), external Agent and input data conversion System Promote, and external Agent and output data conversion System Promote, generally confirm whether the "time sequence factor relationship" is formed.

[0176] In the part of interfacing external Agent, after correcting the "time sequence factor relationship", the "corresponding factor value" is further confirmed, taking numbers as an example, that is, corresponding to non-continuous 1:1 factor relationship, such as 1 corresponding to certain factor L, that is, determining that Bayes 100% corresponds.

[0177] In summary, the embodiments of the present application correct the learning materials in more dimensions through one or more learning logics, which meets the current requirements of the artificial intelligence field for the accuracy of large model output, and also enables the large model after training to meet the demand of iterative learning, and is more applicable. At the same time, through the learning and correction of the Agent instruction learning module and other Promote modules, the large model can be iterated each time to meet the processing logic or input demand of the external Agent, interface with multiple external Agents, form a specified internal and external loop, sequential logic deduction, utilize the advantages of various Agents and external software, and be applicable to more diversified outputs.

[0178] In the description of the present specification, the description referring to the terms "some possible embodiments", "some embodiments", "examples", "specific examples", or "some examples" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiments or examples are contained in at least one embodiment or example of the present application, and the above terms do not necessarily represent the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0179] Regarding the method flowchart of the embodiments of the present application, some operations are described as different steps executed in a certain order. Such flowcharts are illustrative rather than limiting. Some steps described herein can be grouped together and executed in a single operation, or some steps can be divided into multiple sub-steps, and some steps can be executed in an order different from that shown herein. Each step shown in the flowchart can be implemented in any way by any circuit structure and / or tangible mechanism (for example, by software running on a computer device, hardware (for example, processor or chip implemented logic function), etc., and / or any combination thereof) in any manner.

[0180] Those skilled in the art can understand that in the method described in the above specific embodiments, the writing order of each step does not mean a strict execution order, and the specific execution order of each step should be determined by its function and possible inherent logic.

[0181] Based on the same inventive concept, the embodiments of the present application also provide a large model correction device corresponding to the large model correction method. Since the principle of solving problems in the device of the embodiments of the present application is similar to the above-mentioned large model correction method of the embodiments of the present application, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described again.

[0182] Referring to FIG. 2, a schematic diagram of a large model correction device provided by an embodiment of the present application is shown. The device includes an acquisition module, a correction module 202; wherein,

[0183] The acquisition module 201 is configured to acquire a preset original learning data set for a large model.

[0184] The correction module 202 is configured to correct the algorithm of the large model and the original learning data set based on at least one of the pre-constructed geometric logic learning module, mathematical logic learning module, and Agent instruction learning module, to obtain a corrected large model algorithm and a corrected learning data set.

[0185] The corrected large model algorithm and the corrected learning data set are used to conform to the processing logic of the corresponding learning module, and can make the trained large model meet the demand of iterative operation.

[0186] With the above large model correction device, when the original learning data set preset for the large model is acquired, the algorithm of the large model and the original learning data set can be corrected based on at least one of the pre-constructed geometric logic learning module, mathematical logic learning module, and Agent instruction learning module, to obtain a corrected large model algorithm and a corrected learning data set. Since the learning data is corrected in more dimensions by one or more learning logics, the corrected learning data set not only conforms to the processing logic of the learning module, but also makes the trained large model meet the demand of iterative operation, which meets the current requirements of the artificial intelligence field for the accuracy of large model output, and is more applicable.

[0187] Optionally, the correction module 202 is configured to construct the geometric logic learning module according to the following steps:

[0188] A first factor relationship for representing a geometric logic sequential relationship is acquired. The geometric logic sequential relationship is extracted by the Agent model after verifying the learning data set that has been screened and designed.

[0189] A compatible algorithm is extracted in the Agent model algorithm based on the first factor relationship, and corresponding learning data is extracted in the data set corresponding to the Agent model.

[0190] Optionally, the correction module 202 is specifically configured to correct the algorithm of the large model and the original learning data set according to the following steps:

[0191] The learning data extracted from the data set corresponding to the Agent model is input into the large model for iterative operation to determine whether the output result conforms to the geometric logic sequential relationship in the first factor relationship.

[0192] The learning data extracted from the Agent model corresponding data set is mixed with the original learning data set preset by the large model, and is input into the large model for iterative calculation to determine whether the geometric logical sequence relationship in the first factor relationship appears;

[0193] If the geometric logical sequence relationship in the first factor relationship is met, a geometric logical Promote relationship label is added for the corresponding target learning data.

[0194] If the geometric logical sequence relationship in the first factor relationship is not met, the learning data in the original learning data set of the large model that does not meet the first factor relationship is deleted, and / or the large model algorithm is added.

[0195] The optional correction module 202 is configured to construct a mathematical logic learning module according to the following steps:

[0196] A second factor relationship for representing a mathematical logic sequence relationship is obtained, which is extracted by the Agent after completing verification on the screened and designed learning data set;

[0197] Based on the second factor relationship, a compatible algorithm is extracted in the Agent model algorithm, and corresponding learning data is extracted in the data set corresponding to the Agent model.

[0198] The optional correction module 202 is specifically configured to correct the algorithm and the original learning data set of the large model according to the following steps:

[0199] The learning data extracted from the data set corresponding to the Agent model is input into the large model for iterative calculation to determine whether the mathematical logical sequence relationship in the second factor relationship is met.

[0200] The learning data extracted from the data set corresponding to the Agent model is mixed with the original learning data set preset by the large model, and is input into the large model for iterative calculation to determine whether the mathematical logical sequence relationship in the first factor relationship appears.

[0201] If the mathematical logical sequence relationship in the second factor relationship is met, a mathematical logical Promote relationship label is added for the corresponding target learning data.

[0202] If the second factor relationship is not met, the learning data in the original learning data set of the large model that does not meet the second factor relationship is deleted, and / or the large model algorithm is added.

[0203] The optional correction module 202 is configured to construct an Agent instruction learning module according to the following steps:

[0204] obtaining a third factor relationship for characterizing the agent instruction sequential relationship; the agent instruction sequential relationship is obtained by a large model connected with an external agent model after the large model is verified by a filtered and designed learning data set meeting the requirements of the connected external agent model;

[0205] Based on the third factor relationship, a compatible algorithm is extracted in the large model connected with the external agent model, and corresponding learning data is extracted in the data set of the connected large model.

[0206] Optionally, the correction module 202 is specifically configured to correct the algorithm of the large model and the original learning data set according to the following steps:

[0207] The learning data extracted from the data set of the large model connected with the external agent model is input into the large model for iterative calculation to determine whether the output result meets the third factor relationship of the agent instruction learning module;

[0208] If the third factor relationship of the agent instruction learning module is met, an agent instruction promote relationship label is added to the corresponding target learning data;

[0209] If the third factor relationship is not met, the learning data in the original learning data set that does not meet the third factor relationship is deleted, and / or the addition of the algorithm of the large model is performed.

[0210] Optionally, the correction module 202 is further configured to:

[0211] When the correction is based on the factor relationship, a compatible algorithm or a specialized algorithm required by the factor relationship is added; the sequential relationship is determined according to one or more of the following ways:

[0212] Each learning data in the original learning data set is grouped to obtain a plurality of grouped learning data groups; statistical tests are performed on the plurality of learning data groups, and the sequential relationship is determined based on whether the factor relationship or the factor corresponding value of each group is significantly different and whether the corresponding geometric logic, mathematical logic, and agent instruction logic are met;

[0213] The sequential relationship is determined based on a target algorithm having a causal inference or a time or sequential data fitting and testing characteristic, and the target algorithm includes one of a Bayesian algorithm, a Markov chain algorithm, a decision tree regression, a random forest regression, a breakpoint test, and a T-test.

[0214] Optionally, it further includes:

[0215] The deduction module 203 is configured to acquire the Promote learning data set added with the Promote relationship label and the factor relationship thereof; the Promote relationship label comprises at least one of a geometric logic Promote relationship label, a mathematical logic Promote relationship label, and an Agent instruction Promote relationship label; the Promote relationship label comprises a corrected emergent logic Promote or other logic Promote confirmed by an external Agent; and it is determined whether the factor relationship of the output data set obtained by calculation of any Promote learning data in the Promote learning data set coincides with or coexists with the input factor relationship of the Promote learning data itself or other Promote learning data.

[0216] If yes, it is determined that the logical deduction between the Promote learning data forms a cycle or a sequence.

[0217] Optionally, the deduction module 203 is further configured to:

[0218] combine the determined cycle or sequence logical deduction direction to form a new learning data set and a factor relationship, and form a new cycle or sequence logical deduction direction; or

[0219] form a large model according to the new cycle or sequence logical deduction direction and the input sequence of different Agent models, and simultaneously perform corresponding cycle or sequence logical deduction with multiple Agents; or

[0220] break or stop the cycle or sequence logical deduction by modifying all or part of the Promote learning data, so that the output factor relationship obtained by calculation of any Promote learning data does not coincide with or coexist with the input factor relationship of the Promote learning data itself or other Promote learning data; or

[0221] form or break the cycle or sequence logical deduction direction of the large model by combining the learning data set.

[0222] It should be noted that the device in the embodiments of the present application can realize each process of the embodiments of the foregoing method, and achieve the same effects and functions, which will not be described here.

[0223] The embodiment of the present application also provides an electronic device, as shown in Figure 3, which is a structural schematic diagram of an electronic device provided by the embodiment of the present application, and includes a processor 301, a memory 302, and a bus 303. The memory 302 stores machine readable instructions executable by the processor 301 (for example, execution instructions corresponding to the acquisition module 201 and the correction module 202 in the device in Figure 2, etc.), when the electronic device is running, the processor 301 and the memory 302 communicate through the bus 303, and the machine readable instructions are executed by the processor 301 to perform the following processing:

[0224] acquire a preset original learning data set for the large model;

[0225] correct the algorithm of the large model and the original learning data set based on at least one of the pre-constructed geometric logic learning module, the mathematical logic learning module, and the Agent instruction learning module, to obtain a corrected large model algorithm and a corrected learning data set;

[0226] The corrected large model algorithm and the corrected learning data set are used to meet the processing logic of the corresponding learning module, and can make the trained large model meet the demand of iterative operation.

[0227] The embodiment of the present application also provides a computer readable storage medium, and the computer readable storage medium stores a computer program, and the computer program is run by a processor to execute the steps of the large model correction method described in the method embodiment. The storage medium can be a volatile or non-volatile computer readable storage medium.

[0228] The embodiment of the present application also provides a computer program product, and the computer program product carries program code, and the instructions included in the program code can be used to execute the steps of the large model correction method described in the method embodiment, and specific implementation can be referred to the method embodiment, and details are not described herein.

[0229] The computer program product can be specifically implemented by hardware, software or a combination thereof. In an optional embodiment, the computer program product is specifically embodied as a computer storage medium, and in another optional embodiment, the computer program product is specifically embodied as a software product, for example, a software development kit (Software Development Kit, SDK) and the like.

[0230] The various embodiments in the present application are described in progressive manner, and the same or similar parts among the various embodiments can be cross-referenced. Each embodiment focuses on the differences from other embodiments. In particular, for the apparatus, device and computer-readable storage medium embodiments, since they are basically similar to the method embodiments, the description is simplified, and the relevant parts can be referred to the description of the method embodiments.

[0231] The apparatus, device and computer-readable storage medium provided by the embodiments in the present application correspond to the method, and therefore, the apparatus, device and computer-readable storage medium also have the similar beneficial technical effects as the method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the apparatus, device and computer-readable storage medium will not be described here again.

[0232] Those skilled in the art should understand that the embodiments in the present application can be implemented in the form of a method and apparatus (device or system), or a computer-readable storage medium. Therefore, the present application can be implemented in the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware. Moreover, the present application can be in the form of a computer-readable storage medium implemented on one or more computer-readable storage media containing computer-usable program codes, including but not limited to magnetic disk storage, CD-ROM storage, optical storage, etc.

[0233] The present application is described with reference to flowcharts and / or block diagrams of the method, apparatus (device or system) and computer-readable storage medium according to the embodiments in the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts 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, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce the functions specified in one or more flows in the flowcharts and / or one or more blocks in the block diagrams.

[0234] These computer program instructions can also be stored in a computer-readable memory that can cause the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including instruction apparatus, wherein the instruction apparatus implements the functions specified in one or more flows in the flowcharts and / or one or more blocks in the block diagrams.

[0235] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks and / or flowchart diagrams in one or more flowcharts.

[0236] In one typical arrangement, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0237] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory, non-volatile memory, such as read-only memory (ROM), EPROM, and / or flash memory. The memory is an example of computer readable media.

[0238] Computer readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology for storing information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer readable 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, read-only memory, electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape magnetic disk storage or other magnetic storage device, or any other non-transmission medium that can be used to store information accessible by a computing device. In addition, although the operations of the methods of the present application are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in that particular order, or that all of the illustrated operations must be performed to achieve the desired result. In addition, certain steps can be omitted, combined into a single step, and / or divided into multiple steps.

[0239] Although the spirit and principles of the present application have been described above with reference to several specific embodiments, it should be understood that the present application is not limited to the disclosed specific embodiments, and the division of aspects does not mean that the features in these aspects cannot be combined. The present application is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the appended claims.

Claims

1. A large model correction method, comprising: obtaining a preset original learning data set for a large model; correcting an algorithm and the original learning data set of the large model based on at least one of a geometric logic learning module, a mathematical logic learning module, and an Agent instruction learning module, to obtain a corrected large model algorithm and a corrected learning data set; wherein the corrected large model algorithm and the corrected learning data set are used to conform to the processing logic of the corresponding learning module, and can make the trained large model meet the demand of iterative operation.

2. The method of claim 1, wherein, The geometric logic learning module is constructed according to the following steps: obtaining a first factor relationship for representing a geometric logic sequential relationship; the geometric logic sequential relationship is extracted by an Agent model after verifying a screened and designed learning data set; extracting a compatible algorithm in the Agent model algorithm and corresponding learning data in the data set corresponding to the Agent model based on the first factor relationship.

3. The method of claim 2, wherein, The algorithm and the original learning data set of the large model are respectively corrected, including: determining whether the output result conforms to the geometric logic sequential relationship in the first factor relationship by inputting the learning data extracted from the data set corresponding to the Agent model into the large model for iterative operation; determining whether the geometric logic sequential relationship in the first factor relationship appears by mixing the learning data extracted from the data set corresponding to the Agent model with the preset original learning data set of the large model and inputting them into the large model for iterative operation; if the geometric logic sequential relationship in the first factor relationship is met, adding a geometric logic Promote relationship label to the corresponding target learning data; if the geometric logic sequential relationship in the first factor relationship is not met, deleting the learning data in the original learning data set of the large model that does not conform to the first factor relationship, and / or adding the algorithm of the large model.

4. The method of any one of claims 1 to 3, wherein, The mathematical logic learning module is constructed according to the following steps: obtaining a second factor relationship for representing a mathematical logic sequential relationship; the mathematical logic sequential relationship is extracted by an Agent after verifying a screened and designed learning data set; extracting a compatible algorithm in the Agent model algorithm and corresponding learning data in the data set corresponding to the Agent model based on the second factor relationship.

5. The method of claim 4, wherein, The algorithm and the original learning data set of the large model are respectively corrected, including: determining whether the output result conforms to the mathematical logic sequential relationship in the second factor relationship by inputting the learning data extracted from the data set corresponding to the Agent model into the large model for iterative operation; determining whether the mathematical logic sequential relationship in the first factor relationship appears by mixing the learning data extracted from the data set corresponding to the Agent model with the preset original learning data set of the large model and inputting them into the large model for iterative operation; if the mathematical logic sequential relationship in the second factor relationship is met, adding a mathematical logic Promote relationship label to the corresponding target learning data; If the second factor relationship is not met, the learning data in the large model original learning data set that does not meet the second factor relationship is deleted, and / or the large model algorithm is added.

6. The method of any one of claims 1 to 5, wherein, The Agent instruction learning module is constructed according to the following steps: A third factor relationship for representing an Agent instruction sequence relationship is obtained; the Agent instruction sequence relationship is extracted from a large model connected with an external Agent model after the large model is verified by using a learning data set that is screened and designed to meet the requirements of the external Agent model; Based on the third factor relationship, a compatible algorithm is extracted from the large model connected with the external Agent model, and corresponding learning data is extracted from the data set of the connected large model.

7. The method of claim 6, wherein, The algorithm and the original learning data set of the large model are corrected respectively, including: The learning data extracted from the data set of the large model connected with the external Agent model is input into the large model for iterative calculation to determine whether the output result meets the third factor relationship of the Agent instruction learning module; If the third factor relationship is met, an Agent instruction Promote relationship label is added to the corresponding target learning data; If the third factor relationship is not met, the learning data in the original learning data set that does not meet the third factor relationship is deleted, and / or the large model algorithm is added.

8. The method of any one of claims 3, 5, 7, wherein, When the factor relationship is corrected, it also includes: Adding compatible algorithms or specialized algorithms required by the factor relationship; And / or, the order relationship of the factor relationship is determined in one or more of the following ways: Grouping each learning data in the original learning data set to obtain a plurality of grouped learning data groups; statistical test is performed on the plurality of learning data groups, and the order relationship is determined based on whether the factor relationship or the factor corresponding value of each group is significantly different and whether the corresponding geometric logic, mathematical logic, Agent instruction logic is met; The order relationship is determined based on a target algorithm with causal inference or time or sequence data fitting and test characteristics, and the target algorithm includes one of Bayesian algorithm, Markov chain algorithm, decision tree regression, random forest regression, breakpoint test, and T-test.

9. The method of any one of claims 1 to 7, wherein, It also includes: Obtaining a Promote learning data set with a Promote relationship label and its factor relationship; The Promote relationship label includes at least one of geometric logic Promote relationship label, mathematical logic Promote relationship label, and Agent instruction Promote relationship label; the Promote relationship label includes a corrected emerging logic Promote or other logic Promote confirmed by an external Agent model; It is determined whether the factor relationship of the data set obtained by calculating any Promote learning data in the Promote learning data set forms overlap or coexistence with the input factor relationship of itself or other Promote learning data. If yes, determine the logical deduction of the cycle or sequence between each Promote learning material.

10. The method of claim 9, wherein, The method further comprises: Combining the new learning material set and the factor relationship in the determined cycle or sequence logical deduction direction to form a new cycle or sequence logical deduction direction; or, According to the new cycle or sequence logical deduction direction, cooperate with the input sequence of different Agent models to form a large model, and simultaneously interface with multiple Agent models and perform corresponding cycle or sequence logical deduction; or, By modifying all or part of the Promote learning material, so that the output factor relationship obtained by any Promote learning material through calculation does not form overlap or coexist with the input factor relationship of itself or other Promote learning materials, to break or stop the cycle or sequence logical deduction; or, Through the combination of the learning material set, a method of forming or breaking the cycle or sequence logical deduction direction of the large model is formed.

Images