Large model Lora fine tuning method based on self-distillation micro transfer learning technology
Through self-distillation technology, the multi-round dialogue data set of vertical multi-round dialogue is solved, and the model is efficiently adaptable and multi-round dialogue stability in vertical scenarios is achieved.
Patent Information
- Application Number
- CN202411047320.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-12-13
AI Technical Summary
In vertical scenarios, the Lora fine-tuning model lacks multi-round dialogue annotation data, resulting in the model losing its ability to multi-round dialogue during fine-tuning, resulting in the problem of pattern crash.
Through self-distillation technology, a multi-round dialogue data set and a single-round vertical dialogue data set are constructed, and a multi-round vertical dialogue data set is constructed. It is used for Lora fine-tuning of the basic model to ensure that the model can learn vertical in vertical scenarios. Knowledge and maintain multi-round dialogue ability.
It realizes efficient adaptability and multi-round dialogue stability of large models in vertical scenarios, solves the problem of pattern collapse and loss of multi-round dialogue ability caused by Lora's single-round fine-tuning, and improves the practicality and adaptability of the model.
Smart Images
Figure CN119129700A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing technology, and in particular to a large-model Lora fine-tuning method, system, device and medium based on self-distillation micro-transfer learning technology. Background Art
[0002] As a method of fine-tuning large models, Lora fine-tuning is mainly used in vertical fields, such as pet scenarios. This technology fine-tunes professional vertical dialogue datasets on basic large models (such as chatglm, chatgpt, etc.) to improve the adaptability of the model in specific vertical scenarios. Vertical fields usually contain specific knowledge and data distribution, so fine-tuning is needed to enable the model to better understand and respond to user needs in the vertical field.
[0003] In Lora fine-tuning, the idea of introducing low-rank matrices is the key. This method effectively reduces the resources required for fine-tuning and improves the efficiency of fine-tuning by reducing the dimension of model parameters. Fine-tuning of large models for vertical scenarios involves in-depth learning of models in specific fields to better handle users' multi-round dialogue needs. The application scenarios of vertical fields are rich and diverse, including the pet industry, medical industry, and agricultural scenarios. However, in vertical scenarios, Lora fine-tuning technology faces some challenges. First, the data in vertical fields is usually limited and precious, and most of them exist in the knowledge base in the form of single-round questions and answers. Due to the expensiveness of multi-round dialogue annotation data, the fine-tuning process can often only use single-round question and answer data, resulting in a decrease in the performance of the model in multi-round dialogue. In addition, there is a problem with Lora fine-tuning, that is, the ability of the model before fine-tuning in multi-round dialogue may be lost during the fine-tuning process, causing the fine-tuned model to fall into mode collapse in user dialogue. This is due to the lack of multi-round dialogue annotation data for vertical scenarios, which causes the model to lose its original dialogue ability while learning vertical knowledge. Summary of the invention
[0004] In view of the defects in the prior art, the embodiments of the present invention provide a large-model Lora fine-tuning method, system, device and medium based on self-distillation micro-transfer learning technology to solve the technical problems existing in the prior art.
[0005] In a first aspect, an embodiment of the present invention provides a method for fine-tuning a large model Lora based on a self-distillation micro-transfer learning technology, comprising:
[0006] The basic model is self-distilled to obtain a public domain multi-round dialogue dataset;
[0007] Get the vertical single-round data set;
[0008] Construct a vertical multi-round dialogue dataset based on the public domain multi-round dialogue dataset and vertical single-round dataset;
[0009] The vertical multi-round dialogue dataset is used to fine-tune the Lora of the basic model to obtain the fine-tuned model.
[0010] In a second aspect, an embodiment of the present invention provides a large model Lora fine-tuning method based on a self-distillation micro-transfer learning technology, including:
[0011] The basic model is self-distilled to obtain a public domain multi-round dialogue dataset;
[0012] Get the vertical single-round data set;
[0013] The public domain multi-round dialogue dataset and the vertical category single-round dataset are merged to obtain a new dataset;
[0014] Use the new dataset to perform Lora fine-tuning on the base model to obtain the fine-tuned model.
[0015] In a third aspect, an embodiment of the present invention provides a system for fine-tuning a large model Lora based on a self-distillation micro-transfer learning technology, including: a self-distillation module, a data acquisition module, a data set processing module, and a Lora fine-tuning module;
[0016] The self-distillation module is used to perform self-distillation processing on the basic model to obtain a public domain multi-round dialogue dataset;
[0017] The data acquisition module is used to acquire a vertical single-round data set;
[0018] The data set processing module is used to merge the public domain multi-round dialogue data set and the vertical category single-round data set to obtain a new data set or to construct a vertical category multi-round dialogue data set based on the public domain multi-round dialogue data set and the vertical category single-round data set;
[0019] The Lora fine-tuning module uses a new dataset or a vertical multi-round dialogue dataset to perform Lora fine-tuning on the basic model to obtain a fine-tuned model.
[0020] In a fourth aspect, an embodiment of the present invention provides an intelligent device, comprising a processor, an input device, an output device and a memory, wherein the processor is connected to the input device, the output device and the memory, respectively, the memory is used to store a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute the method described in the above embodiment.
[0021] In a fifth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the method described in the above embodiment.
[0022] Beneficial effects of the present invention:
[0023] The embodiment of the present invention provides a large model Lora fine-tuning method based on self-distillation micro-transfer learning technology. By self-distilling the basic model, a public domain multi-round dialogue dataset is obtained. The self-distillation method adopted does not require additional costs, making data acquisition more economical and efficient. The rich dataset has more significant performance and practical value in the fine-tuning of the large model and its application in vertical scenarios. A vertical multi-round dialogue dataset is constructed based on the public domain multi-round dialogue dataset and the vertical single-round dataset, which improves the adaptability of the large model in vertical scenarios and successfully solves the problem of mode collapse and loss of multi-round dialogue ability caused by Lora single-round fine-tuning.
[0024] The embodiments of the present invention provide a large-model Lora fine-tuning system, device and medium based on self-distillation micro-transfer learning technology, which are based on the same inventive concept as the above-mentioned large-model Lora fine-tuning method based on self-distillation micro-transfer learning technology and have the same beneficial effects. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.
[0026] Figure 1 A flowchart of a large model Lora fine-tuning method based on a self-distillation micro-transfer learning technology provided by the first embodiment of the present invention is shown;
[0027] Figure 2 A flowchart of a large model Lora fine-tuning method based on a self-distillation micro-transfer learning technology provided by another embodiment of the present invention is shown;
[0028] Figure 3 A structural block diagram of a large model Lora fine-tuning system based on a micro-transfer learning technology based on self-distillation provided by another embodiment of the present invention is shown;
[0029] Figure 4 A structural block diagram of an intelligent device provided by another embodiment of the present invention is shown. DETAILED DESCRIPTION
[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0031] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.
[0032] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.
[0033] It should be further understood that the term "and / or" used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0034] As used in this specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.
[0035] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.
[0036] The embodiments of the present invention aim to solve the key problems existing in the prior art in fine-tuning vertical scenarios:
[0037] The problem of imbalance in multi-round dialogue capabilities before and after fine-tuning: In the prior art, due to the lack of multi-round dialogue annotation data for vertical scenarios, the fine-tuning process may cause the model to lose its ability in multi-round dialogue, causing it to experience mode collapse in user interaction. The goal of the present invention is to provide a technical solution to balance the learning of vertical knowledge and the maintenance of multi-round dialogue capabilities during the fine-tuning process, ensuring that the fine-tuned model can both understand the specific needs of vertical scenarios and effectively conduct multi-round dialogues.
[0038] Lack of multi-round dialogue datasets: Due to the lack of multi-round dialogue annotated data for vertical scenarios, the present invention focuses on how to construct an effective multi-round dialogue dataset during the fine-tuning process. When faced with limited multi-round dialogue data for vertical scenarios, the present invention provides an innovative method to construct based on existing vertical single-round dialogue datasets, thereby enriching the fine-tuning data and improving the performance of the model in multi-round dialogues.
[0039] like Figure 1 As shown, a flowchart of a large model Lora fine-tuning method based on self-distillation micro-transfer learning technology provided by another embodiment of the present invention is shown, and the method includes:
[0040] The basic model is self-distilled to obtain a public domain multi-round dialogue dataset;
[0041] Get the vertical single-round data set;
[0042] Construct a vertical multi-round dialogue dataset based on the public domain multi-round dialogue dataset and vertical single-round dataset;
[0043] The vertical multi-round dialogue dataset is used to fine-tune the Lora of the basic model to obtain the fine-tuned model.
[0044] Among them, model self-distillation: self-distillation of the basic model is to conduct self-dialogue through the basic model (such as chatglm, chatgpt, etc.) to simulate conversation situations in different scenarios. This process produces a rich and diverse multi-round dialogue dataset, which we call the public domain multi-round dialogue dataset. The public domain multi-round dialogue dataset contains multi-round dialogue data generated by the basic model in different scenarios. This dataset, as a key part of micro-transfer learning, provides a basis for fine-tuning of vertical scenarios. The goal of this step is to distill the performance capabilities of the basic model in multi-round dialogues into the dataset to provide a solid foundation for fine-tuning.
[0045] In order to better adapt to vertical scenarios, a vertical single-round dialogue dataset is introduced. This dataset focuses on single-round dialogues in vertical fields and provides key information for the subsequent construction of vertical multi-round dialogue datasets.
[0046] Construction of vertical multi-round dialogue dataset: Construct a vertical multi-round dialogue dataset based on the public domain multi-round dialogue dataset and the vertical single-round dataset.
[0047] Specifically, for each vertical single-round dialogue data, a multi-round dialogue data is randomly selected from the public domain multi-round dialogue data set, and a dialogue round is randomly selected from the multi-round dialogue data to obtain the vertical multi-round dialogue data set, successfully constructing the vertical multi-round dialogue data set, making the data set closer to the vertical scene while retaining the structure of the multi-round dialogue. The vertical multi-round dialogue data set plays a key role in the fine-tuning stage, making the model more applicable to vertical scenes.
[0048] Micro-transfer learning: Use the constructed vertical multi-round dialogue dataset to fine-tune the Lora of the basic model to obtain a fine-tuned model. The fine-tuned model can fully learn the knowledge of the vertical scenario and maintain stability and efficiency in multi-round dialogues.
[0049] like Figure 2 As shown, a flowchart of a large model Lora fine-tuning method based on a self-distillation micro-transfer learning technology provided by an embodiment of the present invention is shown, and the method includes:
[0050] The basic model is self-distilled to obtain a public domain multi-round dialogue dataset;
[0051] Get the vertical single-round data set;
[0052] The public domain multi-round dialogue dataset and the vertical category single-round dataset are merged to obtain a new dataset;
[0053] Use the new dataset to perform Lora fine-tuning on the base model to obtain the fine-tuned model.
[0054] The difference from the first embodiment is the generation of a new data set and micro-transfer learning.
[0055] In this embodiment, the public domain multi-round dialogue dataset and the vertical single-round dataset are merged to generate a new dataset. Micro-transfer learning: In the fine-tuning stage, the new dataset is used for Lora fine-tuning of the basic model to obtain a fine-tuned model. Lora fine-tuning ensures that the large model can better adapt to vertical scenarios after fine-tuning, while maintaining the fluency and stability of multi-round dialogues. The fine-tuned model is the basic model that has been fine-tuned for vertical scenarios, forming a more adaptable model that can maintain multi-round dialogue capabilities.
[0056] Advantages of a large model Lora fine-tuning method based on self-distillation micro-transfer learning technology provided by an embodiment of the present invention:
[0057] 1. Self-distillation does not require additional costs:
[0058] The self-distillation method is adopted to construct a public domain multi-round dialogue dataset by using the basic model to conduct self-dialogue in different scenarios. This process does not require additional annotation costs or manual intervention, but utilizes the capabilities of the basic model itself, providing an efficient and economical data foundation for subsequent fine-tuning. Compared with traditional data annotation methods, self-distillation is more flexible and cost-effective in data acquisition, and has significant advantages over existing technologies.
[0059] 2. The constructed vertical multi-round dialogue data is of high value:
[0060] Through innovative model self-distillation and vertical multi-round dialogue dataset construction methods, we have successfully constructed a more abundant and closer to actual vertical multi-round dialogue dataset. Compared with traditional single-round dialogue data, this dataset contains more complex and varied multi-round dialogue scenarios, providing more challenging and valuable training data for fine-tuning large models. This rich dataset helps the model understand and adapt to vertical scenarios more comprehensively, improving the performance of the model in practical applications.
[0061] 3. Solved the problem of Lora's single-round fine-tuning causing mode collapse and loss of multi-round dialogue capabilities:
[0062] In response to the challenges of using Lora single-round fine-tuning in vertical scenario fine-tuning, two micro-transfer learning schemes are proposed. By merging the public domain multi-round dialogue dataset and the vertical single-round dataset for Lora fine-tuning, or directly using the constructed vertical multi-round dialogue dataset for fine-tuning, the problem of the model's multi-round dialogue ability being lost during the fine-tuning process is successfully solved. This enables the model to not only learn the knowledge of vertical scenarios, but also maintain stability and fluency in multi-round dialogues, improving the practicality and adaptability of the model.
[0063] In the above-mentioned first embodiment, a large model Lora fine-tuning method based on self-distillation micro-transfer learning technology is provided. Correspondingly, the present application also provides a large model Lora fine-tuning system based on self-distillation micro-transfer learning technology. Please refer to Figure 3 , which is a structural block diagram of a large model Lora fine-tuning system based on self-distillation micro-transfer learning technology provided by another embodiment of the present invention. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described below is only illustrative.
[0064] like Figure 3As shown, another embodiment of the present invention provides a large-model Lora fine-tuning system based on self-distillation micro-transfer learning technology, including: a self-distillation module, a data acquisition module, a dataset processing module and a Lora fine-tuning module; the self-distillation module is used to perform self-distillation processing on the basic model to obtain a public domain multi-round dialogue dataset; the data acquisition module is used to obtain a vertical single-round dataset; the dataset processing module is used to merge the public domain multi-round dialogue dataset and the vertical single-round dataset to obtain a new dataset or construct a vertical multi-round dialogue dataset based on the public domain multi-round dialogue dataset and the vertical single-round dataset; the Lora fine-tuning module uses the new dataset or the vertical multi-round dialogue dataset to perform Lora fine-tuning of the basic model to obtain a fine-tuned model.
[0065] Among them, the self-distillation module includes a public dataset construction unit, which is used to obtain the self-dialogue data of the basic model in different scenarios, and construct a public domain multi-round dialogue dataset based on the self-dialogue data.
[0066] The dataset processing module includes a vertical multi-round dialogue dataset construction unit. For each vertical single-round dialogue data, the vertical multi-round dialogue dataset construction unit randomly selects a multi-round dialogue data from the public domain multi-round dialogue dataset, and randomly selects a dialogue round in the multi-round dialogue data to obtain the vertical multi-round dialogue dataset.
[0067] The above is a description of an embodiment of a large-model Lora fine-tuning system based on self-distillation micro-transfer learning technology provided by another embodiment of the present invention.
[0068] The embodiment of the present invention provides a large model Lora fine-tuning system based on self-distillation micro-transfer learning technology. The public domain multi-round dialogue dataset is obtained by self-distilling the basic model. The self-distillation method adopted does not require additional cost, making data acquisition more economical and efficient. The rich dataset has more significant performance and practical value in the fine-tuning of the large model and application in vertical scenarios. A vertical multi-round dialogue dataset is constructed based on the public domain multi-round dialogue dataset and the vertical single-round dataset, which improves the adaptability of the large model in vertical scenarios and successfully solves the problem of mode collapse and loss of multi-round dialogue ability caused by Lora single-round fine-tuning.
[0069] like Figure 4 As shown, a structural block diagram of an intelligent device according to another embodiment of the present invention is shown, the device includes a processor, an input device, an output device and a memory, the processor is connected to the input device, the output device and the memory respectively, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the method described in the above embodiment.
[0070] It should be understood that in the embodiments of the present invention, the processor referred to may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.
[0071] Input devices may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint direction information), a microphone, etc., and output devices may include a display (LCD, etc.), a speaker, etc.
[0072] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.
[0073] In a specific implementation, the processor, input device, and output device described in the embodiments of the present invention may execute the implementation described in the method embodiment provided in the embodiments of the present invention, or may execute the implementation described in the system embodiment described in the embodiments of the present invention, which will not be described in detail here.
[0074] The present invention further provides an embodiment of a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the method described in the above embodiment.
[0075] The computer-readable storage medium may be an internal storage unit of the terminal described in the foregoing embodiments, such as a hard disk or memory of the terminal. The computer-readable storage medium may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Further, the computer-readable storage medium may also include both an internal storage unit of the terminal and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by the terminal. The computer-readable storage medium may also be used to temporarily store data that has been output or is to be output.
[0076] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0077] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the terminals and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here.
[0078] In the several embodiments provided in the present application, it should be understood that the disclosed terminal and method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.
[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention 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 replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.
Claims
1. A large model Lora fine-tuning method based on self-distillation micro-transfer learning technology, characterized in that: include: The basic model is self-distilled to obtain a public domain multi-round dialogue dataset; Get the vertical single-round data set; Construct a vertical multi-round dialogue dataset based on the public domain multi-round dialogue dataset and vertical single-round dataset; The vertical multi-round dialogue dataset is used to fine-tune the Lora of the basic model to obtain the fine-tuned model.
2. The method according to claim 1, characterized in that The specific method of performing self-distillation processing on the basic model includes: Obtain self-conversation data of the basic model in different scenarios; Construct a multi-round dialogue dataset based on self-dialogue data.
3. The method according to claim 1 or 2, characterized in that The specific method of constructing a vertical multi-round dialogue dataset based on a public domain multi-round dialogue dataset and a vertical single-round dialogue dataset includes: For each vertical category of single-round dialogue data, a multi-round dialogue data set is randomly selected from the public domain multi-round dialogue data set, and a dialogue round is randomly selected from the multi-round dialogue data to obtain the vertical category multi-round dialogue data set.
4. A large model Lora fine-tuning method based on self-distillation micro-transfer learning technology, characterized in that: include: The basic model is self-distilled to obtain a public domain multi-round dialogue dataset; Get the vertical single-round data set; The public domain multi-round dialogue dataset and the vertical single-round dataset are merged to obtain a new dataset; Use the new dataset to perform Lora fine-tuning on the base model to obtain the fine-tuned model.
5. The method according to claim 4, characterized in that The specific method of performing self-distillation processing on the basic model includes: Obtain self-conversation data of the basic model in different scenarios; Construct a public domain multi-turn dialogue dataset based on self-dialogue data.
6. A large-model Lora fine-tuning system based on self-distillation micro-transfer learning technology, characterized in that: include: Self-distillation module, data acquisition module, dataset processing module and Lora fine-tuning module; The self-distillation module is used to perform self-distillation processing on the basic model to obtain a public domain multi-round dialogue dataset; The data acquisition module is used to acquire a vertical single-round data set; The data set processing module is used to merge the public domain multi-round dialogue data set and the vertical category single-round data set to obtain a new data set or to construct a vertical category multi-round dialogue data set based on the public domain multi-round dialogue data set and the vertical category single-round data set; The Lora fine-tuning module uses a new dataset or a vertical multi-round dialogue dataset to perform Lora fine-tuning on the basic model to obtain a fine-tuned model.
7. The system according to claim 6, characterized in that The self-distillation module includes a public dataset construction unit, which is used to obtain self-dialogue data of the basic model in different scenarios, and construct a public domain multi-round dialogue dataset based on the self-dialogue data.
8. The system according to claim 6 or 7, characterized in that The data set processing module includes a vertical multi-round dialogue data set construction unit. For each vertical single-round dialogue data, the vertical multi-round dialogue data set construction unit randomly selects a multi-round dialogue data from a public domain multi-round dialogue data set, and randomly selects a dialogue round in the multi-round dialogue data to obtain a vertical multi-round dialogue data set.
9. An intelligent device, comprising a processor, an input device, an output device and a memory, wherein the processor is connected to the input device, the output device and the memory respectively, and the memory is used to store a computer program, wherein the computer program comprises program instructions, and wherein: The processor is configured to call the program instructions to execute the method according to any one of claims 1-5.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 5.