Large model correction method

Abstract

The invention provides a large model correction method, and the method comprises the steps: obtaining a preset original learning material set for a large model; the algorithm and the original learning material set are corrected based on at least one of a pre-constructed geometric logic learning module, a mathematical logic learning module and an Agent instruction learning module, and the learning materials are corrected in more dimensions through one or more learning logics. The requirement of the current artificial intelligence field for the output accuracy of the large model is met, the trained large model can meet the requirement of iterative learning, and the method is more suitable. Meanwhile, through learning and correction of the Agent instruction learning module and other Promote modules, large model iteration can be carried out at each time, the processing logic or input requirements of external Agents are met, a plurality of external Agents are connected at the same time, specified internal and external circulation and sequential logic deduction are formed, the advantages of various Agents and external software are utilized, and the method is suitable for more multivariate output and the like.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a large model correction method. Background Art

[0002] With the rapid development of artificial intelligence (AI), large models are widely used in all walks of life. However, in practical applications, especially for logical fitting or calculations implemented by agents and Markov chains, large models have the problem of low output accuracy.

[0003] Taking Agent as an example, as a hardware and software system with autonomous, social, responsive and proactive capabilities, Agent can generally only be used in closed, specific business-domain artificial intelligence application scenarios (such as playing chess). That is, Agent is usually used as a closed model.

[0004] However, when using closed models such as Agent to implement iterative optimization of large models, the connected logic often fails due to drift and other problems. Here, as users use it, the large model will incorporate new data and perform feedback iterations, causing model drift or slight changes in convolution methods, causing factor value changes, which makes the logical fitting of the relevant logic modules in the large model, such as basic addition, subtraction, multiplication and division operations, prone to errors.

[0005] Based on this, there is an urgent need for a correction solution to improve the output accuracy of large models. Summary of the invention

[0006] The disclosed embodiments at least provide a large model calibration method to improve the accuracy of the large model output by calibrating the learning data set and its algorithm.

[0007] In a first aspect, an embodiment of the present disclosure provides a large model calibration method, comprising:

[0008] Get the original learning data set preset for large models;

[0009] Based on at least one of the pre-built 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 respectively corrected to obtain a corrected large model algorithm and a corrected learning data set;

[0010] Among them, the corrected large model algorithm and the corrected learning data set are used to comply with the processing logic of the corresponding learning module, and can make the trained large model meet the needs of iterative operations.

[0011] In a possible implementation, the geometric logic learning module is constructed according to the following steps:

[0012] Obtaining a first factor relationship for characterizing a geometric logic sequence relationship; the geometric logic sequence relationship is extracted by the Agent model after completing verification of a screened and designed learning data set;

[0013] Based on the first factor relationship, a compatible algorithm is extracted from the Agent model algorithm and corresponding learning data is extracted from the data set corresponding to the Agent model.

[0014] In a possible implementation, the algorithm of the large model and the original learning data set are calibrated respectively, including:

[0015] Inputting the learning data extracted from the data set corresponding to the Agent model into the large model for iterative calculation to determine whether the output result conforms to the geometric logic sequence relationship in the first factor relationship;

[0016] Mixing the learning data extracted from the data set corresponding to the Agent model with the original learning data set preset in the large model, and inputting them into the large model for iterative calculation to determine whether the geometric logical sequence relationship in the first factor relationship appears;

[0017] If the geometric logic order relationship in the first factor relationship is met, a geometric logic Promote relationship tag is added to the corresponding target learning material;

[0018] If the geometric logical order relationship in the first factor relationship is not met, the learning materials that do not meet the first factor relationship in the original learning data set of the large model are deleted, and / or the large model algorithm is added.

[0019] In a possible implementation, the mathematical logic learning module is constructed according to the following steps:

[0020] Obtaining a second factor relationship for representing a mathematical logic sequence relationship; the mathematical logic sequence relationship is extracted by the Agent after completing verification of a screened and designed learning data set;

[0021] Based on the second factor relationship, a compatible algorithm is extracted from the Agent model algorithm and corresponding learning data is extracted from the data set corresponding to the Agent model.

[0022] In a possible implementation, the algorithm of the large model and the original learning data set are calibrated respectively, including:

[0023] Inputting the learning data extracted from the data set corresponding to the Agent model into the large model for iterative calculation to determine whether the output result conforms to the mathematical logic sequence relationship in the second factor relationship;

[0024] Mixing 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 inputting them into the large model for iterative calculation to determine whether the mathematical logic sequence relationship in the first factor relationship appears;

[0025] If the mathematical logic order relationship in the second factor relationship is met, a mathematical logic Promote relationship tag is added to the corresponding target learning material;

[0026] If the second factor relationship is not met, the learning materials that do not meet the second factor relationship in the original learning data set of the large model are deleted, and / or the large model algorithm is added.

[0027] In a possible implementation, the Agent instruction learning module is constructed according to the following steps:

[0028] Obtaining a third factor relationship for characterizing the sequential relationship of Agent instructions; the Agent instruction sequential relationship is extracted by the large model connected to the external Agent model after completing verification of a learning data set that has been screened and designed to meet the requirements of connecting to the external Agent model;

[0029] Based on the third factor relationship, a compatible algorithm is extracted from the large model connected to the external Agent model and corresponding learning data is extracted from the data set of the connected large model.

[0030] In a possible implementation, the algorithm of the large model and the original learning data set are calibrated respectively, including:

[0031] The learning data extracted from the data set of the large model connected to 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;

[0032] If the third factor relationship of the Agent instruction learning module is met, then the Agent instruction Promote relationship tag is added to the corresponding target learning material;

[0033] If the third factor relationship is not met, the learning materials in the original learning data set that do not meet the third factor relationship are deleted, and / or a large model algorithm is added.

[0034] In a possible implementation, when the correction is performed based on the factor relationship, the method further includes:

[0035] Add the required compatible algorithms or specialized algorithms for factor relations;

[0036] The order relationship is determined in one or more of the following ways:

[0037] Grouping the learning materials in the original learning material set to obtain a plurality of grouped learning material groups; performing statistical tests on the plurality of learning material groups, and determining based on whether the factor relationships or corresponding factor values ​​of the groups are significantly different, and whether they conform to the judgment results of corresponding geometric logic, mathematical logic, and Agent instruction logic;

[0038] The sequential relationship is determined based on a target algorithm with causal inference or temporal or sequential data fitting and verification characteristics, wherein 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.

[0039] In a possible implementation, it further includes:

[0040] Acquire a Promote learning data set and its factor relationships with Promote relationship tags added thereto; the Promote relationship tags include at least one of a geometric logic Promote relationship tag, a mathematical logic Promote relationship tag, and an Agent instruction Promote relationship tag; the Promote relationship tags include logic Promote that emerges after correction, or other logic Promote that has been confirmed by an external Agent model;

[0041] Determine whether the factor relationship of the data set output by calculating any Promote learning data in the Promote learning data set overlaps or coexists with the input factor relationship of itself or other Promote learning data;

[0042] If so, determine the logical deduction between each Promote learning material to form a cycle or sequence.

[0043] In a possible implementation, the method further includes:

[0044] With a certain cyclic or sequential logical deduction direction, a new set of learning data and factor relationships are combined to form a new cyclic or sequential logical deduction direction; or,

[0045] According to the new cyclic or sequential logic deduction direction, in conjunction with the input order of different Agent models, a large model is formed, which is connected with multiple Agent models at the same time and performs corresponding cyclic or sequential logic deduction; or,

[0046] Modify all or part of the Promote learning materials so that the output factor relationship of any Promote learning material does not overlap or coexist with the input factor relationship of itself or other Promote learning materials, so as to interrupt or terminate the cyclic or sequential logical deduction; or

[0047] A method for forming or interrupting a large model cycle or sequential logic deduction direction by combining the learning data sets.

[0048] By adopting the above-mentioned large model correction method, when the original learning data set preset for the large model is obtained, the algorithm and the original learning data set of the large model can be corrected based on at least one of the pre-built geometric logic learning module, mathematical logic learning module, and 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 through one or more learning logics here, the corrected learning data set can not only conform to the processing logic of the learning module, but also enable the trained large model to meet the needs of iterative operations. This meets the current requirements for the accuracy of large model output in the field of artificial intelligence and is more applicable.

[0049] Other advantages of the present disclosure will be explained in more detail in conjunction with the following description and drawings.

[0050] It should be understood that the above description is only an overview of the technical solution of the present disclosure, so that the technical means of the present disclosure can be generally understood and then implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present disclosure more obvious and easy to understand, the specific implementation methods of the present disclosure are specifically described below with examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly introduced below. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the drawings only illustrate certain embodiments of the present disclosure and therefore should not be regarded as limiting the scope of protection. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work. Moreover, the same reference numerals are used to represent the same components throughout the drawings. In the drawings:

[0052] Figure 1 A flow chart of a large model correction method provided by an embodiment of the present disclosure is shown;

[0053] Figure 2 A schematic diagram of a large model correction device provided by an embodiment of the present disclosure is shown;

[0054] Figure 3 A schematic diagram of an electronic device provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0055] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0056] In the description of the embodiments of the present disclosure, it should be understood that terms such as "including" or "having" are intended to indicate the presence of the disclosed features, numbers, steps, behaviors, components, parts, or a combination thereof in the specification, and do not exclude the possibility of the presence of one or more other features, numbers, steps, behaviors, components, parts, or a combination thereof.

[0057] Unless otherwise specified, “ / ” means or. For example, A / B can mean A or B. The “and / or” in this article is merely a way to describe the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0058] The terms "first", "second", etc. are used only to distinguish the same or similar technical features for the convenience of description, and should not be understood as indicating or implying the relative importance or quantity of these technical features. Thus, the features defined by "first", "second", etc. may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present disclosure, unless otherwise specified, the term "plurality" means two or more than two.

[0059] Research has found that the existing technology is mainly based on logical agents, but after each iteration of the large model, problems are likely to occur when connecting to external agents due to changes in the convolutional layer or model drift, so a systematic correction or alignment method is needed.

[0060] At present, when a single logic agent is used, it only fits historical questions and answers, and cannot form a cyclic deduction within the logic or even between logics. For example, it can only answer specific questions, such as the number of digits of pi or the upper limit of addition calculation. Mathematical logic operations will be limited by the calculation ability of the agent or the preset data set. In addition, it is impossible to share the large model iteration and the advantages of each agent, and it is impossible to form a cyclic, continuous, and cross-use from problem identification to various logics or output tools.

[0061] In order to at least partially solve one or more of the above problems and other potential problems, the present disclosure provides a large model logic correction method. In this solution, a method for correcting a large model can be based on other agent logics or geometric and mathematical logic learning modules verified by other agents. A correction method can also be used to form a docking with an external agent after iteration of the large model. At the same time, the Promote design and correction of cyclic or sequential logic deduction of multiple agent logics and external agents is more applicable.

[0062] The present disclosure adopts a more systematic calibration method, which enables the large model to learn the logic of multiple agents at the same time, connect to multiple external agents, form a cycle or sequential deduction, and perform calibration again after each iteration of the large model to ensure the connection with the external agent and conform to the relationship logic with the internal learning data set.

[0063] The external agent here can be extended to an external macro, packaged software, or other hardware and software; the mathematical logic and geometric logic here can be extended to other mapping logics (such as related mapping logic required by the external agent input).

[0064] To facilitate understanding of this embodiment, a large model correction method disclosed in the embodiment of the present disclosure is first introduced in detail. The execution subject of the large model correction method provided in the embodiment of the present disclosure is generally an electronic device with certain computing capabilities, and the electronic device includes, for example: a terminal device or a server or other processing device, and 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 implementations, the large model correction method can be implemented by a processor calling a computer-readable instruction stored in a memory.

[0065] See also Figure 1 , which shows a flow chart of a large model correction method provided by an embodiment of the present disclosure, the method comprising the following steps S101-S102:

[0066] S101: Obtaining an original learning data set preset for a large model;

[0067] S102: Based on at least one of the pre-built geometric logic learning module, mathematical logic learning module, and Agent instruction learning module, the algorithm and original learning data set of the large model are respectively corrected 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 comply with the processing logic of the corresponding learning module, and can make the trained large model meet the needs of iterative operations.

[0068] In order to facilitate the understanding of the large model correction method provided by the embodiment of the present disclosure, the application scenario of the method is briefly introduced below. The large model correction method provided by the embodiment of the present disclosure can be mainly used in the field of artificial intelligence technology, for example, it can be applied to various artificial intelligence sub-fields such as image recognition and face detection, without specific restrictions. Especially in the large model learning stage and the correction stage after going online, the corrected large model has better model performance, so that it can adapt to more complex and changeable downstream application scenarios.

[0069] When the original learning data set preset for the large model is obtained, the disclosed embodiment can calibrate the algorithm and the original learning data set of the large model based on at least one of the pre-built geometric logic learning module, mathematical logic learning module, and Agent instruction learning module. Here, the learning data is corrected 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 can also make the trained large model meet the needs of iterative learning and be more applicable. At the same time, through the learning and correction of the Agent instruction learning module and other Promote modules, each iteration of the large model can meet the processing logic or input requirements of the external Agent, connect to multiple external Agents at the same time, form specified internal and external loops, and sequential logic deduction, and use the advantages of various Agents and external software to apply more diverse outputs.

[0070] The disclosed embodiment not only calibrates the algorithm of the large model, but also calibrates the original learning data set itself, so that the calibrated large model algorithm and the calibrated learning data set conform to the processing logic of the corresponding learning module, and can make the trained large model meet the requirements of iterative operations.

[0071] The geometric logic learning module is mainly used to emphasize that the correction results conform to the geometric logic sequence relationship, the mathematical logic learning module is mainly used to emphasize that the correction results conform to the mathematical logic sequence relationship, and the Agent instruction learning module is mainly used to emphasize that the correction results conform to the Agent instruction sequence relationship.

[0072] The algorithms and / or learning data sets obtained through correction based on the above-mentioned various learning modules can better meet the needs of iterative operations. This is mainly due to the fact that in the prior art, when training large models, convolution migration and other situations may often cause the inability to meet iterative operations, resulting in high output accuracy. Based on this, the above-mentioned correction scheme guided by geometric logic learning, mathematical logic learning and Agent instruction learning is provided to improve the accuracy of subsequent large model outputs.

[0073] Taking into account the different impacts of the above-mentioned learning modules on the correction process, the large model correction method provided by the embodiment of the present disclosure will be specifically described in combination with each learning module.

[0074] First, in the actual process of constructing the geometric logic learning module, the following steps are included:

[0075] Step 1: Obtain the first factor relationship used to characterize the geometric logic order relationship; the geometric logic order relationship is extracted by the Agent model after completing the verification of the screened and designed learning data set;

[0076] Step 2: extracting a compatible algorithm from the Agent model algorithm based on the first factor relationship and extracting corresponding learning data from the data set corresponding to the Agent model.

[0077] The first factor relationship here can be a factor relationship that conforms to the geometric logic order relationship. For example, the learning data set consists of various data such as "3 squares are extended from a right triangle, two right-angled sides are painted with the same color squares, and the third side squares are painted with different colors". After being extracted by the Agent model, if the learning data set has the specified factor relationship covariance of "right triangle" and "area of ​​squares connected by two sides (blue) = area of ​​squares connected by the third side (red)" (corresponding to the above first factor relationship), then the Pythagorean theorem 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.

[0078] In addition to the Pythagorean theorem operation logic, area fitting can be achieved using the Bayesian algorithm random point number, the same number of red and blue points and other operation logics.

[0079] For example, the learning data set consists of various images with triangles of different shapes with the same base and height. At this time, if the "factor relationship" includes "triangle" (3 convex points), "same area" (Bayesian algorithm fitting, all triangles with the same base and height, the same number of random points have the same area), "equal height" (the third point of linear regression is fitted on a line, and the eigenvalue of this line vector is fitted as a line parallel to the third side). If the learning data is a continuous data learning data set of equal-height triangles (A》B), A is a triangle of different shapes with the same base and height, and B is a right-angled triangle with the same base and height, then it can be used as a Promote instruction learning data set, and can be used to convert the triangle into a right-angled triangle with the same base, and logical Promote instruction learning.

[0080] For example, the three angles of an equilateral triangle are equal, and the factor relationships include "triangle" (3 convex point tests), the three sides of the triangle are equal (Bayesian test), and the three angles are equal (angle fitting such as vector sin+icos fitting and matrix eigenvalue fitting) appear at the same time.

[0081] The geometric logic order relationship in the first factor relationship includes, but is not limited to, determination based on Bayesian area test, sin+icos fitting, Fourier transform convolution, matrix equivalent transformation convolution fitting, convex point test, etc., as specifically described as follows:

[0082] Convex point test (turning point) is generally done by using differential algorithm. When there is a significant differential difference and then it stabilizes, it is considered as a turning point (it can be a right angle, an obtuse angle, or an acute angle), or it can be differentiated again (arc angle). Bayesian area test can also be used for length test. That is, randomly select points on the image. If the number of points selected on the two images is the same, it means that the two areas are similar. If the two lines in the image are widened into a rectangle, it can also be used to verify the length approximation; sin+icos fitting, sin+icos can refer to any point in the plane space, and is particularly suitable for fitting some natural graphics such as spirals 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 are mainly approximation algorithms; in addition, there are irrational number fitting such as pi and e, vector conversion, Bayesian similarity modules, etc., which can be flexibly used in the equivalent transformation logic, which will not be elaborated here.

[0083] Based on the geometric logic learning module constructed above, the algorithm of the large model and the original learning data set can be calibrated according to the following steps:

[0084] Step 1: Input the learning data extracted from the data set corresponding to the Agent model into the large model for iterative calculation to determine whether the output result conforms to the geometric logical order relationship in the first factor relationship;

[0085] Step 2: 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 calculation to determine whether the geometric logical sequence relationship in the first factor relationship appears;

[0086] Step 3: If the geometric logic order relationship in the first factor relationship is met, a geometric logic Promote relationship label is added to the corresponding target learning material; if the geometric logic order relationship in the first factor relationship is not met, the learning material that does not meet the first factor relationship in the original learning material set of the large model is deleted, and / or the large model algorithm is added.

[0087] The geometric logic here is mostly equivalent transformation logic, such as the same area of ​​equal-height triangles, Pythagorean theorem, proportional lengths of similar triangles, topology, etc. If there is a corresponding algorithm module and a suitable "learning data set", the specified factor relationship (i.e., the first factor relationship) of these theorems will covariate (i.e., "covariant factor relationship"). Based on this, the learning data extracted from the data set corresponding to the Agent model can be mixed with the original learning data set preset by the large model and input into the large model for iterative calculation to determine whether the geometric logic sequence relationship in the first factor relationship appears, and exclude the data set that does not conform to the geometric theorem. At this time, the large model completes the learning and confirmation (correction) process of the geometric equivalent model.

[0088] Since the large model correction method provided by the embodiment of the present disclosure excludes data sets that do not conform to geometric logic, the data sets of the entire model now conform to the fitted geometric theorem. For example, if there is a fitted heart line figure, then the picture data in the data set, like a round cup exposed to the sun, will have a heart line pattern of light and shadow on the cup mouth. At the same time, the large model data set will be marked with the corresponding equivalent factor relationship label (i.e., geometric logic Promote relationship label), so that the large model can quickly and effectively produce logical output.

[0089] Equivalent geometric changes do not necessarily have a time series relationship, but rather are factors that covary or coexist. If it is an equivalent transformation logic, sometimes the output can be filtered through label screening (but may be limited by the data set), that is, through the correction of the geometric equivalent factor relationship, and then the correction of the large model data set, it can have a great impact on the logic of the large model output (by increasing the algorithm of the large model, increasing the richness of the verification of the large model, and using the learning data set as an aid to learn).

[0090] Here, the disclosed embodiment can also be used to conduct System Promote instruction design learning, establish some commonly used equivalent transformation directions, and use the covariation factors of the output and input of different learning data sets to achieve the purpose of logical continuous deduction. Further optimize and expand the logical deduction capabilities and scope of large models. The embedded geometric System Promote still has the ability to iteratively emerge more logical deduction applications, and has the opportunity to fit a ductile compatible factor relationship, while having a higher model iteration stability.

[0091] The large model learning correction of geometry and mathematical logic is a mapping from equivalent logic learning correction to logic SystemPromote learning correction. Through the correction of logically correct data sets and the learning of continuously deducible logic SystemPromote instructions, the characteristics of large model learning correction related to geometry and mathematical logic are extracted. In geometry learning correction, different equivalent transformation factor relationships appear simultaneously to perform learning correction, which is also used as the learning basis for the subsequent commonly used equivalent deduction logic System Promote.

[0092] The second aspect: In the actual process of constructing the mathematical logic learning module, the specific steps include:

[0093] Step 1: Obtain the second factor relationship used to represent the mathematical logic sequence relationship; the mathematical logic sequence relationship is extracted by the Agent after completing the verification of the screened and designed learning data set;

[0094] Step 2: extracting a compatible algorithm from the Agent model algorithm based on the second factor relationship and extracting corresponding learning data from the data set corresponding to the Agent model.

[0095] The second factor relationship here can be a factor relationship that conforms to the order relationship of mathematical logic, and the question bank for students to learn mathematics is used as an example of a learning data set. The data set includes time series image data, and each image is attached with corresponding formulas (omitted =) in addition to equivalent data. For example, the data set before the addition data set includes the following groups of data: single-digit addition without carry (data set 1), and data set of addition without carry after carry to tens (data set 2), where data set 1 is single-digit images such as 1, 2, 3, and 4. Fitting the emergent carry logic and convolution layer only requires a limited data set, but it is extensional and sufficient to perform addition exceeding the number of digits in this data set.

[0096] For example, (sample A first, sample B later) 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, we first analyze the data, determine the minimum sample number based on the second factor relationship, and then find the factor sample type based on the cluster relationship, determine the minimum sample number based on the estimation, and determine the logical factor type.

[0097] For example, here there are three types of causes: 123+89, 0132+0080, and 0122+10+0080, and the types of effects are 0132+0080, 0122+10+0080, and 0212+0000. The data of the types of causes need to meet the minimum number of samples, and the data of the effects also need to meet the minimum number of samples, so that the mathematical logic order relationship can be identified, and then the Markov chain model of the three types of causes can be fitted. If there is a type that the data cause does not have, the logical calculation ends, and then the difference in factors from effect to cause appears to be concentrated in a specific type and the stability reaches 100% or a specific proportion, then logical ductility appears.

[0098] The algorithm here is the fitting result of the above mapping relationship calculation. At this time, it can effectively and stably fit the specified factor relationship that can be extended through convolution.

[0099] Among them, the addition learning data set and the algorithm can form or compose different compatible factor relationships, as long as they are logically closed and extended in combination with convolution. For example, because the carry of single-digit addition will not exceed 1 digit, but when there are consecutive 9999 after the carry, two data sets with non-9 carry and consecutive 9 after the carry can be established, for example: including {(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 matching data set with the previous digit being 9 and consecutive carry.

[0100] Among them, {(9998+37->09998+37), (09998+37->10005+30).....} can be included as a data set, and the corresponding two types of logical operations can be fitted by cluster analysis. The algorithm is implemented by mapping relationship calculation fitting. If the fitting is successful and two sets of logical sequential relationships appear, the specified factor relationship that can be extended through convolution can be effectively and stably fitted. Based on this logic, the corresponding learning data set is extracted and fitted into two types of mathematical mapping logic and convolution combinations, which can produce circular inference extension through closed logical relationships.

[0101] It should be noted that the compatible algorithm here contains specialized algorithms corresponding to the learning data set, thus forming a combination of mathematical mapping logic and convolution. However, the learning data set that does not achieve the above combination can still have the mathematical logic capabilities limited by the general data set of the external agent (for example, only a few additions or multiplications). At this time, by embedding mathematical logic into the large model and fitting a scalable compatible factor relationship, the following criteria need to be met:

[0102] First, the learning materials should be composed of the data within the large model. For example, it is best if 1+1 is composed of existing images in the data set within the large model. Secondly, the quality of the data must meet certain standards to facilitate calibration. Furthermore, for A-》B sequential data similar to the System Promote learning data set, A and B must form a logical closure. Here, logical closure means that when B is input as a SystemPromote instruction, there can be corresponding data in the A data set, and the data not included in A is the final answer of the logical deduction. It can be extended through convolution (such as the carry of addition and the logical relationship of the ten-breaking method of subtraction can be extended through convolution). Finally, if the fitting here is a non-continuous mapping relationship, it can avoid the influence of model iteration and drift to the greatest extent. At the same time, the calculation fitting here can effectively and stably fit the mathematical logic sequence relationship that can be extended through convolution.

[0103] Based on the mathematical logic learning module constructed above, the algorithm of the large model and the original learning data set can be calibrated according to the following steps:

[0104] Step 1: Input the learning data extracted from the data set corresponding to the Agent model into the large model for iterative calculation to determine whether the output result conforms to the mathematical logic sequence relationship in the second factor relationship;

[0105] Step 2: 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 calculation to determine whether the mathematical logic sequence relationship in the first factor relationship appears;

[0106] Step 3: 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 material; if the second factor relationship is not met, the learning material that does not meet the second factor relationship in the original learning material set of the large model is deleted, and / or the large model algorithm is added.

[0107] Similar to the process of achieving correction based on geometric logical sequence relationships, since the large model correction method provided by the embodiment of the present disclosure excludes data sets that do not conform to mathematical logic, the data set of the entire model now conforms to the fitted mathematical logic; at the same time, the large model data set will be marked with corresponding equivalent factor relationship labels (i.e., mathematical logic Promote relationship labels), so that logical output can be made quickly and effectively with the large model.

[0108] The third aspect: In the actual process of constructing the Agent instruction learning module, the specific steps include:

[0109] Step 1: Obtain the third factor relationship used to characterize the sequential relationship of Agent instructions; the sequential relationship of Agent instructions is extracted by the large model connected to the external Agent model after completing the verification of the learning data set that has been screened and designed to meet the requirements of connecting to the external Agent model;

[0110] Step 2: Based on the third factor relationship, a compatible algorithm is extracted from the large model connected to the external Agent model and corresponding learning data is extracted from the data set of the connected large model.

[0111] The Agent command sequence 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 corresponding input of the external Agent, the data format meets the requirements of the external Agent, C is the corresponding input of the external Agent-》D is the internal data set format of the large model, or the Promote command format.

[0112] Based on the Agent instruction learning module constructed above, the algorithm of the large model and the original learning data set can be calibrated according to the following steps:

[0113] Step 1: Input the learning data extracted from the data set of the large model connected to the external Agent model into the large model for iterative calculation to determine whether the output result meets the third factor relationship of the Agent instruction learning module;

[0114] Step 2: If the third factor relationship of the Agent instruction learning module is met, then the Agent instruction Promote relationship label is added to the corresponding target learning material; if the third factor relationship is not met, then the learning material that does not meet the third factor relationship in the original learning material set is deleted, and / or the large model algorithm is added.

[0115] Similar to the process of achieving correction based on geometric logic sequence relationship, that is, mathematical logic sequence relationship, since the large model correction method provided by the embodiment of the present disclosure excludes the data set that does not conform to the Agent instruction sequence relationship, the data set of the entire model now conforms to the fitted Agent instruction sequence; at the same time, the large model data set will be marked with the corresponding equivalent factor relationship label (that is, the Agent instruction Promote relationship label), so that a logical output can be made quickly and effectively with the large model.

[0116] Regardless of which of the above calibration methods is used, in the actual calibration process, the large model calibration method provided by the embodiment of the present disclosure can also add a compatible algorithm or a specialized algorithm required for the factor relationship, and determine the order relationship according to one or more of the following methods:

[0117] First, the learning materials in the original learning material set are grouped to obtain multiple learning material groups; statistical tests are performed on the multiple learning material groups to determine whether the factor relationships or corresponding factor values ​​of each group are significantly different, and whether they conform to the corresponding geometric logic, mathematical logic, and Agent instruction logic judgment results;

[0118] Second, the sequential relationship is determined based on a target algorithm with causal inference or time or sequential data fitting and verification characteristics, and the target algorithm includes one of the following: Bayesian algorithm, Markov chain algorithm, decision tree regression, random forest regression, breakpoint test, and T test.

[0119] Here, an example is given of an "instruction learning sequence data set" consisting of pictures of Arabic numerals corresponding to "numbers 1-9".

[0120] Among them, specify all Arabic numeral picture sets, and specify the corresponding test method of cluster fitting, such as T test). From the large model factor relationship table, find the covariate factor relationship, exclude 0 and unchanged or random factors.

[0121] The data is divided into groups (1, 2, 3), (4, 5, 6), (1, 2, 3, 12, 23), (45, 54, 64), and similarity factors and 0 (no correlation factor) are excluded. Each group needs to meet the minimum number of samples. At this time, if there is a one-to-one 100% mapping relationship between 0 and 1-9 Arabic numerals, it can be determined through clustering data.

[0122] If a one-to-one 100% mapping relationship (factor relationship mapping characteristics) cannot be formed, then see which numbers are not working and increase the learning data set. For example, 9 and 8, 7 and 1 are similar in shape. Learn to recognize them by discovering errors and practicing many times. The fitting here is that the eigenvalues ​​are not obvious or are not convoluted. Increase the number of samples and then fit the large model.

[0123] Among them, the "A-》B Markov factor relationship" used for correction can be divided into the factor relationship characteristics of A, the factor relationship characteristics of B, and then the A-》B mapping relationship. The A-》B mapping relationship can fit the corresponding mapping logic relationship after A grouping and B grouping (such as Markov chain fitting, which can determine the mapping relationship characteristics of A and B).

[0124] The "simple factor relationship" here means that the algorithm needs to fit significantly and pass the test based on the System Promote instruction. The "logical order factor relationship (the specified factor relationship here)" is to fit significantly and pass the test, and have the specified logical characteristics, such as the corresponding number is a 100% mapping relationship.

[0125] The algorithm here can be a method of fitting and testing time or sequential data, such as Markov (Bayesian) fitting, decision tree regression, random forest regression, and corresponding tests (such as breakpoint test, T test), etc. Many large models have corresponding modules. In the embodiments of the present disclosure, the main emphasis is on correcting or adding algorithms that are compatible with the learning data or the calibration procedure.

[0126] For the same (or compatible) algorithm, when the System Promote instruction uses the learning data set as the learning data set, after iterative operation of the aforementioned algorithm, a compatible "logical order factor relationship (here the specified factor relationship)" will be run out to complete the learning and correction.

[0127] For example, the relationship between the factor changes of the previous and subsequent data of the Markov fitting time series can be a functional relationship or a linkage change relationship of the factors. After grouping, the probability, continuous function or mapping change of different changes can be fitted. If it is a logical mapping change, then the "logical order factor relationship (here the specified factor relationship)" feature corresponds to a 100% mapping relationship or all possible changes.

[0128] If no sequential relationship emerges, then the learning data set is increased and / or the specific algorithm is modified.

[0129] Regarding the learning and confirmation of geometric logic, the geometric logic specifies the factor relationship as a "covariant factor relationship"; regarding the learning and confirmation of mathematical logic, the mathematical logic specifies the factor relationship as a "time series relationship", and its time series relationship can be formed through the composition of convolutional layers and other statistical algorithms to form a compatible specified factor relationship. If the iterative learning of the above logics also has the following technical features, it can increase the effectiveness and stability of the logic:

[0130] One is to form a logical mapping relationship. The "specified factor relationship" is a specified logical feature after the fitting is significant and passes the test, such as the corresponding number is a 100% mapping relationship.

[0131] The second is to specify a simple factor relationship algorithm and correction. Simple factor relationship, "specified factor relationship" embodiment: The "A-》B Markov factor relationship" used for correction can be divided into the factor relationship characteristics of A, the factor relationship characteristics of B, and then the A-》B mapping relationship. The A-》B mapping relationship can fit the corresponding mapping logic relationship after A is grouped and B is grouped.

[0132] At this time, the "simple factor relationship" means that the algorithm needs to be significantly fitted and passed the test based on the Promote instruction. The compatible algorithm (extensible convolution and logical combination) is designed and modified by specifying the simple factor relationship, so that the originally designed "logical order factor relationship" convolution and mapping logical combination can be easily fitted.

[0133] The third is specialized algorithms such as convolutional layer combinations. Use specified cluster comparisons to determine whether factors and logical factors are extensional and whether there are compatible specific modules. Specifying simple factor relationships can only be related, but specialized algorithms refer to a specific combination of convolution and regression relationships in compatible relationships.

[0134] The large model correction method provided in the embodiment of the present disclosure can also be combined with an external Agent model to implement cyclic or sequential logic deduction, which specifically includes the following steps:

[0135] Step 1: Obtain a Promote learning data set and its factor relationships with Promote relationship tags added thereto; the Promote relationship tags include at least one of a geometric logic Promote relationship tag, a mathematical logic Promote relationship tag, and an agent instruction Promote relationship tag; the Promote relationship tag includes a logic Promote that emerges after correction, or other logic Promote that has been confirmed by an external agent;

[0136] Step 2: Determine whether the factor relationship of the data set output by calculating any Promote learning data in the Promote learning data set overlaps or coexists with the input factor relationship of itself or other Promote learning data;

[0137] Step 3: If yes, determine the logical deduction between each Promote learning material to form a cycle or sequence.

[0138] After the relevant fitting modules are placed, according to the definition of the big model, with appropriate modules and sufficient data, equivalent logic should emerge. However, in actual applications, some factors may cause the logic not to emerge, or other system or non-system problems may cause the relationship to emerge. In addition, some may be caused by systemic problems such as data format, so reconfirmation is required.

[0139] For example, after the algorithm is used to train the entire large model data set, if other covariate factor relationships emerge, the atlas can be output for confirmation. After confirmation, the learning data set can be output for future model calibration or stabilization, and can also be used to design the System Promote learning data set.

[0140] At this time, it is also possible to calculate other logical sequence relationships through the sequence data of the large model data set, and provide manual confirmation whether it is a new System Promote instruction that has emerged.

[0141] Regarding the confirmation of the emerging underlying relationship, this is because the underlying layer of the large model is quantitative data and mathematical logic algorithms, so it is easy to have a large number of significant continuous and discontinuous changes. If the Markov chain (Bayesian algorithm) is used for fitting, various continuous data logics may appear. For example, A1->A2 (10%), A3 (90%), at this time (A1->A2, A1->A3) is composed of (1:9).

[0142] The mapping logic here is not easily affected by model iteration, and generally a large number of "overfitted" logical relationships can be excluded. Therefore, other relationships fitted by the mapping logic-related fitting method can be manually confirmed and used for subsequent model stabilization and other model correction.

[0143] Here, the emergence relationship and Promote design can be integrated. The continuity of geometric logic relies on the factor relationship of the output learning data set, which is the A factor relationship of A-》B of another Promote or geometric logic data set. Therefore, the factor relationship forms coexistence, but the data sets may or may not be the same.

[0144] However, the continuous part of mathematical logic relies on the overlap of the output learning data set and the input learning data set and the coordination of the convolutional layer, such as single-digit addition and carry; and the Agent docking data set is a logical mapping relationship, which is mainly overlap. If cross-deduction is required, or as long as these factor relationships exist, even if they are not designed or the same learning data set, if such relationships emerge after iteration, a cycle or sequential deduction will be formed in the large model.

[0145] Mathematical logic can form an extendable circular deduction through algorithmic convolution and specific factor relationships. For example, when two sets of logical sequential relationships appear through additive fitting, the specified factor relationship that can be extended through convolution can be effectively and stably fitted.

[0146] In addition, most of the geometric logic is equivalent transformation logic. According to the current situation of the big model, with appropriate fitting modules and sufficient data, geometric equivalent logic will emerge. That is, through various equivalent tests and fittings, the equivalent logic of covariance can be determined, and then the big model logic can be naturally embedded, so that it can be integrated into the big model logic, such as the same area of ​​equal-height triangles.

[0147] In the disclosed embodiment, the emergent relationship can use the same learning correction method as the aforementioned various logics, such as proposing a factor relationship data set with a Bayesian probability of 95%. After confirming this logical relationship, the large model learning data that affects the inability to form a 100% mapping relationship is removed to complete the correction.

[0148] When actually conducting logical deduction, one can first combine new learning data sets and factor relationships according to a certain cyclic or sequential logical deduction direction to form a new cyclic or sequential logical deduction direction; the second is to form a large model according to the new cyclic or sequential logical deduction direction and in coordination with the input sequence of different Agent models, and simultaneously connect with multiple Agents to conduct corresponding cyclic or sequential logical deduction; the third is to modify all or part of the Promote learning data so that the output factor relationship obtained by calculating any Promote learning data does not overlap or coexist with the input factor relationship of itself or other Promote learning data, so as to interrupt or terminate the cyclic or sequential logical deduction; the fourth is to form or interrupt the cyclic or sequential logical deduction direction of the large model by combining learning data sets.

[0149] In the disclosed embodiment, a geometric logic System Promote learning data set can be established based on common geometric logic deduction requirements, that is, a geometric logic System Promote is established. Different geometric logic System Promote learning data sets have the same covariant "logical order factor relationship" (herein, the specified factor relationship) between the input and output.

[0150] At this point, the two geometric logic System Promotes can form a continuous logic deduction (perform geometric equivalent logic deduction). The implementation example of establishing the geometric logic System Promote is as follows:

[0151] Establish System Promote A-》B, System Promote C-》D, where A and B are equivalent data, C and D are equivalent data, but B has the covariation characteristics of C and D on the basis of being equivalent to A. In this way, continuous geometric equivalent logic deduction can be achieved through learning and correction of System Promote instructions. For example, in System Promote learning data A-》B, A and B are equal-height triangles, and B is a right triangle. In System Promote learning data C-》D, C is a right triangle, and D is an equal-base rectangle with half the height (equal area). If the large-model learning and correction of the above learning data set is completed, you can now input "a triangle and the specified final System Promote (equal-area rectangle)" into the large model, thereby realizing the logical deduction of transforming a triangle into an equal-area rectangle.

[0152] The mathematical logic System Promote algorithm here includes a convolutional layer and a time series relationship algorithm, where the time series algorithm can be a Bayesian algorithm or a Markov chain algorithm (both have existing functions). The time series factor relationship between the convolutional layer and the time series algorithm forms a compatible specified factor relationship, and learning and correction are completed.

[0153] Example 1: For example, if the learning data set is A--》B, after the Bayesian algorithm is fitted, when the data set A is an equal-height triangle, the data set B is a corresponding equal-height right triangle. After the Bayesian algorithm is fitted, the data set changes from the A covariate factor relationship to the "A covariate factor relationship + right angle" time series Bayesian probability is 100%.

[0154] Example 2: For example, if the learning data set A--》B, A is composed of a stepped 1-digit convolution layer and a single-digit addition (such as 1+1, 9+8), and B is composed of a stepped 2-digit convolution layer and the result of a single-digit addition (2, 18), after the Bayesian algorithm is fitted, the data set changes from "A covariate factor relationship" to a time series with a Bayesian probability of 100% for the answer, that is, if 1+1-》2, the Bayesian probability of the time series is 100%.

[0155] For the logic deduction of external Agent Promote, data sets A and B are logical Promote, and (text C—》logic D) conforms to the logical D output. Through the specified embedded logic fitting, the order of external Agent and input data conversion Promote, the data set data format conversion, or the built-in order and data set data format conversion, the continuous logic automatic fitting of the embedded logic module and the external Agent is established when the data is input. At this time, it can be explained that Agent and embedding each have their own advantages. If they can be used in a compatible order, the best effect can be achieved.

[0156] In addition, for the establishment of external agent and input data conversion System Promote, external agent and output data conversion System Promote, such as external agent and input data conversion System Promote data set (A-》B1), external agent and input data output conversion System Promote data set (C1-》D1), A is the internal data set of the large model, B1 is the corresponding input requirement of the external agent, the data format meets the requirements of the external agent, and can start the calculation of the external agent.

[0157] Among them, C1 is the corresponding output of the external Agent, and D1 is the corresponding internal data set format of the large model, or the SystemPromote instruction format. Therefore, inputting D1 into the large model will form a circular deduction.

[0158] For the logic deduction of multiple external Agent System Promote, such as the logical System Promote logic deduction sequence data set (A-》B2). B2 belongs to the compatible data set A, which will form a circular deduction. When B2 meets the input of another logic SystemPromote (text C2—》logic D2), the logic D2 output is formed. Therefore, the order of System Promote can be converted by specifying embedded logic fitting, external Agent and input data, by data set data format conversion, or built-in sequence and data set data format conversion, to establish the continuous logic automatic fitting of embedded logic and external Agent when data is input.

[0159] Among them, the priority calculation data is transformed based on the built-in logic deduction order, and the format of the data conforms to the input of the external agent or 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, it will not start even if it is first input into the external agent, and the Roman numerals will be input into the large model to form a logical deduction.

[0160] In the disclosed embodiment, the built-in logic deduction order factor table can be used to convert the logic instructions of the external agent in priority to the embedded equivalent logic, and then output to various external agents first. Whether completed or not, they are converted back through the external agent and the output data conversion system Promote, and then fitted with the built-in logic.

[0161] If A-》may be B1 or B2, the external agent is input in sequence first. If A-》B1 is input, the external agent will be started, and after the output conversion, it will be input into the large model with D1. If A-》B2 is input, the agent will not be started, so B2 will be input into the large model again, forming a logical System Promote logical deduction sequence.

[0162] 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, and the System Promote loop logic has no restrictions, at this time, 10+10 and 10000+10000 can be input into the System Promote loop logic, but because the Agent is input first, the Agent will output 20 and 10000+10000 to the large model, and the large model will then output 20 and 20000.

[0163] In the disclosed embodiment, System Promote instruction learning data sets of different logics can also be mixed. For example, in order to be compatible with various external and embedded modules, multiple layers may be required, or a "logical deduction sequence learning data set" may be required. For another example, in order to be compatible with the input forms of various embedded and external logic modules, it is used in the output and re-input of System Promote instructions. For example, in the "Digital and Code Conversion Learning Sequence Data Set" (see equivalent logic), when the code input external module data does not match, it is converted into a digital format and then input through the "Digital and Code Conversion Learning Module". The "Digital and Code Conversion Learning Data Set" is also a part of the System Promote "System Promote Instruction Learning Data Set", and the correction is completed using the same algorithm and correction method. It can be seen that the entire large model System Promote instruction can automatically perform logical deduction.

[0164] 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 subsequent corresponding System Promote instruction and logical element instruction. B needs to be compatible or correspond to the learning data format requirements of input A of the "equivalent logic and sequential logic learning data set" or the input A of the logic agent.

[0165] A and B can be designed in multiple layers, with the first layer being the System Promote instruction learning dataset. B should include as much usage-oriented design as possible for people, things, time, and places, which can maximize the subsequent application of automatic logic deduction. It also needs to be compatible with the instructions required by the Agent, or compatible with the large model logic meta-instructions. The quality of the A dataset needs to be able to identify specific datasets, but does not need to cover all data categories. Here, "zero-sample learning" or large model iterative evolution can be used to expand the recognition type.

[0166] In addition, in the first-level deduction of the large model System Promote learning data set, it is recommended that B has elements such as "people, things, time, place, and objects" so that System Promote can be used to perform more generalized automatic logical deduction of various texts in the future.

[0167] In the disclosed embodiment, different logical System Promote "instruction learning sequence data sets" can also be mixed. For example, the A-》B mapping relationship is used as an example. Part A of the System Promote "instruction learning sequence data set" about addition can be a Chinese text question bank about addition for primary school students. The artificially marked time-continuous result B needs to be compatible with the "instructions required by the Agent" or compatible with the "large model logic element instructions".

[0168] For example, A("There are 3 apples and 2 oranges on the table at home in the morning. If 2 apples are eaten, how many apples are left?")-》B(morning (time), home (place), 3-2 apples (things)).

[0169] In the "Digital and Code Conversion Learning Sequential Data Set", if the Agent needs data in a certain format, such as "There is..."》5-3 (picture), the data format needs to be compatible with the instruction "5-3" (text) required by the Agent, or compatible with the large model logic element instruction. If text is the input element of the Agent, a layer of System Promote learning data set can be added if necessary, 5-3 (picture)》5-3 (text).

[0170] The "sequential logic learning data set" includes: 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 must meet the requirements of continuous input for System Promote instructions.

[0171] Then, through "zero-data learning" and "few-sample learning", the large model iteration price logic expansion is realized. Input A ("Based on the following morning photos and dialogues, how many apples are there at home in the morning?", dialogue "There are 3 apples on the table, 2 oranges eat 2 apples" and there is also a photo of an apple tree in front of the house with three apples on it) -》This question drives the recognition to form B (morning (time), home (place), 3-2+1+1+1 apples (things)". At this time, the logic program of the loop input System Promote is entered, and the calculation logic order is first identified, and the order is identified, and the mathematical logic result fitting is entered in this order. Finally, 4 apples are output.

[0172] The disclosed embodiment can also set a "logic deduction sequential learning data set" to learn to get inputs such as "1+13*2". At this time, you can first execute the multiplication logic System Promote instruction, or intercept the 13*2 input Agent, and if necessary, there must be an Agent output how to embed the "digital and code conversion learning module" back into the System Promote instruction, so as to form a more generalized logic automatic deduction; or you can directly output to Excel-Pyron for drawing; or then use the convolution layer to transform the fitting logic deduction results... and other operations.

[0173] The logic System Promote (text-->logic), logic System Promote (logic deduction order), geometry and mathematical logic (System Promote), external agent and input data conversion System Promote, and external agent and output data conversion System Promote in the embodiments of the present disclosure are only for forming a System Promote loop deduction, that is, each data set is (A-->B), B is compatible with A or other System Promote input data sets, so it can form a loop deduction when re-input.

[0174] Among them, for logical System Promote (text -> logic), logical System Promote (logical deduction order), logical System Promote (logical deduction order), external agent and input data conversion System Promote, external agent and output data conversion System Promote, generally they all confirm whether a "time sequence factor relationship" is formed.

[0175] In the part of connecting to the external agent, after correcting the "time sequence factor relationship", further confirm the "corresponding factor value". Taking numbers as an example, the corresponding non-continuous 1 to 1 factor relationship, such as 1 corresponds to a certain factor L, which means that Bayesian 100% correspondence is determined.

[0176] In summary, the disclosed embodiments use one or more learning logics to perform multi-dimensional corrections on learning materials, which meets the current requirements for large model output accuracy in the field of artificial intelligence, and can also make the trained large model meet the needs of iterative learning and be more applicable. At the same time, through the learning and correction of the Agent instruction learning module and other Promote modules, each large model iteration can meet the processing logic or input requirements of the external Agent, connect to multiple external Agents at the same time, form specified internal and external loops, and sequential logic deductions, and use the advantages of various Agents and external software to apply more diverse outputs.

[0177] In the description of this specification, the description with reference to the terms "some possible embodiments", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure, and the above terms do not necessarily represent the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0178] About the method flow chart of the embodiment of the present disclosure, some operations are described as different steps performed in a certain order. Such a flow chart is illustrative and not restrictive. Some steps described in this article can be grouped together and performed in a single operation, or some steps can be divided into multiple sub-steps, and some steps can be performed in an order different from that shown in this article. The various steps shown in the flow chart 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, a logical function implemented by a processor or a chip), etc., and / or any combination thereof).

[0179] Those skilled in the art will appreciate that, in the method described in the above specific implementation, 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 internal logic.

[0180] Based on the same inventive concept, a large model correction device corresponding to the large model correction method is also provided in the embodiment of the present disclosure. Since the principle of solving the problem by the device in the embodiment of the present disclosure is similar to the above-mentioned large model correction method in the embodiment of the present disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0181] Reference Figure 2 , which is a schematic diagram of a large model correction device provided by an embodiment of the present disclosure, the device includes: an acquisition module, a correction module 202; wherein,

[0182] An acquisition module 201 is used to acquire an original learning data set preset for a large model;

[0183] A correction module 202, used to correct the algorithm of the large model and the original learning data set based on at least one of the pre-built 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;

[0184] Among them, the calibrated large model algorithm and the calibrated learning data set are used to comply with the processing logic of the corresponding learning module, and can make the trained large model meet the needs of iterative operations.

[0185] By adopting the above-mentioned large model correction device, when the original learning data set preset for the large model is obtained, the algorithm and the original learning data set of the large model can be corrected based on at least one of the pre-built geometric logic learning module, mathematical logic learning module, and 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 through one or more learning logics here, the corrected learning data set can not only conform to the processing logic of the learning module, but also enable the trained large model to meet the needs of iterative operations. This meets the current requirements for the output accuracy of large models in the field of artificial intelligence and is more applicable.

[0186] In a possible implementation, the correction module 202 is used to construct a geometric logic learning module according to the following steps:

[0187] Obtaining the first factor relationship used to characterize the geometric logic order relationship; the geometric logic order relationship is extracted by the Agent model after completing the verification of the screened and designed learning data set;

[0188] Based on the first factor relationship, a compatible algorithm is extracted from the Agent model algorithm and corresponding learning data is extracted from the data set corresponding to the Agent model.

[0189] In a possible implementation, the correction module 202 is specifically used to calibrate the algorithm of the large model and the original learning data set respectively according to the following steps:

[0190] Input the learning data extracted from the data set corresponding to the Agent model into the large model for iterative calculation to determine whether the output result conforms to the geometric logical order relationship in the first factor relationship;

[0191] 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 input into the large model for iterative calculation to determine whether the geometric logical sequence relationship in the first factor relationship appears;

[0192] If the geometric logic order relationship in the first factor relationship is met, a geometric logic Promote relationship label is added to the corresponding target learning material;

[0193] If the geometric logical order relationship in the first factor relationship is not met, the learning materials that do not meet the first factor relationship in the original learning data set of the large model are deleted, and / or the large model algorithm is added.

[0194] In a possible implementation, the correction module 202 is used to construct a mathematical logic learning module according to the following steps:

[0195] Obtain the second factor relationship used to represent the mathematical logic sequence relationship; the mathematical logic sequence relationship is extracted by the Agent after completing the verification of the screened and designed learning data set;

[0196] Based on the second factor relationship, a compatible algorithm is extracted from the Agent model algorithm and corresponding learning data is extracted from the data set corresponding to the Agent model.

[0197] In a possible implementation, the correction module 202 is specifically used to calibrate the algorithm of the large model and the original learning data set respectively according to the following steps:

[0198] 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 sequence relationship in the second factor relationship;

[0199] 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 input into the large model for iterative calculation to determine whether the mathematical logic sequence relationship in the first factor relationship appears;

[0200] 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 material;

[0201] If the second factor relationship is not met, the learning data that does not meet the second factor relationship in the original learning data set of the large model is deleted, and / or the large model algorithm is added.

[0202] In a possible implementation, the correction module 202 is used to construct an Agent instruction learning module according to the following steps:

[0203] Obtaining a third factor relationship for characterizing the sequential relationship of Agent instructions; the sequential relationship of Agent instructions is extracted by the large model connected to the external Agent model after completing the verification of the learning data set that has been screened and designed to meet the requirements of connecting to the external Agent model;

[0204] Based on the third factor relationship, a compatible algorithm is extracted from the large model connected to the external Agent model and corresponding learning data is extracted from the data set of the connected large model.

[0205] In a possible implementation, the correction module 202 is specifically used to calibrate the algorithm of the large model and the original learning data set respectively according to the following steps:

[0206] The learning data extracted from the data set of the large model connected to 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;

[0207] If it meets the third factor relationship of the Agent instruction learning module, then add the Agent instruction Promote relationship tag to the corresponding target learning material;

[0208] If the third factor relationship is not met, the learning materials in the original learning data set that do not meet the third factor relationship are deleted, and / or the large model algorithm is added.

[0209] In a possible implementation, the correction module 202 is further configured to:

[0210] When performing correction based on factor relationships, add a compatible algorithm or a specialized algorithm required for the factor relationships; determine the order relationship in one or more of the following ways:

[0211] The learning materials in the original learning material set are grouped to obtain multiple learning material groups; statistical tests are performed on the multiple learning material groups to determine whether the factor relationships or corresponding factor values ​​of each group are significantly different, and whether they conform to the judgment results of corresponding geometric logic, mathematical logic, and Agent instruction logic;

[0212] The sequential relationship is determined based on a target algorithm with causal inference or temporal or sequential data fitting and verification 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.

[0213] In a possible implementation, it further includes:

[0214] The deduction module 203 is used to obtain a Promote learning data set with Promote relationship tags added and its factor relationships; the Promote relationship tags include at least one of a geometric logic Promote relationship tag, a mathematical logic Promote relationship tag, and an Agent instruction Promote relationship tag; the Promote relationship tag includes a logic Promote that emerges after correction, or other logic Promote that has been confirmed by an external agent; determine whether the factor relationship of the data set output by calculating any Promote learning data in the Promote learning data set overlaps or coexists with the input factor relationship of itself or other Promote learning data;

[0215] If so, determine the logical deduction between each Promote learning material to form a cycle or sequence.

[0216] In a possible implementation, the deduction module 203 is further configured to:

[0217] With a certain cyclic or sequential logical deduction direction, a new set of learning data and factor relationships are combined to form a new cyclic or sequential logical deduction direction; or,

[0218] According to the new cyclic or sequential logic deduction direction, in conjunction with the input sequence of different Agent models, a large model is formed, which is connected with multiple Agents at the same time and performs corresponding cyclic or sequential logic deduction; or,

[0219] Modify all or part of the Promote learning materials so that the output factor relationship of any Promote learning material does not overlap or coexist with the input factor relationship of itself or other Promote learning materials, so as to interrupt or terminate the cyclic or sequential logical deduction; or

[0220] A method of forming or interrupting the direction of large model cycles or sequential logical deduction by combining learning data sets.

[0221] It should be noted that the device in the embodiment of the present disclosure can implement each process of the embodiment of the aforementioned method and achieve the same effects and functions, which will not be repeated here.

[0222] The present disclosure also provides an electronic device, such as Figure 3 , which is a schematic diagram of the structure of an electronic device provided by an embodiment of the present disclosure, 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, Figure 2When the electronic device is running, the processor 301 communicates with the memory 302 via the bus 303, and when the machine-readable instructions are executed by the processor 301, the following processing is performed:

[0223] Get the original learning data set preset for large models;

[0224] Based on at least one of the pre-built 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 respectively corrected to obtain a corrected large model algorithm and a corrected learning data set;

[0225] Among them, the calibrated large model algorithm and the calibrated learning data set are used to comply with the processing logic of the corresponding learning module, and can make the trained large model meet the needs of iterative operations.

[0226] The present disclosure also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the large model correction method described in the above method embodiment are executed. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0227] The present disclosure also provides a computer program product that carries a program code. The program code includes instructions that can be used to execute the steps of the large model correction method described in the above method embodiment. For details, please refer to the above method embodiment, which will not be repeated here.

[0228] The computer program product may be implemented in hardware, software or a combination thereof. In one optional embodiment, the computer program product is implemented as a computer storage medium. In another optional embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0229] Each embodiment in the present disclosure is described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and computer-readable storage medium embodiments, since they are basically similar to the method embodiments, their descriptions are simplified, and the relevant parts can be referred to the partial description of the method embodiments.

[0230] The apparatus, equipment and computer-readable storage medium provided in the embodiments of the present disclosure correspond one-to-one to the method. Therefore, the apparatus, equipment and computer-readable storage medium also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the apparatus, equipment and computer-readable storage medium will not be repeated here.

[0231] It should be understood by those skilled in the art that the embodiments of the present disclosure can be implemented in the form of methods and apparatus (devices or systems), or computer-readable storage media. Therefore, the present disclosure can be implemented in a complete hardware implementation, a complete software implementation, or an implementation that combines software and hardware. Moreover, the present disclosure can be implemented in the form of a computer-readable storage medium implemented on one or more computer-usable storage media (including but not limited to disk storage, read-only compact disk storage (CD-ROM), optical storage, etc.) containing computer-usable program code.

[0232] The present disclosure is described with reference to the flowchart and / or block diagram of the method, device (equipment or system) and computer-readable storage medium according to the embodiment of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, and the combination of the process and / or box in the flowchart and / or block diagram 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 a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0233] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, wherein the instruction device implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0234] These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0235] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0236] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash memory (Flash RAM). The memory is an example of a computer-readable medium.

[0237] Computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, 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 disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. In addition, although the operations of the disclosed method are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. In addition, some steps can be omitted, multiple steps can be combined into one step, and / or one step can be decomposed into multiple sub-steps.

[0238] Although the spirit and principle of the present disclosure have been described above with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A large model calibration method, characterized in that: include: Get the original learning data set preset for large models; Based on at least one of the pre-built 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 respectively corrected to obtain a corrected large model algorithm and a corrected learning data set; Among them, the corrected large model algorithm and the corrected learning data set are used to comply with the processing logic of the corresponding learning module, and can make the trained large model meet the needs of iterative operations.

2. The method according to claim 1, characterized in that Follow the steps below to build the geometric logic learning module: Obtaining a first factor relationship for characterizing a geometric logic sequence relationship; the geometric logic sequence relationship is extracted by the Agent model after completing verification of a screened and designed learning data set; Based on the first factor relationship, a compatible algorithm is extracted from the Agent model algorithm and corresponding learning data is extracted from the data set corresponding to the Agent model.

3. The method according to claim 2, characterized in that The algorithm and the original learning data set of the large model are calibrated respectively, including: Inputting the learning data extracted from the data set corresponding to the Agent model into the large model for iterative calculation to determine whether the output result conforms to the geometric logic sequence relationship in the first factor relationship; Mixing the learning data extracted from the data set corresponding to the Agent model with the original learning data set preset in the large model, and inputting them into the large model for iterative calculation to determine whether the geometric logical sequence relationship in the first factor relationship appears; If the geometric logic order relationship in the first factor relationship is met, a geometric logic Promote relationship tag is added to the corresponding target learning material; If the geometric logical order relationship in the first factor relationship is not met, the learning materials that do not meet the first factor relationship in the original learning data set of the large model are deleted, and / or the large model algorithm is added.

4. The method according to claim 1, characterized in that: Follow the steps below to build the mathematical logic learning module: Obtaining a second factor relationship for representing a mathematical logic sequence relationship; the mathematical logic sequence relationship is extracted by the Agent after completing verification of a screened and designed learning data set; Based on the second factor relationship, a compatible algorithm is extracted from the Agent model algorithm and corresponding learning data is extracted from the data set corresponding to the Agent model.

5. The method according to claim 4, characterized in that The algorithm and the original learning data set of the large model are calibrated respectively, including: Inputting the learning data extracted from the data set corresponding to the Agent model into the large model for iterative calculation to determine whether the output result conforms to the mathematical logic sequence relationship in the second factor relationship; Mixing 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 inputting them into the large model for iterative calculation to determine whether the mathematical logic sequence relationship in the first factor relationship appears; If the mathematical logic order relationship in the second factor relationship is met, a mathematical logic Promote relationship tag is added to the corresponding target learning material; If the second factor relationship is not met, the learning materials that do not meet the second factor relationship in the original learning data set of the large model are deleted, and / or the large model algorithm is added.

6. The method according to claim 1, characterized in that Follow the steps below to build the Agent instruction learning module: Obtaining a third factor relationship for characterizing the sequential relationship of Agent instructions; the Agent instruction sequential relationship is extracted by the large model connected to the external Agent model after completing verification of a learning data set that has been screened and designed to meet the requirements of connecting to the external Agent model; Based on the third factor relationship, a compatible algorithm is extracted from the large model connected to the external Agent model and corresponding learning data is extracted from the data set of the connected large model.

7. The method according to claim 6, characterized in that The algorithm and the original learning data set of the large model are calibrated respectively, including: The learning data extracted from the data set of the large model connected to 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; If the third factor relationship of the Agent instruction learning module is met, then the Agent instruction Promote relationship tag is added to the corresponding target learning material; If the third factor relationship is not met, the learning materials in the original learning data set that do not meet the third factor relationship are deleted, and / or a large model algorithm is added.

8. The method according to any one of claims 3, 5 and 7, characterized in that: When making corrections based on factor relationships, this also includes: Add the required compatible algorithms or specialized algorithms for factor relations; And / or, determine the order relationship of the factor relationship according to one or more of the following methods: Grouping the learning materials in the original learning material set to obtain a plurality of grouped learning material groups; performing statistical tests on the plurality of learning material groups, and determining based on whether the factor relationships or corresponding factor values ​​of the groups are significantly different, and whether they conform to the judgment results of corresponding geometric logic, mathematical logic, and Agent instruction logic; The sequential relationship is determined based on a target algorithm with causal inference or temporal or sequential data fitting and verification characteristics, wherein 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.

9. The method according to any one of claims 1 to 7, characterized in that Also includes: Acquire a Promote learning data set and its factor relationships with Promote relationship tags added thereto; the Promote relationship tags include at least one of a geometric logic Promote relationship tag, a mathematical logic Promote relationship tag, and an Agent instruction Promote relationship tag; the Promote relationship tags include logic Promote that emerges after correction, or other logic Promote that has been confirmed by an external Agent model; Determine whether the factor relationship of the data set output by calculating any Promote learning data in the Promote learning data set overlaps or coexists with the input factor relationship of itself or other Promote learning data; If so, determine the logical deduction between each Promote learning material to form a cycle or sequence.

10. The method according to claim 9, characterized in that The method further comprises: With a certain cyclic or sequential logical deduction direction, a new set of learning data and factor relationships are combined to form a new cyclic or sequential logical deduction direction; or, According to the new cyclic or sequential logic deduction direction, in conjunction with the input order of different Agent models, a large model is formed, which is connected with multiple Agent models at the same time and performs corresponding cyclic or sequential logic deduction; or, Modify all or part of the Promote learning materials so that the output factor relationship of any Promote learning material does not overlap or coexist with the input factor relationship of itself or other Promote learning materials, so as to interrupt or terminate the cyclic or sequential logical deduction; or A method for forming or interrupting a large model cycle or sequential logic deduction direction by combining the learning data sets.

Images