Data labeling method, model training method, display method, device, storage medium and program product

Abstract

The embodiment of the present application provides a data annotation method, a model training method, a display method, a device, a storage medium and a program product, which are applicable to the field of data annotation and applied to the annotation system. The method includes: in response to the annotation request triggered for the target task, using the first large model to perform a multi-step reasoning operation for the target task; for the current reasoning step, generating the reasoning content of the reasoning step, and generating the reasoning result according to the reasoning content; outputting the reasoning content and the reasoning result; in response to the confirmation request, judging whether the current reasoning step is the final reasoning step; when the current reasoning step is an intermediate reasoning step, performing the next reasoning operation; in the case of the final reasoning step, based on the reasoning content and reasoning results corresponding to the multi-step reasoning steps, respectively, the training data corresponding to the target task is obtained. The scheme of the embodiment of the present application improves the efficiency of data annotation.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of data annotation, and in particular, to a data annotation method, a model training method, a display method, a device, a storage medium, and a program product. Background Art

[0002] With the development of artificial intelligence technology, large models have become the core of the artificial intelligence field. Large models refer to machine learning models with large-scale parameters and complex computing structures. They have been widely used in natural language processing, intelligent question answering, multimodal processing, computer vision and other scenarios. When dealing with complex problems, large models have the ability to gradually decompose the problem and solve each sub-problem one by one through multi-step reasoning operations, and finally get a complete answer, so as to better understand user intentions and needs.

[0003] At present, large models usually adopt a training mode of pre-training plus fine-tuning training, and the excellent reasoning ability of large models depends on high-quality labeled training data. However, in the existing technology, training data is manually labeled by labelers, which is costly and inefficient. Summary of the invention

[0004] The embodiments of the present application provide a data labeling method, a model training method, a display method, a device, a storage medium and a program product to solve the problem of low efficiency and high cost of data labeling caused by manual labeling in the prior art.

[0005] In a first aspect, the present application proposes a data annotation method, which is applied to an annotation system, and the method comprises:

[0006] In response to a labeling request triggered for a target task, performing a multi-step reasoning operation for the target task using the first large model;

[0007] For the current reasoning step, generate the reasoning content of the reasoning step, and generate the reasoning result according to the reasoning content;

[0008] Outputting the inference content and the inference result;

[0009] In response to the confirmation request, determining whether the current reasoning step is a final reasoning step;

[0010] When the current reasoning step is an intermediate reasoning step, the next reasoning operation is performed;

[0011] When the current reasoning step is the final reasoning step, the training data corresponding to the target task is obtained based on the reasoning contents and reasoning results corresponding to the multiple reasoning steps respectively; the training data is used to train the second largest model.

[0012] In the second aspect, the present application proposes a model training method, comprising:

[0013] Acquire training data for a target task, the training data being obtained based on the reasoning content and reasoning results respectively corresponding to the multiple reasoning steps corresponding to the target task; the reasoning content and reasoning results respectively corresponding to the multiple reasoning steps are generated and confirmed by the first large model performing a multiple reasoning operation on the sample object for the target task;

[0014] Training a second model using the training data;

[0015] The second large model is used to respond to a processing request for a target task of an object to be processed and perform a multi-step reasoning operation to obtain a processing result of the object to be processed.

[0016] In a third aspect, the present application proposes a display method, comprising:

[0017] Display the annotation prompt information of the target task in the annotation interface;

[0018] In response to a trigger operation for the annotation prompt information, a annotation request is generated; the annotation request is used to trigger the first large model to perform a multi-step reasoning operation for the target task, generate a reasoning content for the reasoning step for the current reasoning step, and generate a reasoning result according to the reasoning content;

[0019] Displaying the reasoning content and reasoning result of the current reasoning step in the annotation interface;

[0020] In response to the confirmation operation, a confirmation request is generated; the confirmation request is used to trigger the first large model to determine whether the current reasoning step is the final reasoning step, and when the current reasoning step is an intermediate reasoning step, perform the next reasoning operation, and when the current reasoning step is the final reasoning step, based on the reasoning content and reasoning results corresponding to the multiple reasoning steps, obtain the training data corresponding to the target task; the training data is used to train the second large model.

[0021] In a fourth aspect, the present application proposes a computing device, comprising a storage component and a processing component; the storage component stores one or more computer program instructions, the computer program instructions are called and executed by the processing component, and the processing component executes the one or more computer program instructions to implement the data labeling method described in the first aspect, or the model training method described in the second aspect, or the display method described in the third aspect.

[0022] In a fifth aspect, the present application proposes a computer-readable storage medium storing a computer program, wherein the computer program is executed by a computer to implement the data labeling method described in the first aspect, or the model training method described in the second aspect, or the display method described in the third aspect.

[0023] In a sixth aspect, the present application proposes a computer program product storing a computer program, which, when executed by a computer, implements the data labeling method described in the first aspect, or the model training method described in the second aspect, or the display method described in the third aspect.

[0024] The scheme of the embodiment of the present application, in response to the labeling request triggered for the target task, can use the first large model to perform multi-step reasoning operations on the target task. During the multi-step reasoning operation, for the current reasoning step, the reasoning content under the reasoning step can be generated, and the reasoning result is generated according to the reasoning content, and the reasoning content and the reasoning result are output together. In response to the confirmation request, it can be determined whether the current reasoning step is the final reasoning step. If the current reasoning step is an intermediate reasoning step, the next reasoning operation is continued. If the current reasoning step is the final reasoning step, the training data corresponding to the target task is obtained based on the reasoning content and reasoning results corresponding to the multi-step reasoning steps. By adopting a human-computer collaborative labeling method, with the help of the reasoning ability of the first large model to assist in the multi-step reasoning operation, it is only necessary to manually confirm the reasoning content and reasoning results of each reasoning step, without manually writing the multi-step reasoning steps and the reasoning content and results of each reasoning step, which greatly shortens the data labeling time, improves the data labeling efficiency, and reduces the labor cost.

[0025] These and other aspects of the present application will become more apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 A flow chart of an embodiment of a data annotation method provided by the present application is shown;

[0028] Figure 2 A flow chart of an embodiment of a model training method provided by the present application is shown;

[0029] Figure 3A flow chart of an embodiment of a data reasoning method provided by the present application is shown;

[0030] Figure 4 A flow chart of an embodiment of a display method provided by the present application is shown;

[0031] Figure 5 A schematic diagram showing annotation prompt information displayed in an annotation interface in an actual application is shown;

[0032] Figure 6 A schematic diagram showing the reasoning content and reasoning results of the current reasoning step in a annotation interface in an actual application is shown;

[0033] Figure 7 A schematic diagram showing update prompt information displayed in a annotation interface in an actual application is shown;

[0034] Figure 8 A schematic diagram showing the reasoning content and reasoning results of the final reasoning step in a annotation interface in an actual application is shown;

[0035] Fig. 9 A schematic diagram of a system architecture in a practical application is shown;

[0036] Fig.10 A schematic diagram showing the structure of an embodiment of a data labeling device provided by the present application is shown;

[0037] Fig.11 A schematic diagram showing the structure of an embodiment of a model training device provided by the present application is shown;

[0038] Fig.12 A schematic diagram showing the structure of an embodiment of a data inference device provided by the present application is shown;

[0039] Fig.13 A schematic structural diagram of an embodiment of a display device provided by the present application is shown;

[0040] Fig.14 A schematic diagram of the structure of an embodiment of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0042] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0043] The technical solution of the embodiment of the present application can be applied to data annotation scenarios, especially annotation scenarios for training data of large models. As described above, when a large model deals with complex problems, it specifically decomposes the problem step by step through multi-step reasoning operations and solves each sub-problem one by one, and finally obtains a complete answer. When annotating the training data of a large model, it is usually divided into multiple reasoning steps, and each reasoning step is annotated, and the final annotation result is obtained by the annotation results corresponding to multiple reasoning steps.

[0044] However, in the prior art, this is achieved by manual labeling by labelers, who manually divide and write multiple reasoning steps and manually write labeling results for each reasoning step. This takes a long time, has low efficiency, and has high labor costs.

[0045] In order to solve the above problems, the inventors thought that if a general large model can be used to assist the annotators in annotation, that is, the general large model can assist in performing multi-step reasoning operations, generate reasoning content and reasoning results for each step of reasoning operations, and only the generated reasoning content and reasoning results need to be manually confirmed, the annotation time can be greatly shortened, the annotation efficiency can be improved, and the cost can be reduced. Based on this, the technical solution of the present application is proposed, including: in response to a labeling request triggered for a target task, using the first large model to perform multi-step reasoning operations for the target task; for the current reasoning step, generating the reasoning content of the reasoning step, and generating the reasoning result according to the reasoning content; outputting the reasoning content and the reasoning result; in response to the confirmation request, determining whether the current reasoning step is the final reasoning step; when the current reasoning step is an intermediate reasoning step, performing the next reasoning operation; when the current reasoning step is the final reasoning step, based on the reasoning content and reasoning results corresponding to the multi-step reasoning steps, obtaining the training data corresponding to the target task; the training data is used to train the second large model.

[0046] By adopting a human-machine collaborative labeling method and relying on the reasoning ability of the first large model to assist in multi-step reasoning operations, it is only necessary to manually confirm the reasoning content and reasoning results of each reasoning step, without manually writing multiple reasoning steps and the reasoning content and results of each reasoning step. This greatly shortens the data labeling time, improves data labeling efficiency, and reduces labor costs.

[0047] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0048] The technical solution of the embodiment of the present application can be applied to the architecture of the annotation system including the client and the server. The client and the server can establish a connection through a network. The network can include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0049] Optionally, the client may include a marking client and a processing client. The marking client may refer to a client used by the marking personnel, which may interact with the server through the network to send tasks or receive prompt information, etc. The processing client may refer to a client corresponding to the object to be processed. For example, when the object to be processed is a merchant store, the processing client may be a client used by relevant personnel of the merchant store.

[0050] Among them, the client is provided with a user interface, and prompt information can be displayed in the user interface and user operations can be sensed. The user end can be a browser, an APP (Application), or a web application such as H5 (Hyper Text Markup Language5, Hypertext Markup Language Version 5) application, or a light application (also known as a mini-program, a lightweight application) or a cloud application, etc. The user end can be deployed in an electronic device and needs to rely on the device to run or some APPs in the device to run. For example, the electronic device can have a display screen and support information browsing, such as a personal mobile terminal such as a mobile phone, a tablet computer, a personal computer, etc. Various other types of applications can also be configured in the electronic device, such as human-computer dialogue applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0051] The service end may include servers that provide various services, such as a server that processes tasks sent by a client to obtain training data, a server that uses training data to train a large model, etc.

[0052] It should be noted that the server can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. The server can also be a server of a distributed system, or a server combined with blockchain. The server can also be a cloud server, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0053] It should be noted that the data annotation method, model training method and data reasoning method provided in the embodiments of the present application are generally performed by the server, and the display method is generally performed by the client. However, in other embodiments of the present application, the client may also have similar functions to the server, thereby executing the data annotation method, model training method and data reasoning method provided in the embodiments of the present application. In other embodiments, the data annotation method, model training method, or data reasoning method provided in the embodiments of the present application can also be performed jointly by the client and the server.

[0054] It should be noted that the embodiments of the present application may involve the use of user data. In actual applications, user-specific personal data can be used in the scheme described herein within the scope permitted by applicable laws and regulations, subject to the requirements of applicable laws and regulations of the country where the user is located (for example, with the user's explicit consent, effective notification to the user, etc.).

[0055] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0056] It should be noted that the technical solution of the embodiment of the present application is applicable to a network virtual environment. The users described generally refer to "virtual users". Real users can register a user account in the server through registration to obtain a user identity in the network environment.

[0057] like Figure 1 FIG. 1 is a flowchart of an embodiment of a data annotation method provided by the present application. The method is applied to an annotation system and can be specifically executed by a server in the annotation system. The method can include the following steps:

[0058] 101: In response to a labeling request triggered for a target task, use a first large model to perform a multi-step reasoning operation for the target task.

[0059] Among them, the target task can refer to the task to be processed, such as the merchant store data analysis task, the product sales data analysis task, the product purchase recommendation task, etc. in the e-commerce scenario.

[0060] Optionally, before this, the server can provide multiple tasks, respond to the labeling request for any task, determine the target task, and use the first large model to perform multi-step reasoning operations on the target task. Among them, multiple tasks can be provided to the labeler through the labeling client, and the labeler selects the target task and triggers the generation of the labeling request.

[0061] The large model (also called the foundation model, i.e., FoundationModel) involved in the embodiments of the present application, including the first large model and the second large model involved below, refers to a machine learning model with a large number of parameters and a complex structure, which can process massive data and complete various complex tasks, such as natural language processing, computer vision, speech recognition, etc. It is an AI (Artificial Intelligence) model. The large model can be implemented using LLM (Large Language Model) or MLM (Multimodal Large Model), for example, a generative pre-training model can be used, and this application does not limit this.

[0062] Optionally, in response to a labeling request triggered for a target task, a multi-step reasoning operation may be performed on the target task using the first large model for the sample object.

[0063] Among them, the sample object can refer to the processing object corresponding to the target task. For example, when the target task is a merchant store data analysis task, the sample object can refer to the merchant store. For another example, when the target task is a commodity sales data analysis task, the sample object can refer to the commodity.

[0064] 102: Generate reasoning content of the current reasoning step, and generate reasoning results according to the reasoning content.

[0065] 103: Output the inference content and inference result.

[0066] Specifically, during the multi-step reasoning operation, for the current reasoning step, the first large model can be used to generate the reasoning content under the reasoning step, and the reasoning result can be generated according to the reasoning content. The server can send the reasoning content and reasoning result corresponding to the current reasoning step to the annotation client for display, so that the annotation personnel can view it.

[0067] Among them, in each reasoning step, the first large model can perform reasoning operations according to the reasoning paradigm. The reasoning paradigm is used to instruct the large model to combine reasoning and behavior, so that the large model can complete complex tasks through multi-step reasoning operations, and each reasoning step can be implemented with the help of tools. Among them, the reasoning paradigm can be further divided into four main parts: thought information (Thought), action (Action), action parameters (Action Input) and reasoning results (Observation). Each reasoning step of the first large model can be divided into the above four parts. Thought information indicates the thinking process of the current reasoning step; action indicates the operation to be performed in the current reasoning step, such as querying information, calling tools, generating intermediate results, etc.; action parameters indicate the input information used for the operation to be performed in the current reasoning step, such as query keywords, tool names, etc.; reasoning results indicate the results or feedback obtained after executing the operation.

[0068] The reasoning content generated by reasoning using the first large model may include, for example, thinking information and action information, wherein the action information may include action content and action parameters, etc. The first large model can generate reasoning results according to the thinking information and action information. The specific implementation will be described in subsequent embodiments.

[0069] 104: In response to the confirmation request, determine whether the current reasoning step is the final reasoning step. If the determination result is no, return to step 102 to continue executing to perform the next reasoning operation; if the determination result is yes, execute step 105.

[0070] The confirmation request may be generated by the annotation client based on the confirmation operation of the annotator.

[0071] 105: Based on the reasoning contents and reasoning results corresponding to the multiple reasoning steps, obtain training data corresponding to the target task.

[0072] That is to say, when the current reasoning step is an intermediate reasoning step, the next reasoning operation can be performed, that is, the reasoning content and reasoning result in the next reasoning step are generated and output for confirmation. Among them, the intermediate reasoning step includes the first reasoning step.

[0073] When the current reasoning step is the final reasoning step, the training data corresponding to the target task can be obtained based on the reasoning content and reasoning results corresponding to the multiple reasoning steps. Among them, the training data may include the reasoning content and reasoning results corresponding to each reasoning step. The training data can be used to train the second largest model, and the trained second largest model can be used for data reasoning. The second largest model and the first largest model can be the same large model, or of course, different large models, and this application does not limit this.

[0074] In this embodiment, in response to the labeling request triggered for the target task, the first large model can be used to perform a multi-step reasoning operation on the target task. During the multi-step reasoning operation, for the current reasoning step, the reasoning content under the reasoning step can be generated, and the reasoning result is generated according to the reasoning content, and the reasoning content and the reasoning result are output together. In response to the confirmation request, it can be determined whether the current reasoning step is the final reasoning step. If the current reasoning step is an intermediate reasoning step, the next reasoning operation is continued. If the current reasoning step is the final reasoning step, the training data corresponding to the target task is obtained based on the reasoning content and reasoning results corresponding to the multi-step reasoning steps. By adopting a human-computer collaborative labeling method, the multi-step reasoning operation is assisted by the reasoning ability of the first large model. It is only necessary to manually confirm the reasoning content and reasoning results of each reasoning step, without manually writing the multi-step reasoning steps and the reasoning content and results of each reasoning step, which greatly shortens the data labeling time, improves the data labeling efficiency, and reduces the labor cost.

[0075] In practical applications, in order to improve the accuracy of reasoning, in some embodiments, in response to a labeling request triggered for a target task, performing a multi-step reasoning operation on the target task using the first large model may include:

[0076] In response to a labeling request triggered for a target task, determining a processing example corresponding to the target task;

[0077] The first large model is used in combination with processing samples to perform multi-step reasoning operations on the target task.

[0078] The processing sample can be pre-annotated by the annotator for the target task, and the processing sample can include multiple reasoning steps and the reasoning content and reasoning results corresponding to each reasoning step. It can be understood that the reasoning steps divided in the processing sample and the reasoning content and reasoning results annotated for each reasoning step are reference samples for the first large model to perform multi-step reasoning operations to improve the reasoning efficiency and accuracy of the first large model, but are not the final training data.

[0079] In order to further improve the accuracy of reasoning, in some embodiments, for the current reasoning step, generating the reasoning content of the reasoning step may include:

[0080] For the current reasoning step, the reasoning content and reasoning results of the historical reasoning steps are combined to generate the reasoning content of the reasoning step.

[0081] Among them, when the current reasoning step is the first reasoning step, the first large model can be used in combination with the processing sample to generate the reasoning content and reasoning result under the reasoning step. When the current reasoning step is not the first reasoning step, the first large model can be used in combination with the reasoning content and reasoning results corresponding to all historical reasoning steps to generate the reasoning content and reasoning result under the current reasoning step. Optionally, the first large model can be used in combination with the processing sample, and the reasoning content and reasoning results corresponding to all historical reasoning steps to generate the reasoning content and reasoning result under the current reasoning step.

[0082] By combining the reasoning content and reasoning results corresponding to all historical reasoning steps, the reasoning operation of the current reasoning step is performed, which enhances the logical closeness between multiple reasoning steps and improves the accuracy of reasoning.

[0083] In actual applications, in order to improve the accuracy of training data, data is usually acquired through tool calls for reasoning, where the tools may include API (Application Programming Interface), search engines, etc. That is, in the reasoning process, the generated reasoning content may include thinking information and action information, and executing the corresponding action according to the instructions of the action information may refer to calling the indicated tool to obtain the corresponding data. In the traditional solution, the annotator manually searches and determines the tool to be called, which places high demands on the annotator and has low accuracy, thus affecting the accuracy of the reasoning results and training data.

[0084] Therefore, in order to solve this technical problem, in some embodiments, the annotation system may integrate multiple candidate tools corresponding to the target task. The candidate tools may include APIs, search engines, and the like.

[0085] On this basis, for the current reasoning step, generating the reasoning content of the reasoning step, and generating the reasoning result according to the reasoning content may include:

[0086] For the current reasoning step, the thinking information and action information of the reasoning step are generated; the action information may include a tool identification of the target tool; and according to the target tool and action parameters indicated by the action information, the target tool is called to obtain target data that meets the action parameters.

[0087] Among them, the action parameters may refer to the data conditions for calling the tool to obtain data, which may include data source, data time, data format, etc. The data source may, for example, include the identification information of the data provider, such as the merchant store ID, etc. The data time may, for example, include the time range for obtaining data, such as obtaining data within the current week, etc.

[0088] At this time, the target data obtained is the reasoning result of the current reasoning step.

[0089] Furthermore, the output of the inference content and the inference result may include:

[0090] Output thinking information, action information and target data.

[0091] In this embodiment, by integrating multiple candidate tools corresponding to the target task in the annotation system, it is possible to call the target tool and action parameters indicated by the action information in the reasoning content during the reasoning operation to obtain the target data that meets the action parameters, so as to obtain the reasoning result under the current reasoning step. There is no need to manually search for tools, which reduces the requirements for annotation personnel and improves the accuracy of data acquisition, thereby improving the accuracy of training data. When the second largest model is trained with high-accuracy training data, the reasoning accuracy of the second largest model can also be improved.

[0092] In some embodiments, the plurality of candidate tools may be data interfaces. In response to a labeling request triggered for a target task, performing a multi-step reasoning operation for the target task using the first large model may include:

[0093] In response to a labeling request triggered for a target task, a multi-step reasoning operation is performed on the sample object using the first large model for the target task.

[0094] At this time, according to the target tool and the action parameters indicated by the action information, calling the target tool to obtain the target data that meets the action parameters may include:

[0095] According to the target interface and action parameters indicated by the action information, the target interface is called to obtain target data that meets the action parameters from the data set of the sample object.

[0096] Taking the sample object as a merchant store as an example, the data interface can be a data acquisition interface provided by the merchant store, such as a visitor data acquisition interface, a commodity sales data acquisition interface, etc. For example, the target interface indicated by the action information is a visitor data acquisition interface, and the action parameters include the merchant store ID and the time range for obtaining visitor data, such as within three days from the current time, then the visitor data interface can be called to obtain visitor data within three days from the current time from the visitor data set of the merchant store based on the merchant store ID.

[0097] In this embodiment, by integrating multiple data interfaces corresponding to the target task in the annotation system, it is possible to achieve, during the reasoning operation, with the help of the target interface and action parameters indicated by the action information in the reasoning content, call the target interface to obtain the target data that meets the action parameters from the data set of the sample object, so as to obtain the reasoning result under the current reasoning step, thereby improving the accuracy of data acquisition, and then improving the accuracy of the training data and the reasoning accuracy of the second largest model.

[0098] In order to further improve the accuracy of the training data, the reasoning content and reasoning results corresponding to the current reasoning step generated by the first large model may also be updated. Therefore, in some embodiments, after outputting the reasoning content and reasoning results, the method further includes:

[0099] In response to the update request, the inference content and the inference result are updated.

[0100] The update request may be generated by the annotation client based on the update operation of the annotator, and may include the updated reasoning content and / or reasoning result.

[0101] In actual applications, after the server updates the reasoning content and reasoning results of the current reasoning step, the updated reasoning content and reasoning results can be sent to the annotation client again. Optionally, the reasoning content and reasoning results in a reasoning step can be updated multiple times until confirmed, and the server can respond to the confirmation request and perform subsequent operations.

[0102] In this embodiment, by adopting a human-computer collaborative labeling method, the reasoning ability of the first large model is used to assist in multi-step reasoning operations, and the reasoning content and reasoning results of each reasoning step are manually updated and confirmed. On the basis of improving data labeling efficiency, the accuracy of training data is improved.

[0103] The update implementation is described below.

[0104] In some embodiments, when the current reasoning step is an intermediate reasoning step, in response to an update request, updating the reasoning content and the reasoning result may include:

[0105] In response to the update request, updating the inference content;

[0106] Based on the updated reasoning content, the first model is used to update the reasoning result.

[0107] In this embodiment, the update request may be generated for an update operation triggered by the inference content, and the update request may include the updated inference content. At this time, the server may update the inference result based on the updated inference content using the first large model, that is, re-infer using the first large model to obtain the inference result.

[0108] In this embodiment, the reasoning content of the current reasoning step is updated in response to an update request, and the reasoning result is updated using the first large model based on the updated reasoning content, thereby improving the rationality and accuracy of the reasoning content and the reasoning result.

[0109] Optionally, based on the updated reasoning content, the reasoning content and reasoning results corresponding to all historical reasoning steps can be combined, and the first large model can be used to update the reasoning results under the current reasoning step to further enhance the reasoning accuracy.

[0110] In some embodiments, the reasoning content may include thinking information, action content, and action parameters. At this time, in response to the update request, updating the reasoning content may include:

[0111] In response to an update request triggered by thinking information, providing selection prompt information of a plurality of candidate thinking information, updating the thinking information based on the selected target candidate thinking information, and updating the action content and / or action parameters based on the updated thinking information;

[0112] In response to an update request triggered for an action content, providing selection prompt information of a plurality of candidate action contents, updating the action content based on the selected target candidate action content, and updating the action parameters based on the updated action content;

[0113] In response to an update request triggered for an action parameter, selection prompt information for a plurality of candidate parameters is provided, and the action parameter is updated based on the selected target candidate parameter.

[0114] In this embodiment, the update request may be specifically generated for an update operation on one or more of thinking information, action content, and action parameters.

[0115] Specifically, when updating the thinking information, selection prompt information including multiple candidate thinking information can be provided to the annotator for selection. The multiple candidate thinking information is also generated by reasoning under the current reasoning step using the first large model. Thus, in response to the annotator's selection request for any candidate thinking information, the target candidate thinking information can be determined, and the thinking information under the current reasoning step can be updated based on the target candidate thinking information. For example, the target candidate thinking information can be directly used as the updated thinking information. Based on the updated thinking information, the action content and / or action parameters can be updated using the first large model.

[0116] On this basis, the first large model can be used to update the reasoning results based on the updated thinking information, action content and action parameters, and the updated reasoning results can be sent together with the thinking information, action content and action parameters to the labeling personnel for updating or confirmation.

[0117] Optionally, based on the updated thinking information, action content and action parameters, the reasoning content and reasoning results corresponding to all historical reasoning steps can be combined to update the reasoning results using the first large model.

[0118] When updating the action content, selection prompt information including multiple candidate action contents can be provided to the annotator. The multiple candidate action contents are also inferred and generated using the first large model under the current reasoning step. In response to the annotator's selection request for any candidate action content, the target candidate action content can be determined, and the action content under the current reasoning step can be updated based on the target candidate action content. For example, the target candidate action content can be directly used as the updated action content. Based on the updated action content, the action parameters can be updated using the first large model. At this time, the thinking information is no longer updated.

[0119] On this basis, the first large model can be used to update the reasoning results based on the thinking information, updated action content and action parameters, and the thinking information, updated reasoning results, action content and action parameters can be sent to the labeling personnel for updating or confirmation.

[0120] Optionally, based on the thinking information, updated action content and action parameters, the reasoning content and reasoning results corresponding to all historical reasoning steps can also be combined to update the reasoning results using the first large model.

[0121] When updating the action parameters, selection prompt information including multiple candidate parameters can be provided to the annotator. The multiple candidate parameters are also inferred and generated using the first large model under the current reasoning step. In response to the annotator's selection request for one or more candidate parameters, the target candidate parameters can be determined, and the action parameters under the current reasoning step can be updated based on the target candidate parameters. For example, the target candidate parameters can be directly used as the updated action parameters. At this time, the thinking information and action content are no longer updated.

[0122] On this basis, the first large model can be used to update the reasoning results based on the thinking information, action content, and updated action parameters, and the thinking information, action content, updated reasoning results and action parameters can be sent to the labeling personnel for updating or confirmation.

[0123] Optionally, based on the thinking information, action content, and updated action parameters, the reasoning content and reasoning results corresponding to all historical reasoning steps can be combined to update the reasoning results using the first large model.

[0124] In this embodiment, by responding to the update request triggered by the annotator for the reasoning content, the annotator is provided with selection prompt information including multiple candidate thinking information, candidate action content, and / or candidate action parameters for selection by the annotator, and the annotator does not need to manually write updates, thereby improving the update efficiency and further improving the data annotation efficiency. In addition, by following the chain update relationship of updating the action content and / or action parameters based on the thinking information and updating the action parameters based on the action content, the update efficiency is further improved.

[0125] In actual applications, when the annotation system integrates multiple candidate tools, the action content is to call the tool to obtain data. In order to facilitate the annotation personnel to better select the target candidate tool, the annotation personnel can also be provided with description information corresponding to the multiple candidate tools. The description information can describe the functions and usage of the candidate tools, etc., wherein the description information can be generated when the candidate tools are integrated. For example, when the candidate tool is a data interface, the corresponding description information can be the function of the data interface, such as the interface is used to obtain visitor data, the interface is used to obtain product sales data, and the interface is used to sort data.

[0126] By providing the annotation personnel with the descriptive information corresponding to the candidate tools, so that the annotation personnel can clearly understand the calling functions and calling methods of the candidate tools, it will help to better select and update the candidate tools and improve the updating efficiency and accuracy.

[0127] Of course, in actual applications, it is also possible to support manual update writing by annotators. Optionally, the update request triggered by the thinking information can be generated based on the thinking information input by the annotator. At this time, the thinking information under the current reasoning step can be updated according to the thinking information input by the annotator, for example, the input thinking information is directly used as the updated thinking information, and the action content and / or action parameters are updated using the first large model, as well as the reasoning result.

[0128] In some embodiments, the above method may further include:

[0129] Providing first saving prompt information and second saving prompt information;

[0130] In response to the first save request triggered by the first save prompt information, the updated inference content is saved.

[0131] The first saving prompt information may be used to prompt the annotator to only save the updated reasoning content instead of performing the update operation.

[0132] Based on the updated reasoning content, updating the reasoning result using the first model may include:

[0133] In response to the second save request triggered by the second save prompt information, based on the updated reasoning content, the reasoning result is updated using the first large model, and the updated reasoning content and reasoning result are saved.

[0134] The second saved information can be used to prompt the annotator to save the updated reasoning content. At the same time, based on the updated reasoning content, the first large model is used to update the reasoning results, and the updated reasoning content and reasoning results are saved.

[0135] By providing two different update saving methods, different update requirements can be met. For example, for update content such as text modification and phrase replacement, without affecting the reasoning results, you can only perform the save operation without updating the reasoning results, thereby improving the update and reasoning efficiency. For update content such as tool switching and action parameter changes, you can use the save and then re-reasoning operation to ensure the accuracy of the reasoning results.

[0136] One or more of the above embodiments record the implementation method for updating the reasoning content and reasoning results in the intermediate reasoning step. When the current reasoning step is the final reasoning step, in some embodiments, the reasoning content may include thinking information. At this time, in response to the update request, updating the reasoning content and reasoning results may include:

[0137] In response to an update request for the thinking information, the thinking information is updated, and based on the updated thinking information, the inference result is updated using the first large model;

[0138] In response to an update request triggered for the inference result, the inference result is updated.

[0139] Among them, during the reasoning operation of the final reasoning step, the generated reasoning content can include thinking information, and the tool can no longer be called to execute the action content. Based on the thinking information, the first large model can be used to combine the reasoning content and reasoning results corresponding to all historical reasoning steps to generate the reasoning result corresponding to the final reasoning step. The server can send the thinking information and reasoning results to the annotation personnel for update or confirmation.

[0140] Specifically, in response to the labeler's request to update the thinking information of the final reasoning step, the thinking information can be updated, and based on the updated thinking information, the reasoning result can be updated using the first large model. Optionally, based on the updated thinking information, the reasoning content and reasoning results corresponding to all historical reasoning steps can be combined to update the reasoning result using the first large model.

[0141] In response to an update request from the annotator for the reasoning result of the final reasoning step, the reasoning result may be updated.

[0142] Optionally, if in response to the labeler's request to update the thinking information under the final reasoning step, it is confirmed to increase the execution action content, for example, the labeler updates the thinking information to add a data acquisition operation, and analyzes the acquired data, etc., the reasoning content can be updated using the first large model. The updated reasoning content can include thinking information, action content and action parameters, and the reasoning results are generated according to the thinking information, action content and action parameters, and sent to the labeler for updating or confirmation, so as to continue the next reasoning operation after confirmation. In other words, the current reasoning has not ended, and the next reasoning step still needs to be executed.

[0143] After the reasoning is completed, the thinking information and reasoning results of the final reasoning step, and the reasoning content and reasoning results corresponding to all historical reasoning steps can be used as training data for the target task.

[0144] In this embodiment, when the current reasoning step is the final reasoning step, the thinking information is updated in response to an update request for the thinking information under the final reasoning step, and the reasoning result is updated using the first large model based on the updated thinking information, and the reasoning result is updated in response to an update request for the reasoning result under the final reasoning step, thereby improving the rationality and accuracy of the reasoning result under the final reasoning step.

[0145] After obtaining the training data, in order to improve the accuracy of the training data, the generated training data may also be evaluated. Therefore, in some embodiments, the above method may further include:

[0146] In response to an evaluation request for the training data, obtaining an evaluation result;

[0147] If the evaluation is passed, the training data and the target tasks are stored in the database accordingly; the tasks and corresponding training data stored in the database are used to train the second largest model.

[0148] Among them, manual sampling or using a large model to evaluate the training data can be used to obtain the evaluation results. The evaluation method can be set according to actual needs, and this application does not limit this.

[0149] In this embodiment, the training data is evaluated, and if the evaluation is passed, the training data is stored corresponding to the target task to train the second largest model, thereby improving the accuracy of the training data and further improving the reasoning accuracy of the second largest model.

[0150] like Figure 2As shown, it is a flowchart of an embodiment of a model training method provided by the present application. The method is applied to a labeling system and can be specifically executed by a server in the labeling system. The method may include the following steps:

[0151] 201: Obtain training data for the target task.

[0152] The training data can be obtained based on the reasoning content and reasoning results corresponding to the multi-step reasoning steps corresponding to the target task; the reasoning content and reasoning results corresponding to the multi-step reasoning steps are generated and confirmed by the first large model for the sample object and the target task. Figure 1 The corresponding description has been given in the illustrated embodiment and will not be repeated here.

[0153] 202: Train the second largest model using the training data.

[0154] Among them, the second largest model can be used to respond to a processing request for a target task of the object to be processed, and perform a multi-step reasoning operation to obtain a processing result of the object to be processed.

[0155] In this embodiment, training data is obtained by adopting a human-computer collaborative labeling method, and multi-step reasoning operations are assisted by the reasoning ability of the first large model. It is only necessary to manually confirm the reasoning content and reasoning results of each reasoning step, without manually writing multiple reasoning steps and the reasoning content and results of each reasoning step. This greatly shortens the data labeling time, improves the efficiency of generating training data, and reduces labor costs, thereby improving training efficiency and reducing costs.

[0156] Optionally, a tool set for the target task may be integrated into the annotation system, and the tool set includes multiple candidate tools corresponding to the target task. Specifically, the second largest model may be trained using the training data and the tool set. During the multi-step reasoning operation, the second largest model may request to call the tool indicated in the generated reasoning content to obtain the corresponding data to obtain the output reasoning result, compare the output result with the training data, and adjust the model parameters of the second largest model to train the second largest model. The specific training process will not be repeated here.

[0157] like Figure 3 FIG. 1 is a flowchart of an embodiment of a data reasoning method provided by the present application. The method is applied to a tagging system and can be specifically executed by a server in the tagging system. The method can include the following steps:

[0158] 301: Get a processing request for a target task of an object to be processed.

[0159] The server may obtain a processing request for a target task of an object to be processed sent by the processing client.

[0160] 302: Use the second largest model to perform a multi-step reasoning operation based on the target task to obtain a processing result of the object to be processed.

[0161] The second largest model is obtained by training the training data. The training data can be obtained based on the reasoning content and reasoning results corresponding to the multi-step reasoning steps corresponding to the target task; the reasoning content and reasoning results corresponding to the multi-step reasoning steps are generated and confirmed by the first largest model for the sample object and the target task through multi-step reasoning operations. Figure 1 The corresponding description has been given in the illustrated embodiment and will not be repeated here.

[0162] 303: Output processing results.

[0163] The server can send the processing results to the processing client to be displayed in the user interface of the processing client.

[0164] In this embodiment, the second largest model is obtained based on training data. Since the training data is obtained by human-computer collaborative labeling, the reasoning ability of the first largest model is used to assist in multi-step reasoning operations. It is only necessary to manually confirm the reasoning content and reasoning results of each reasoning step, without manually writing multiple reasoning steps and the reasoning content and results of each reasoning step. This greatly shortens the data labeling time, improves the efficiency of generating training data, and reduces labor costs, thereby improving processing efficiency and reducing costs.

[0165] like Figure 4 FIG. 1 is a flowchart of an embodiment of a display method provided by the present application. The method is applied to a marking system and can be specifically executed by a marking client in the marking system. The method can include the following steps:

[0166] 401: Display the annotation prompt information of the target task in the annotation interface.

[0167] Optionally, multiple tasks can also be displayed in the annotation interface. In response to the selection operation of the annotator on any task, the target task is determined, and the annotation prompt information of the target task is displayed in the annotation interface. The annotation prompt information may include annotation controls, text prompt information, etc.

[0168] Optionally, at least one training data corresponding to the target task may also be displayed in the annotation interface, where the at least one training data is obtained by processing different sample objects.

[0169] 402: In response to a triggering operation for marking prompt information, a marking request is generated.

[0170] The annotation request can be used to trigger the first large model to perform multi-step reasoning operations on the target task, generate reasoning content for the current reasoning step, and generate reasoning results according to the reasoning content. Figure 1 The corresponding description has been given in the illustrated embodiment and will not be repeated here.

[0171] 403: Display the reasoning content and reasoning result of the current reasoning step in the annotation interface.

[0172] 404: In response to the confirmation operation, a confirmation request is generated.

[0173] The confirmation request can be used to trigger the first large model to determine whether the current reasoning step is the final reasoning step, and to perform the next reasoning operation if the current reasoning step is an intermediate reasoning step, and to obtain the training data corresponding to the target task based on the reasoning content and reasoning results corresponding to the multiple reasoning steps if the current reasoning step is the final reasoning step. The training data can be used to train the second large model. The specific update process and training data generation are described in Figure 1 There are corresponding descriptions in the illustrated embodiment, which will not be repeated here.

[0174] In this embodiment, by adopting a human-computer collaborative labeling method, the reasoning ability of the first large model is used to assist in multi-step reasoning operations. It is only necessary to manually confirm the reasoning content and reasoning results of each reasoning step, without manually writing multiple reasoning steps and the reasoning content and results of each reasoning step. This greatly shortens the data labeling time, improves data labeling efficiency, and reduces labor costs.

[0175] For ease of understanding, Figure 5 FIG. 1 shows a schematic diagram of displaying annotation prompt information in an annotation interface in an actual application. The annotation prompt information may include text prompt information and annotation controls. Figure 5 As shown, multiple tasks are displayed in the annotation interface 50, namely, Task 1, Task 2, Task 3, etc. Taking Task 1 as an example, the annotation interface 50 displays an annotation control 501, and the annotator can click the annotation control 501. Multiple annotation data corresponding to Task 1 and a text prompt message 502 of "unannotated" are displayed in the annotation interface 50. The annotator can click the text box corresponding to the text prompt message 502 to trigger an annotation request for Task 1.

[0176] Figure 6 FIG. 1 shows a schematic diagram of displaying the reasoning content and reasoning results of the current reasoning step in a annotation interface in an actual application. The reasoning content includes thinking information, action content and action parameters. Figure 6As shown, the annotation interface 50 displays the target task, namely, task 1, as well as thinking information 601, action content 602, action parameters 603 and reasoning results 604.

[0177] Alternatively, if Figure 6 As shown, the annotation interface 50 also displays the current reasoning step 605, such as Figure 6 As shown, the current reasoning step is the third step, so that the annotator can clearly understand the multi-step reasoning progress.

[0178] In some embodiments, after displaying the reasoning content and reasoning result of the current reasoning step in the annotation interface, the method may further include:

[0179] In response to the update operation, an update request is generated; the update request can be used to trigger the first large model to update the inference content and inference results.

[0180] In some embodiments, the display method may further include:

[0181] Display confirmation prompt information in the annotation interface.

[0182] At this time, in response to the confirmation operation, generating a confirmation request may include:

[0183] In response to a confirmation operation triggered by the confirmation prompt information, a confirmation request is generated.

[0184] In some embodiments, the display method may further include:

[0185] Display update prompt information in the annotation interface.

[0186] At this time, in response to the update operation, generating an update request may include:

[0187] In response to an update operation triggered by the update prompt information, update content is determined, and an update request is generated based on the update content.

[0188] Figure 7 FIG. 1 shows a schematic diagram of displaying update prompt information in a annotation interface in an actual application. The update prompt information may include an edit control for thought information, an edit control for action content, and an edit control for action parameters. Figure 7 As shown, the annotation interface 50 displays an editing control 701 for thinking information, an editing control 702 for action content, and an editing control 703 for action parameters.

[0189] Optionally, confirmation prompt information may also be displayed in the annotation interface. The confirmation prompt information may include a confirmation control and text prompt information. Figure 7As shown, a confirmation control 704 and a text prompt message 705 of “Confirmed, generate next step” are displayed in the annotation interface 50 .

[0190] Optionally, a “previous step” return control 706 may also be displayed in the annotation interface 50 , for returning to the previous step of reasoning to display the reasoning content and reasoning results.

[0191] In some embodiments, the reasoning content may include thought information, action content, and action parameters.

[0192] In response to the update operation triggered by the update prompt information, determining the update content, and generating the update request based on the update content may include:

[0193] In the case where the update prompt information includes selection prompt information of multiple candidate thinking information, in response to the selection operation for any candidate thinking information, the target candidate thinking information is determined, and based on the target candidate thinking information, an update request is generated, and the update request is used to trigger the first large model to update the thinking information, and update the action content and / or action parameters based on the updated thinking information.

[0194] In the case where the update prompt information includes selection prompt information for multiple candidate action contents, in response to a selection operation for any candidate action content, a target candidate action content is determined, and based on the target candidate action content, an update request is generated, the update request is used to trigger the first large model to update the action content, and update the action parameters based on the updated action content.

[0195] In the case where the update prompt information includes selection prompt information for multiple candidate parameters, in response to a selection operation on at least one candidate parameter, a target candidate parameter is determined, and based on the target candidate parameter, an update request is generated, and the update request is used to trigger the first large model to update the action parameters.

[0196] In some embodiments, the display method may further include:

[0197] Displaying the first saving prompt information and the second saving prompt information in the annotation interface;

[0198] In response to a first save operation triggered by the first save prompt information, a first save request is generated, and the first save request is used to trigger the first large model to save the updated inference content.

[0199] In response to the update operation, generating an update request may include:

[0200] In response to the second save operation triggered by the second save prompt information, a second save request is generated, and the second save request is used to trigger the first large model to update the reasoning result based on the updated reasoning content, and save the updated reasoning content and reasoning result.

[0201] Figure 8 A schematic diagram of displaying the reasoning content and reasoning results of the final reasoning step in a annotation interface in an actual application is shown. Among them, the reasoning content under the final reasoning step includes thinking information. Figure 8 As shown, the annotation interface 50 displays the thinking information 801 and the reasoning result 802 under the final reasoning step.

[0202] For ease of understanding, the following Fig. 9 The architecture diagram of the annotation system shown in the figure illustrates the technical solution of this application. Fig. 9 As shown, the annotation system may include a server 901, an annotation client 902, and a processing client 902. The number and form of the annotation client 901 and the processing client 902 may be set according to actual needs, and this application does not limit this.

[0203] The server 901 can provide multiple tasks to the annotation client 902, and determine the target task, such as the merchant store data analysis task, in response to the annotation personnel's selection request for any task. For the sample object, such as the merchant store, the first large model is used to perform a multi-step reasoning operation on the target task. During the multi-step reasoning operation, for the current reasoning step, the reasoning content under the reasoning step is generated, and the reasoning result is generated according to the reasoning content, and the reasoning content and the reasoning result are sent to the annotation client 902. In response to the update operation triggered by the annotation personnel, the annotation client 902 generates an update request and sends it to the server 901, and the server 901 updates the reasoning content and the reasoning result. In response to the confirmation operation triggered by the annotation personnel, the annotation client 902 generates a confirmation request and sends it to the server 901. The server 901 can determine whether the current reasoning step is the final reasoning step. If the current reasoning step is an intermediate reasoning step, the next reasoning operation is continued. If the current reasoning step is the final reasoning step, the training data corresponding to the target task is obtained based on the reasoning content and reasoning results corresponding to the multiple reasoning steps. By adopting a human-machine collaborative labeling method and using the reasoning ability of the first large model to assist in multi-step reasoning operations, labelers only need to update and confirm the reasoning content and reasoning results of each reasoning step, without manually writing multiple reasoning steps and the reasoning content and results of each reasoning step. This greatly shortens the data labeling time, improves data labeling efficiency, and reduces labor costs.

[0204] Optionally, the server 901 can respond to the labeling request triggered by the labeling personnel for the target task, determine the processing sample corresponding to the target task, and use the first large model in combination with the processing sample to perform multi-step reasoning operations on the target task to improve the reasoning efficiency and accuracy of the first large model.

[0205] Optionally, the server 901 may combine the reasoning contents and reasoning results corresponding to all historical reasoning steps to perform the reasoning operation of the current reasoning step, thereby enhancing the logical tightness between multiple reasoning steps and improving the reasoning accuracy.

[0206] Optionally, multiple candidate tools corresponding to the target task are integrated in the annotation system, such as multiple data interfaces. The server 901 can generate thinking information and action information of the reasoning step for the current reasoning step; the action information can include the tool identification of the target tool; and according to the target tool and action parameters indicated by the action information, the target tool is called to obtain target data that meets the action parameters. By integrating multiple candidate tools corresponding to the target task in the annotation system, the target tool is called to obtain target data that meets the action parameters with the help of the target tool and action parameters indicated by the action information in the reasoning content, and there is no need for the annotator to find the tool by himself, which reduces the requirements for the annotator and improves the accuracy of data acquisition, thereby improving the accuracy of the training data. When the second largest model is trained with high-accuracy training data, the reasoning accuracy of the second largest model can also be improved.

[0207] Optionally, after the server 901 sends the reasoning content and reasoning results generated by the current reasoning step to the labeling personnel, it can respond to the update request triggered by the labeling personnel to update the reasoning content of the current reasoning step, and based on the updated reasoning content, use the first large model to update the reasoning results, thereby improving the rationality and accuracy of the reasoning content and reasoning results.

[0208] Optionally, the reasoning content generated by the server 901 using the first large model may include thinking information, action content, and action parameters. In response to the update request triggered by the annotation personnel for the reasoning content, the server 901 provides the annotation personnel with selection prompt information including multiple candidate thinking information, candidate action content, and / or candidate action parameters for the annotation personnel to choose from, without the need for the annotation personnel to manually write updates, thereby improving the update efficiency and further improving the data annotation efficiency. In addition, by updating the action content and / or action parameters based on the thinking information and updating the action parameters based on the action content in a chained update relationship, the update efficiency is further improved.

[0209] Optionally, the server 901 may also provide the annotation personnel with descriptive information corresponding to the candidate tools so that the annotation personnel can clearly understand the calling functions and calling methods of the candidate tools, which helps to better select and update the candidate tools and improve the updating efficiency and accuracy.

[0210] Optionally, the server 901 can provide the annotator with a first save prompt and a second save prompt; in response to a first save request triggered by the first save prompt, save the updated reasoning content; in response to a second save request triggered by the second save prompt, based on the updated reasoning content, use the first large model to update the reasoning result, and save the updated reasoning content and reasoning result. By providing two different update saving methods, different update requirements can be met. For example, for update content such as text modification and phrase replacement, only the save operation can be performed without updating the reasoning result without affecting the reasoning result, thereby improving the update and reasoning efficiency. For update content such as tool switching and action parameter changes, the re-reasoning operation after saving can be adopted to ensure the accuracy of the reasoning result.

[0211] Optionally, after the server 901 obtains the training data, it can use the training data to train the second largest model. After the second largest model is trained, the server 901 can respond to the processing request for the target task of the object to be processed sent by the processing client 903, and use the second largest model to perform multi-step reasoning operations based on the target task to obtain the processing result of the object to be processed, and feed the processing result back to the processing client 903.

[0212] like Fig.10 As shown, it is a structural schematic diagram of an embodiment of a data annotation device provided by the present application, and the device may include the following modules.

[0213] A determination module 1001 is used to respond to a labeling request triggered for a target task and perform a multi-step reasoning operation on the target task using a first large model;

[0214] The first reasoning module 1002 is used to generate reasoning content of the reasoning step for the current reasoning step, and generate a reasoning result according to the reasoning content;

[0215] The sending module 1003 is used to output the inference content and inference results;

[0216] A determination module 1004, configured to determine whether the current reasoning step is a final reasoning step in response to the confirmation request;

[0217] The training data generation module 1005 is used to perform the next reasoning operation when the current reasoning step is an intermediate reasoning step; when the current reasoning step is a final reasoning step, the training data corresponding to the target task is obtained based on the reasoning content and reasoning results corresponding to multiple reasoning steps respectively; the training data is used to train the second largest model.

[0218] In some embodiments, the apparatus may further include:

[0219] The update module is used to update the inference content and the inference result in response to the update request.

[0220] In some embodiments, the annotation system integrates multiple candidate tools corresponding to the target task;

[0221] The first reasoning module 1002 may be specifically used to generate thinking information and action information of the reasoning step for the current reasoning step; the action information includes a tool identification of a target tool; according to the target tool and action parameters indicated by the action information, the target tool is called to obtain target data that meets the action parameters;

[0222] The sending module 1003 can be specifically used to output thinking information, action information and target data.

[0223] In some embodiments, the plurality of candidate tools are data interfaces;

[0224] The determination module 1001 may be specifically configured to respond to a labeling request triggered for a target task and perform a multi-step reasoning operation for the target task using the first large model for the sample object;

[0225] The first reasoning module 1002 may be specifically configured to call the target interface to obtain target data that meets the action parameters from the data set of the sample object according to the target interface and the action parameters indicated by the action information.

[0226] In some embodiments, when the current reasoning step is an intermediate reasoning step, the update module can be specifically used to update the reasoning content in response to an update request; based on the updated reasoning content, the reasoning result is updated using the first large model.

[0227] In some embodiments, the apparatus may further include:

[0228] A first providing module, used for providing first saving prompt information and second saving prompt information;

[0229] A first saving module, configured to save the updated reasoning content in response to a first saving request triggered by the first saving prompt information;

[0230] The update module can be specifically used to respond to the second save request triggered by the second save prompt information, update the inference result based on the updated inference content using the first large model, and save the updated inference content and inference result.

[0231] In some embodiments, the reasoning content may include thinking information, action content, and action parameters;

[0232] The update module can be specifically used to respond to an update request triggered by thinking information, provide selection prompt information of multiple candidate thinking information, update the thinking information based on the selected target candidate thinking information, and update the action content and / or action parameters based on the updated thinking information; respond to an update request triggered by action content, provide selection prompt information of multiple candidate action contents, update the action content based on the selected target candidate action content, and update the action parameters based on the updated action content; respond to an update request triggered by action parameters, provide selection prompt information of multiple candidate parameters, and update the action parameters based on the selected target candidate parameters.

[0233] In some embodiments, when the current reasoning step is the final reasoning step, the reasoning content includes thinking information, and the update module can be specifically used to respond to an update request for the thinking information, update the thinking information, and based on the updated reasoning content, update the reasoning result using the first large model; respond to an update request triggered for the reasoning result, update the reasoning result.

[0234] In some embodiments, the determination module 1001 can be specifically used to determine the processing sample corresponding to the target task in response to a labeling request triggered for the target task; and perform multi-step reasoning operations on the target task using the first large model combined with the processing sample.

[0235] In some embodiments, the first reasoning module 1002 may be specifically configured to generate reasoning content of a current reasoning step in combination with reasoning content and reasoning results of historical reasoning steps.

[0236] Fig.10 The data annotation device shown can be used to perform Figure 1 The data annotation method shown, its implementation principle and technical effect are not repeated here. Fig.10 One or more modules of the data annotation device shown can constitute Fig. 9 The server in the system architecture shown. The specific manner in which each module in the data labeling device in the above embodiment performs operations has been described in detail in the embodiment of the method, and will not be described in detail here.

[0237] like Fig.11As shown, it is a structural schematic diagram of an embodiment of a model training device provided by the present application, and the device may include the following modules.

[0238] The first acquisition module 1101 is used to acquire training data for the target task; wherein the training data can be obtained based on the reasoning content and reasoning results corresponding to the multi-step reasoning steps corresponding to the target task; the reasoning content and reasoning results corresponding to the multi-step reasoning steps are generated and updated or confirmed by the first large model performing multi-step reasoning operations on the sample object and the target task;

[0239] The training module 1102 is used to train the second largest model using the training data.

[0240] Among them, the second largest model can be used to respond to a processing request for a target task of the object to be processed, and perform a multi-step reasoning operation to obtain a processing result of the object to be processed.

[0241] Fig.11 The model training apparatus shown can be used to perform Figure 2 The model training method shown, its implementation principle and technical effect will not be repeated here. Fig.11 One or more modules of the model training device shown can constitute Fig. 9 The server in the system architecture shown. The specific manner in which each module performs operations in the model training device in the above embodiment has been described in detail in the embodiment of the method, and will not be described in detail here.

[0242] like Fig.12 As shown, it is a structural schematic diagram of an embodiment of a data inference device provided by the present application, and the device may include the following modules.

[0243] The second acquisition module 1201 is used to acquire a processing request for a target task of an object to be processed;

[0244] The second reasoning module 1202 is used to perform a multi-step reasoning operation based on the target task using the second largest model to obtain a processing result of the object to be processed;

[0245] The second providing module 1203 is used to output the processing result.

[0246] Fig.12 The data inference apparatus shown can be used to perform Figure 3 The data inference method shown, its implementation principle and technical effect are not repeated here. Fig.12 One or more modules of the data reasoning device shown may constitute Fig. 9The server in the system architecture shown. The specific manner in which each module performs operations in the data reasoning device in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0247] like Fig.13 As shown, it is a schematic diagram of the structure of an embodiment of a display device provided by the present application, and the device may include the following modules.

[0248] The first display module 1301 is used to display the annotation prompt information of the target task in the annotation interface;

[0249] The first generating module 1302 is used to generate a labeling request in response to a triggering operation on the labeling prompt information; the labeling request can be used to trigger the first large model to perform a multi-step reasoning operation on the target task, generate reasoning content of the reasoning step for the current reasoning step, and generate a reasoning result according to the reasoning content;

[0250] The second display module 1303 is used to display the reasoning content and reasoning result of the current reasoning step in the annotation interface;

[0251] The second generation module 1304 is used to generate a confirmation request in response to a confirmation operation; the confirmation request can be used to trigger the first large model to determine whether the current reasoning step is the final reasoning step, and when the current reasoning step is an intermediate reasoning step, perform the next reasoning operation, and when the current reasoning step is the final reasoning step, based on the reasoning content and reasoning results corresponding to the multiple reasoning steps, respectively, obtain training data corresponding to the target task, and the training data can be used to train the second large model.

[0252] In some embodiments, the apparatus may further include:

[0253] The third generation module is used to generate an update request in response to the update operation; the update request can be used to trigger the first large model to update the inference content and inference results.

[0254] In some embodiments, the apparatus may further include:

[0255] The third display module is used to display update prompt information in the annotation interface;

[0256] The third generating module may be specifically configured to determine update content in response to an update operation triggered by the update prompt information, and to generate an update request based on the update content.

[0257] In some embodiments, the apparatus may further include:

[0258] A fourth display module, used to display the first saving prompt information and the second saving prompt information in the annotation interface;

[0259] A fourth generating module, configured to generate a first saving request in response to a first saving operation triggered by the first saving prompt information, wherein the first saving request is used to trigger the first large model to save the updated reasoning content;

[0260] The third generation module can be specifically used to generate a second save request in response to a second save operation triggered by the second save prompt information, and the second save request is used to trigger the first large model to update the reasoning result based on the updated reasoning content, and save the updated reasoning content and reasoning result.

[0261] In some embodiments, the third generation module can be specifically used for, in the case where the update prompt information includes selection prompt information of multiple candidate thinking information, in response to a selection operation on any candidate thinking information, determining the target candidate thinking information, and generating an update request based on the target candidate thinking information, the update request is used to trigger the first large model to update the thinking information, and update the action content and / or action parameters based on the updated thinking information; in the case where the update prompt information includes selection prompt information of multiple candidate action contents, in response to a selection operation on any candidate action content, determining the target candidate action content, and generating an update request based on the target candidate action content, the update request is used to trigger the first large model to update the action content, and update the action parameters based on the updated action content; in the case where the update prompt information includes selection prompt information of multiple candidate parameters, in response to a selection operation on at least one candidate parameter, determining the target candidate parameter, and generating an update request based on the target candidate parameter, the update request is used to trigger the first large model to update the action parameters.

[0262] In some embodiments, the apparatus may further include:

[0263] The fifth display module is used to display confirmation prompt information in the annotation interface;

[0264] The second generating module 1304 may be specifically configured to generate a confirmation request in response to a confirmation operation triggered by the confirmation prompt information.

[0265] Fig.13 The display device shown can be used to perform Figure 4 The display method shown, its implementation principle and technical effect are not described in detail. Fig.13 One or more modules of the display device shown may constitute Fig. 9The annotation client in the system architecture shown. The specific manner in which each module of the display device in the above embodiment performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0266] like Fig.14 , which is a schematic diagram of a structure of an embodiment of a computing device provided by the present application, the device may include a storage component 1301 and a processing component 1302;

[0267] The storage component 1401 can be used to store one or more computer program instructions, wherein the one or more computer program instructions are called and executed by the processing component 1402 to implement Figure 1 The data annotation method shown, or Figure 2 The model training method shown, or Figure 3 The data reasoning method shown, or Figure 4 The display method shown.

[0268] Of course, the above computing device may also include other components, such as input / output interfaces, communication components, etc.

[0269] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc. The communication component is configured to facilitate wired or wireless communication between the computing device and other devices.

[0270] It should be noted that the above computing device implements Figure 1 The data annotation method shown, or Figure 2 The model training method shown, or Figure 3 In the case of the data reasoning method shown, it can be a physical device or an elastic computing host provided by a cloud computing platform, etc. It can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or as a single server or a single terminal device.

[0271] The computing device described above implements Figure 4 In the case of the display method shown, it can be implemented as an electronic device. The electronic device can refer to a device used by a user and having the functions of surfing the Internet, computing, communicating, etc. required by the user, such as a mobile phone, a tablet computer, a personal computer, a wearable device, etc. It can be understood that the above electronic device must also include a display component, an input / output interface, a communication component, and other components, which will not be repeated.

[0272] In one or more of the above embodiments, the processing component may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component can also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0273] The storage component is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0274] The display component may be an electroluminescent (EL) element, a liquid crystal display or a micro display having a similar structure, or a retinal direct display or a similar laser scanning display.

[0275] The present application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a computer, Figure 1 The data annotation method shown, or Figure 2 The model training method shown, or Figure 3 The data reasoning method shown, or Figure 4 The computer readable medium may be included in the computing device described in the above embodiments; or it may exist independently without being assembled into the computing device.

[0276] The computer-readable storage medium may be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above.

[0277] The present application also provides a computer program product, which includes a computer program carried on a computer-readable storage medium, and when the computer program is executed by a computer, Figure 1 The data annotation method shown, or Figure 2 The model training method shown, or Figure 3 The data reasoning method shown, or Figure 4 The display method shown.

[0278] In such an embodiment, the computer program may be downloaded and installed from a network, and / or installed from a removable medium. When the computer program is executed by a processor, various functions defined in the system of the present application are performed.

[0279] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0280] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0281] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0282] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data labeling method, characterized in that: Applied to a labeling system, the method comprises: In response to a labeling request triggered for a target task, performing a multi-step reasoning operation for the target task using the first large model; For the current reasoning step, generate the reasoning content of the reasoning step, and generate the reasoning result according to the reasoning content; Outputting the inference content and the inference result; In response to the confirmation request, determining whether the current reasoning step is a final reasoning step; When the current reasoning step is an intermediate reasoning step, the next reasoning operation is performed; When the current reasoning step is the final reasoning step, the training data corresponding to the target task is obtained based on the reasoning contents and reasoning results corresponding to the multiple reasoning steps respectively; the training data is used to train the second largest model.

2. The method according to claim 1, characterized in that After outputting the inference content and the inference result, the method further includes: In response to the update request, the inference content and the inference result are updated.

3. The method according to claim 1, characterized in that: The annotation system integrates a plurality of candidate tools corresponding to the target task; The step of generating the reasoning content of the reasoning step for the current reasoning step and generating the reasoning result according to the reasoning content includes: For the current reasoning step, generating thinking information and action information of the reasoning step; the action information includes a tool identification of a target tool; According to the target tool and action parameters indicated by the action information, calling the target tool to obtain target data that meets the action parameters; The outputting of the inference content and the inference result includes: The thought information, the action information, and the target data are output.

4. The method according to claim 3, characterized in that The plurality of candidate tools are data interfaces; In response to the labeling request triggered for the target task, performing a multi-step reasoning operation on the target task using the first large model includes: In response to a labeling request triggered for a target task, performing a multi-step reasoning operation for the target task using a first large model for the sample object; According to the target tool and the action parameters indicated by the action information, calling the target tool to obtain the target data that meets the action parameters includes: According to the target interface and action parameters indicated by the action information, the target interface is called to obtain target data that meets the action parameters from the data set of the sample object.

5. The method according to claim 2, characterized in that: In the case where the current reasoning step is an intermediate reasoning step, the updating of the reasoning content and the reasoning result in response to the update request includes: In response to the update request, updating the inference content; Based on the updated reasoning content, the reasoning result is updated using the first large model.

6. The method according to claim 5, characterized in that Also includes: Providing first saving prompt information and second saving prompt information; In response to a first save request triggered by the first save prompt information, save the updated reasoning content; The updating of the reasoning result by using the first large model based on the updated reasoning content includes: In response to a second save request triggered by the second save prompt information, based on the updated reasoning content, the reasoning result is updated using the first large model, and the updated reasoning content and reasoning result are saved.

7. The method according to claim 5, characterized in that The reasoning content includes thinking information, action content and action parameters; In response to the update request, updating the inference content includes: In response to an update request triggered by the thinking information, providing selection prompt information of a plurality of candidate thinking information, updating the thinking information based on the selected target candidate thinking information, and updating the action content and / or action parameter based on the updated thinking information; In response to an update request triggered for the action content, providing selection prompt information of a plurality of candidate action contents, updating the action content based on the selected target candidate action content, and updating the action parameters based on the updated action content; In response to an update request triggered for the action parameter, selection prompt information of a plurality of candidate parameters is provided, and the action parameter is updated based on the selected target candidate parameter.

8. The method according to claim 1, characterized in that In the case where the current reasoning step is the final reasoning step, the reasoning content includes thinking information, and the updating of the reasoning content and the reasoning result in response to the update request includes: In response to an update request for the thinking information, updating the thinking information, and updating the reasoning result using the first large model based on the updated thinking information; In response to an update request triggered for the inference result, the inference result is updated.

9. The method according to claim 1, characterized in that: In response to the labeling request triggered for the target task, performing a multi-step reasoning operation on the target task using the first large model includes: In response to a labeling request triggered for a target task, determining a processing example corresponding to the target task; The first large model is combined with the processing example to perform a multi-step reasoning operation on the target task.

10. The method according to claim 1, characterized in that For the current reasoning step, generating the reasoning content of the reasoning step includes: For the current reasoning step, the reasoning content and reasoning results of the historical reasoning steps are combined to generate the reasoning content of the reasoning step.

11. A model training method, characterized in that: include: Acquire training data for a target task, where the training data is obtained based on reasoning contents and reasoning results respectively corresponding to a plurality of reasoning steps corresponding to the target task; The reasoning contents and reasoning results respectively corresponding to the multi-step reasoning steps are generated and confirmed by the first large model performing multi-step reasoning operations on the sample object and the target task; Training a second model using the training data; The second large model is used to respond to a processing request for a target task of an object to be processed and perform a multi-step reasoning operation to obtain a processing result of the object to be processed.

12. A display method, characterized in that: include: Display the annotation prompt information of the target task in the annotation interface; In response to a triggering operation on the annotation prompt information, generating an annotation request; The annotation request is used to trigger the first large model to perform a multi-step reasoning operation on the target task, generate a reasoning content of the reasoning step for the current reasoning step, and generate a reasoning result according to the reasoning content; Displaying the reasoning content and reasoning result of the current reasoning step in the annotation interface; In response to the confirmation operation, a confirmation request is generated; the confirmation request is used to trigger the first large model to determine whether the current reasoning step is the final reasoning step, and when the current reasoning step is an intermediate reasoning step, perform the next reasoning operation, and when the current reasoning step is the final reasoning step, based on the reasoning content and reasoning results corresponding to the multiple reasoning steps, obtain the training data corresponding to the target task; the training data is used to train the second large model.

13. The method according to claim 12, characterized in that After displaying the reasoning content and the reasoning result of the current reasoning step in the annotation interface, the method further includes: In response to the update operation, an update request is generated; the update request is used to trigger the first large model to update the inference content and the inference result.

14. The method according to claim 13, characterized in that Also includes: Displaying update prompt information in the annotation interface; In response to the update operation, generating an update request includes: In response to an update operation triggered by the update prompt information, update content is determined, and an update request is generated based on the update content.

15. The method according to claim 13, characterized in that Also includes: Displaying first saving prompt information and second saving prompt information in the annotation interface; In response to a first save operation triggered by the first save prompt information, generating a first save request, wherein the first save request is used to trigger the first large model to save the updated inference content; The generating of the update request in response to the triggered update operation comprises: In response to a second save operation triggered by the second save prompt information, a second save request is generated, and the second save request is used to trigger the first large model to update the reasoning result based on the updated reasoning content, and save the updated reasoning content and reasoning result.

16. The method according to claim 14, characterized in that The reasoning content includes thinking information, action content and action parameters; The determining, in response to the update operation triggered by the update prompt information, update content, and generating an update request based on the update content includes: In the case where the update prompt information includes selection prompt information of multiple candidate thinking information, in response to a selection operation for any candidate thinking information, a target candidate thinking information is determined, and an update request is generated based on the target candidate thinking information, the update request is used to trigger the first large model to update the thinking information, and update the action content and / or action parameters based on the updated thinking information; When the update prompt information includes selection prompt information of a plurality of candidate action contents, in response to a selection operation on any candidate action content, a target candidate action content is determined, and an update request is generated based on the target candidate action content, wherein the update request is used to trigger the first large model to update the action content, and the action parameters are updated based on the updated action content; When the update prompt information includes selection prompt information for multiple candidate parameters, a target candidate parameter is determined in response to a selection operation on at least one candidate parameter, and an update request is generated based on the target candidate parameter, and the update request is used to trigger the first large model to update the action parameter.

17. The method according to claim 12, characterized in that Also includes: Displaying confirmation prompt information in the annotation interface; In response to the confirmation operation, generating a confirmation request includes: In response to a confirmation operation triggered by the confirmation prompt information, a confirmation request is generated.

18. A computing device, characterized in that It comprises a storage component and a processing component; the storage component stores one or more computer program instructions, the computer program instructions are called and executed by the processing component, and the processing component executes the one or more computer program instructions to implement the data labeling method according to any one of claims 1 to 10, or the model training method according to claim 11, or the display method according to any one of claims 12 to 17.

19. A computer-readable storage medium, characterized in that: A computer program is stored, and the computer program is executed by a computer to implement the data labeling method according to any one of claims 1 to 10, or the model training method according to claim 11, or the display method according to any one of claims 12 to 17.

20. A computer program product, characterized in that A computer program is stored, and when the computer program is executed by a computer, the data labeling method according to any one of claims 1 to 10, the model training method according to claim 11, or the display method according to any one of claims 12 to 17 is implemented.

Images