Method and device for constructing large model data set based on few samples

Abstract

The invention relates to the technical field of large models, in particular to a method and device for constructing a large model data set based on a small number of samples. According to the method for constructing the large model data set based on the few samples, an original sample is obtained, the original sample is divided into independent data blocks in a user-defined mode, and key data blocks are selected from the independent data blocks in a user-defined mode after preliminary labeling; defining a downstream task instruction, and customizing an instruction format according to a task type; sorting object elements, behavior elements and item elements in the business, and respectively expanding and augmenting the sorted object elements, behavior elements and item elements by using a large model; and generating a positive example and a counter example, and labeling the positive example and the counter example to ensure that the large model can cope with different business situations and edge situations. According to the method and device for constructing the large model data set based on the small number of samples, the accuracy of business processing is improved, the effectiveness of the finally generated industry model during business processing can be ensured, and therefore powerful support is provided for training and application of the large model.

Description

Technical Field

[0001] The present invention relates to the field of large model technology, and in particular to a method and device for constructing a large model data set based on a small number of samples. Background Art

[0002] With the continuous deepening of information and intelligent construction, large models play an increasingly important role in business processing. They have been widely used in many fields such as automated grassroots reporting, telephone label classification, document report generation, and intelligent online customer service.

[0003] However, the business processing capabilities of general large models are highly dependent on high-quality corpus training. Without sufficient industry-specific data, the accuracy of the model's business processing will be greatly reduced. When labeling various types of data, it is necessary to ensure the quality of the data and the accuracy of the labeling to ensure the effectiveness of large-scale model training and the optimization of model performance.

[0004] In the current research field, the annotation and corpus generation of unstructured text to support large-scale model training can be divided into three main categories according to the size of the sample data set: massive sample data annotation methods, medium-sized sample data annotation methods, and small sample data annotation methods.

[0005] Massive sample data labeling method: usually applied to operation and maintenance logs or a large number of log files, aiming to generate a large amount of high-quality sample data through manual labeling. This type of labeling process is usually targeted at specific downstream tasks of large models, with clear and single goals. The labeling work mainly relies on experienced manual labelers who have rich business knowledge. Through simple data cleaning, association, and format reconstruction operations, a data set suitable for downstream tasks of large models (such as incremental pre-training, reinforced feedback, etc.) can be quickly constructed.

[0006] Medium-sized sample data labeling method: Applicable to industry data sets of a certain scale, which usually come from manually reported data in grassroots industry governance business. Unlike massive sample data, these data are not generated in real time, but they are also of a certain scale. Data processing processes, such as classification or response, are usually based on industry work instructions. When processing this type of data, it is necessary to carefully split the work instructions, combine them with existing manually labeled data samples, and use large models to perform appropriate data augmentation to generate a large number of training data sets.

[0007] Method for labeling a small amount of sample data: When the amount of data is extremely limited, such as when there is only one job or business instruction, or even a lack of data samples, it is difficult for business personnel to timely and effectively convert their business processing experience into a corpus for large model training if the above two methods are used. In this case, it is usually necessary to carefully split the job instructions and use a large model with auxiliary annotation to understand the relevant content and generate sample data. However, this process is highly dependent on the accuracy of the splitting and data combing rules, otherwise the generated sample data may deviate significantly from the rules in the business instruction, thereby affecting the accuracy of the large model in business processing.

[0008] Currently, most methods for building datasets for a small number of samples focus on the infrastructure of large models, aiming to maintain their generalization capabilities. A common practice is to use in-context learning to annotate datasets. However, this method often ignores the organization and creation of datasets from a business perspective, and lacks interaction with experienced business personnel. This makes it difficult to build high-quality corpora that meet the training needs of large models when industry data is limited.

[0009] In response to the challenge of limited industry data, the present invention proposes a method and device for constructing a large model dataset based on a small number of samples, aiming to quickly convert limited data or samples into high-quality corpus required for the large model. Summary of the invention

[0010] In order to overcome the defects of the prior art, the present invention provides a simple and efficient method and device for constructing a large model data set based on a small number of samples.

[0011] The present invention is achieved through the following technical solutions:

[0012] A method for constructing a large model dataset based on a small number of samples, comprising the following steps:

[0013] Step S1: Obtaining the original sample

[0014] Business personnel first input the original business processing rules or work instructions as samples. The samples include the specific processes and regulations of business operations and are the basis for building data sets.

[0015] Step S2: Preliminary labeling of samples

[0016] Business personnel customize the segmentation of original samples by chapter or use the existing unstructured text segmenter to extract the feature information of the original samples and convert them into several independent data blocks to generate annotation features;

[0017] Business personnel preliminarily mark the split independent data blocks and select key data blocks from them;

[0018] The key data blocks are steps, regulations or conditions in an independent business processing flow.

[0019] Step S3: Clarify downstream task instructions

[0020] Clarify downstream task instructions and customize the instruction format according to the task type; for example, if an industry big model is used for classification tasks, the instruction format can be "Please classify the following XX" or similar expressions; if it is a question-and-answer task, the instruction format can be "Please answer based on XX question"; if it is an audit task, the instruction format can be "Please audit according to XX rules", etc.

[0021] After clarifying the relevant downstream task instructions, construct a dataset in alpaca format in a targeted manner;

[0022] Step S4: sorting out object elements

[0023] Business personnel sort out the object elements in the business, including the people, time, place and objects involved in the business, and further divide each data block that has been divided according to the dimensions of people, time, place or objects to expand the number of data blocks;

[0024] Step S5: Augmentation of the object element model

[0025] Use the big model to expand and augment the sorted object elements, generate operation behaviors in different situations for different objects, and expand the number of data blocks;

[0026] In step S5, the sorted object elements are expanded and augmented based on the operating behaviors of the object (person) at different time points, the processing steps of the object (object) in different business scenarios, and the behavior model of the object (time) in a specific situation.

[0027] Step S6: Sorting out behavioral elements

[0028] After the big model augments the object elements, business personnel collaborate with the big model to conduct a basic sorting of the behavior elements, clarifying the behavior of each object element one by one, including the behavior of different objects in different situations and their impact; by carefully sorting out the behavior in business scenarios, the coverage of the data set is further expanded, and future data samples are generated and labeled based on the behavior;

[0029] Step S7: Augmentation of the behavioral factor model

[0030] After completing the basic sorting of behavioral elements, the big model further expands the behavioral elements according to the triggering conditions and execution order of the behaviors in the custom scenarios, generates more detailed behaviors, and further expands the number of data blocks;

[0031] In step S7, the large model refines the behavior elements to generate specific operation descriptions under different conditions, and generates behavior objects under different situations based on time and condition changes, thereby realizing behavior element augmentation;

[0032] When expanding behavioral elements, use the Few Shot Prompt method to improve the prompt words based on the basic behaviors previously sorted out.

[0033] Step S8: Sorting out the elements of the matter

[0034] After the behavioral elements are expanded, business personnel collaborate with the big model to sort out the event elements and generate a complete list of operational items by describing in detail the processing steps of each item in the business.

[0035] The item elements refer to specific operational events or process links in the business, which are further refined to clarify the specific items in the business operation;

[0036] Step S9: Augmentation of the event element model

[0037] After manually sorting out the elements of the matter, the big model further augments the elements of the matter according to the specific operational matters and detailed processing steps in the custom scenario;

[0038] In step S9, the big model generates specific operational items based on the sorted time behaviors, generates detailed processing steps for each item, and ensures that all key links in the business process are covered to achieve item element augmentation.

[0039] Step S10: Positive and negative example generation

[0040] The big model further generates positive examples (positive samples) and negative examples (negative samples) based on the event elements; positive examples refer to the standard operating procedures in the business, and negative examples include scenarios where abnormal processing or incorrect operations are discovered;

[0041] While generating positive and negative examples, the data is annotated, labels are generated for positive and negative examples based on classification tasks, and large models are used to generate answers for answering tasks to ensure that the large model can cope with a variety of business scenarios and edge cases.

[0042] A device for constructing a large model data set based on a small number of samples, comprising:

[0043] The original sample input module is responsible for providing a channel for business personnel to input original business processing rules or work instructions as samples. The samples include specific processes and regulations of business operations and are the basis for building data sets.

[0044] The sample preliminary annotation module is responsible for helping business personnel to segment the original sample by chapter or use the existing unstructured text word segmenter, extract the feature information of the original sample, and convert it into several independent data blocks, perform preliminary annotation on the segmented independent data blocks, and select key data blocks from them by definition;

[0045] The key data blocks are steps, regulations or conditions in an independent business processing flow.

[0046] The downstream task instruction clarification module is responsible for clarifying the downstream task instructions and customizing the instruction format according to the task type. For example, if an industry big model is used for classification tasks, the instruction format can be "Please classify the following XX" or similar expressions; if it is a question-and-answer task, the instruction format can be "Please answer according to XX question"; if it is an audit task, the instruction format can be "Please audit according to XX rules", etc.

[0047] After clarifying the relevant downstream task instructions, we specifically build a dataset in the alpaca format;

[0048] The object element sorting module is responsible for sorting out the object elements in the business, including the people, time, place and objects involved in the business, and further dividing each data block that has been divided according to the dimensions of people, time, place or objects to achieve the expansion of the number of data blocks;

[0049] The object element large model augmentation module is responsible for expanding and augmenting the sorted object elements using the large model, generating operation behaviors in different situations for different objects, and expanding the number of data blocks;

[0050] The object element large model augmentation module expands and augments the sorted object elements based on the operating behaviors of objects (people) at different time points, the processing steps of objects (objects) in different business scenarios, and the behavior models of objects (time) in specific situations.

[0051] The behavior factor combing module is responsible for combing the behavior factors after the big model has augmented the object factors. Through the collaboration between business personnel and the big model, the module clarifies the behavior of each object factor one by one, including the behavior of different objects in different situations and their impact. Through the detailed combing of the behavior in the business scenario, the module further expands the coverage of the data set and generates and annotates future data samples based on the behavior.

[0052] The behavior factor big model augmentation module is responsible for further augmenting the behavior factors based on the triggering conditions and execution sequence of behaviors in custom scenarios after completing the basic sorting of the behavior factors, generating more detailed behaviors and further expanding the number of data blocks;

[0053] The behavior factor large model augmentation module uses the large model to refine the behavior factors, generate specific operation descriptions under different conditions, and generate behavior objects under different situations based on time and condition changes to achieve behavior factor augmentation;

[0054] When expanding behavioral elements, use the Few Shot Prompt method to improve the prompt words based on the basic behaviors previously sorted out.

[0055] The item element sorting module is responsible for sorting out the item elements through the collaboration between business personnel and the big model after the behavioral element augmentation is completed, and generating a complete list of operational items by describing the processing steps of each item in the business in detail;

[0056] The item elements refer to specific operational events or process links in the business, which are further refined to clarify the specific items in the business operation;

[0057] The item element big model augmentation module is responsible for further augmenting the item elements based on the specific operation items and detailed processing steps in the custom scenario using the big model after the item elements are manually sorted out;

[0058] The item element big model augmentation module uses the big model to generate specific operation items based on the sorted time behavior, generates detailed processing steps for each item, and ensures that all key links in the business process are covered to achieve item element augmentation.

[0059] The positive and negative example generation module is responsible for using the big model to further generate positive examples (positive samples) and negative examples (negative samples) based on the event elements. Positive examples refer to the standard operating procedures in the business, and negative examples include scenarios where abnormal processing or erroneous operations are discovered.

[0060] While generating positive and negative examples, the data is annotated, labels are generated for positive and negative examples based on classification tasks, and large models are used to generate answers for answering tasks to ensure that the large model can cope with a variety of business scenarios and edge cases.

[0061] A computing device for constructing a large model data set based on a small number of samples, characterized by comprising:

[0062] A computer-readable storage medium storing one or more programs, characterized in that: the one or more programs include instructions, which, when executed by a computing device that constructs a large model data set based on a small number of samples, enable the computing device that constructs a large model data set based on a small number of samples to execute any of the above methods.

[0063] A computing device for constructing a large model data set based on a small number of samples, characterized by comprising:

[0064] One or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the above methods.

[0065] The beneficial effects of the present invention are as follows: the method and device for constructing a large model data set based on a small number of samples can construct data sets for specific business scenarios, quickly generate high-quality corpora that meet the industry's large model training needs, improve the accuracy of business processing, and ensure the effectiveness of the final generated industry model in processing business, thereby providing strong support for the training and application of large models. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0067] Attached Figure 1 Schematic diagram of the method of the present invention for constructing a large model data set based on a small number of samples.

[0068] Attached Figure 2 Schematic diagram of the method for generating business combing and annotation examples for a small number of samples of the present invention. DETAILED DESCRIPTION

[0069] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0070] The method for constructing a large model data set based on a small number of samples includes the following steps:

[0071] Step S1: Obtaining the original sample

[0072] First, input the original business processing rules or work instructions as a sample (data volume: 1). The sample includes the specific process and regulations of business operations, which is the basis for constructing the data set.

[0073] Step S2: Preliminary labeling of samples

[0074] Business personnel customize the segmentation of the original sample by chapter or use the existing unstructured text segmenter to extract the feature information of the original sample and convert it into several independent data blocks to generate annotation features, for example, split it into 10 blocks;

[0075] Business personnel preliminarily mark the split independent data blocks and select several key data blocks from them;

[0076] The key data blocks are steps, regulations or conditions in an independent business processing flow.

[0077] Step S3: Clarify downstream task instructions

[0078] Clarify downstream task instructions and customize the instruction format according to the task type; for example, if an industry big model is used for classification tasks, the instruction format can be "Please classify the following XX" or similar expressions; if it is a question-and-answer task, the instruction format can be "Please answer based on XX question"; if it is an audit task, the instruction format can be "Please audit according to XX rules", etc.

[0079] After clarifying the relevant downstream task instructions, a dataset in the Alpaca format is constructed specifically; the Alpaca format has been widely used in the construction of datasets for fine-tuning large model instruction sets.

[0080] Step S4: sorting out object elements

[0081] Sorting out the object elements in the business, including the people, time, place and objects involved in the business, and clarifying the object elements will help the large model to be further expanded in the subsequent augmentation process. By sorting out these elements, each data block that has been segmented can be further divided according to the dimensions of people, time, place or objects to achieve the expansion of the number of data blocks; for example, assuming there are currently 10 data blocks, after further refinement, the number of data blocks will be expanded to 1,000.

[0082] Step S5: Augmentation of the object element model

[0083] Use the big model to expand and augment the sorted object elements, generate operation behaviors in different situations for different objects, and expand the number of data blocks;

[0084] In step S5, the sorted object elements are expanded and augmented based on the operating behaviors of the object (person) at different time points, the processing steps of the object (object) in different business scenarios, and the behavior model of the object (time) in a specific situation.

[0085] For example, through this step, the number of data blocks can be expanded from 1000 to 2000. The following are examples of prompt words using the large model:

[0086] prompt="""

[0087] You are an XX business book analysis assistant. Please generate the corresponding objects according to the following requirements:

[0088] - Based on the object (person), generate its operation behavior at different time points.

[0089] -Please generate the processing steps of objects (items) in different business scenarios.

[0090] -Generate a model of the behavior of an object (time) in a specific situation.

[0091] """

[0092] Step S6: Sorting out behavioral elements

[0093] After the big model augments the object elements, the business personnel collaborate with the big model to conduct a basic combing of the behavior elements, clarifying the behavior of each object element one by one, including the behavior of different objects in different situations and their impact; by carefully combing the behavior in the business scenario, the coverage of the data set is further expanded; the following is an example of time behavior combing: the time object of XX business - before the declaration time, after the completion time, before the deadline, etc. In this way, future data samples are further generated and labeled based on behavior;

[0094] Step S7: Augmentation of the behavioral factor model

[0095] After completing the basic sorting of behavioral elements, the big model further expands the behavioral elements according to the triggering conditions and execution order of the behaviors in the custom scenarios, generates more detailed behaviors, and further expands the number of data blocks;

[0096] In step S7, the large model refines the behavior elements to generate specific operation descriptions under different conditions, and generates behavior objects under different situations based on time and condition changes, thereby realizing behavior element augmentation;

[0097] When expanding behavioral elements, use the Few Shot Prompt method to improve the prompt words based on the basic behaviors previously sorted out.

[0098] Through the expansion of the big model, the data volume is further expanded to 4,000. The following are examples of prompt words constructed using the big model:

[0099] prompt="""

[0100] You are an XX business book behavior analysis assistant. Please generate the corresponding behavior element augmentation according to the following requirements:

[0101] -Please refine the behavioral elements and generate specific operation descriptions under different conditions.

[0102] -Generate behavioral objects in different situations based on time and condition changes.

[0103] -Please refer to the following behavior example: XXX.XXX

[0104] """

[0105] Step S8: Sorting out the elements of the matter

[0106] After the behavioral elements are expanded, business personnel collaborate with the big model to sort out the event elements and generate a complete list of operational items by describing in detail the processing steps of each item in the business.

[0107] The item elements refer to specific operational events or process links in the business, which are further refined to clarify the specific items in the business operation;

[0108] For example, taking the time behavior previously sorted out as an example, assuming that XXX operation is performed before the deadline for XX business declaration, this operation can be expanded into a clear matter.

[0109] Step S9: Augmentation of the event element model

[0110] After manually sorting out the elements of the matter, the big model further augments the elements of the matter according to the specific operational matters and detailed processing steps in the custom scenario;

[0111] In step S9, the big model generates specific operational items based on the sorted time behaviors, generates detailed processing steps for each item, and ensures that all key links in the business process are covered to achieve item element augmentation.

[0112] The prompt words for the augmentation are as follows:

[0113] prompt="""

[0114] You are an XX business book analysis assistant. Please generate the corresponding event elements according to the following requirements:

[0115] -Generate specific action items based on the sorted time behavior.

[0116] -For example, XXX operation was performed before the XX business declaration deadline. Please expand this operation into specific matters.

[0117] - Please generate detailed processing steps for each matter and ensure that all key links in the business process are covered.

[0118] -Please refer to the following behavior example: XXX.XXX

[0119] """

[0120] Step S10: Positive and negative example generation

[0121] The big model generates specific and visual examples based on abstract matter descriptions. The generated examples should be vivid and clear, showing the complete process and possible results of each matter.

[0122] The big model further generates positive examples (positive samples) and negative examples (negative samples) based on the event elements; positive examples refer to the standard operating procedures in the business, and negative examples include scenarios where abnormal processing or incorrect operations are discovered;

[0123] While generating positive and negative examples, we annotate the data, generate labels for positive and negative examples according to the classification task, and use the big model to generate answers for answering tasks; for example, according to the expansion of the above data, the data volume is eventually expanded to 160,000, ensuring that the big model can cope with a variety of business scenarios and edge cases. The prompt word format used is as follows:

[0124] prompt="""

[0125] You are an assistant for data annotation generation in the XX industry. Please generate positive and negative samples according to the following requirements:

[0126] -Generate positive samples: Please describe the correct operation steps and results in the standard business process.

[0127] - Generate negative samples: Please describe the process deviations or problems that may occur under abnormal or incorrect operations.

[0128] - Ensure that positive and negative samples cover the same business scenarios and show the comparison between correct and incorrect processing respectively.

[0129] """

[0130] Among them, positive and negative examples can also allow the big model to generate specific and visual examples based on abstract event descriptions, for example: XX person went to XX hall on XX day to do XX event, and XX problem occurred. The prompt words used are as follows:

[0131] prompt="""

[0132] You are an assistant for analyzing XX business books. Please generate specific and visual examples based on the abstract description of the matter:

[0133] -For example: "XX person went to XX hall to handle XX matters on XX day, and XX problem occurred."

[0134] -Please describe the event in detail, including participants, time, location, operation steps and results.

[0135] - The examples generated should be vivid and clear, showing the complete process and possible results of each matter.

[0136] """

[0137] The device for constructing a large model data set based on a small number of samples includes:

[0138] The original sample input module is responsible for providing a channel for business personnel to input original business processing rules or work instructions as samples. The samples include specific processes and regulations of business operations and are the basis for building data sets.

[0139] The sample preliminary annotation module is responsible for helping business personnel to segment the original sample by chapter or use the existing unstructured text word segmenter, extract the feature information of the original sample, and convert it into several independent data blocks, perform preliminary annotation on the segmented independent data blocks, and select key data blocks from them by definition;

[0140] The key data blocks are steps, regulations or conditions in an independent business processing flow.

[0141] The downstream task instruction clarification module is responsible for clarifying the downstream task instructions and customizing the instruction format according to the task type. For example, if an industry big model is used for classification tasks, the instruction format can be "Please classify the following XX" or similar expressions; if it is a question-and-answer task, the instruction format can be "Please answer according to XX question"; if it is an audit task, the instruction format can be "Please audit according to XX rules", etc.

[0142] After clarifying the relevant downstream task instructions, we specifically build a dataset in the alpaca format;

[0143] The object element sorting module is responsible for sorting out the object elements in the business, including the people, time, place and objects involved in the business, and further dividing each data block that has been divided according to the dimensions of people, time, place or objects to achieve the expansion of the number of data blocks;

[0144] The object element large model augmentation module is responsible for expanding and augmenting the sorted object elements using the large model, generating operation behaviors in different situations for different objects, and expanding the number of data blocks;

[0145] The object element large model augmentation module expands and augments the sorted object elements based on the operating behaviors of objects (people) at different time points, the processing steps of objects (objects) in different business scenarios, and the behavior models of objects (time) in specific situations.

[0146] The behavior factor combing module is responsible for combing the behavior factors after the big model has augmented the object factors. Through the collaboration between business personnel and the big model, the module clarifies the behavior of each object factor one by one, including the behavior of different objects in different situations and their impact. Through the detailed combing of the behavior in the business scenario, the module further expands the coverage of the data set and generates and annotates future data samples based on the behavior.

[0147] The behavior factor big model augmentation module is responsible for further augmenting the behavior factors based on the triggering conditions and execution sequence of behaviors in custom scenarios after completing the basic sorting of the behavior factors, generating more detailed behaviors and further expanding the number of data blocks;

[0148] The behavior factor large model augmentation module uses the large model to refine the behavior factors, generate specific operation descriptions under different conditions, and generate behavior objects under different situations based on time and condition changes to achieve behavior factor augmentation;

[0149] When expanding behavioral elements, use the Few Shot Prompt method to improve the prompt words based on the basic behaviors previously sorted out.

[0150] The item element sorting module is responsible for sorting out the item elements through the collaboration between business personnel and the big model after the behavioral element augmentation is completed, and generating a complete list of operational items by describing the processing steps of each item in the business in detail;

[0151] The item elements refer to specific operational events or process links in the business, which are further refined to clarify the specific items in the business operation;

[0152] The item element big model augmentation module is responsible for further augmenting the item elements based on the specific operation items and detailed processing steps in the custom scenario using the big model after the item elements are manually sorted out;

[0153] The item element big model augmentation module uses the big model to generate specific operation items based on the sorted time behavior, generates detailed processing steps for each item, and ensures that all key links in the business process are covered to achieve item element augmentation.

[0154] The positive and negative example generation module is responsible for using the big model to further generate positive examples (positive samples) and negative examples (negative samples) based on the event elements. Positive examples refer to the standard operating procedures in the business, and negative examples include scenarios where abnormal processing or erroneous operations are discovered.

[0155] While generating positive and negative examples, the data is annotated, labels are generated for positive and negative examples based on classification tasks, and large models are used to generate answers for answering tasks to ensure that the large model can cope with a variety of business scenarios and edge cases.

[0156] The computing device for building a large model data set based on a small number of samples includes:

[0157] The computer-readable storage medium stores one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device that constructs a large model data set based on a small number of samples, enable the computing device that constructs a large model data set based on a small number of samples to perform any of the above methods.

[0158] The computing device for building a large model data set based on a small number of samples includes:

[0159] One or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the above methods.

[0160] Compared with the existing technology, the method and device for constructing a large model data set based on a small number of samples can effectively split data samples such as business instructions through business combing, and adopt the object-behavior-matter method to promote the interaction between business labeling personnel and the large model for auxiliary labeling, thereby completing the combing of the data set; the method and device not only closely fit the business needs, but can also construct data sets for specific business scenarios and quickly generate high-quality corpus that meets the industry's large model training needs. The data set constructed by this method can also improve the accuracy of business processing and ensure the effectiveness of the final generated industry model in processing business, thereby providing strong support for the training and application of large models.

[0161] The above is a detailed introduction to a method and device for building a large model data set based on a small number of samples in an example of the present invention. This section uses specific examples to illustrate the principles and implementation methods of the invention. The above examples are only used to help understand the core ideas of the present invention. Without departing from the principles of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

Claims

1. A method for constructing a large model data set based on a small number of samples, characterized by: The following steps are involved: Step S1: Obtaining the original sample Business personnel first input the original business processing rules or work instructions as samples. The samples include the specific processes and regulations of business operations and are the basis for building data sets. Step S2: Preliminary labeling of samples Business personnel customize the segmentation of original samples by chapter or use the existing unstructured text segmenter to extract the feature information of the original samples and convert them into several independent data blocks to generate annotation features; Business personnel preliminarily mark the split independent data blocks and select key data blocks from them; Step S3: Clarify downstream task instructions Clarify downstream task instructions and customize the instruction format according to the task type; and after clarifying the relevant downstream task instructions, build a data set in alpaca format in a targeted manner; Step S4: sorting out object elements Business personnel sort out the object elements in the business, including the people, time, place and objects involved in the business, and further divide each data block that has been divided according to the dimensions of people, time, place or objects to expand the number of data blocks; Step S5: Augmentation of the object element model Use the big model to expand and augment the sorted object elements, generate operation behaviors in different situations for different objects, and expand the number of data blocks; Step S6: Sorting out behavioral elements After the big model augments the object elements, business personnel collaborate with the big model to conduct a basic sorting of the behavior elements, clarifying the behavior of each object element one by one, including the behavior of different objects in different situations and their impact; by carefully sorting out the behavior in business scenarios, the coverage of the data set is further expanded, and future data samples are generated and labeled based on the behavior; Step S7: Augmentation of the behavioral factor model After completing the basic sorting of behavioral elements, the big model further expands the behavioral elements according to the triggering conditions and execution order of the behaviors in the custom scenarios, generates more detailed behaviors, and further expands the number of data blocks; Step S8: Sorting out the elements of the matter After the behavioral elements are expanded, business personnel collaborate with the big model to sort out the event elements and generate a complete list of operational items by describing in detail the processing steps of each item in the business. The item elements refer to specific operational events or process links in the business, which are further refined to clarify the specific items in the business operation; Step S9: Augmentation of the event element model After manually sorting out the elements of the matter, the big model further augments the elements of the matter according to the specific operational matters and detailed processing steps in the custom scenario; Step S10: Positive and negative example generation The big model further generates positive and negative examples based on the event elements. Positive examples refer to standard operating procedures in the business, while negative examples include scenarios where abnormal processing or incorrect operations are discovered. While generating positive and negative examples, the data is annotated, labels are generated for positive and negative examples based on classification tasks, and the big model is used to generate answers for answering tasks to ensure that the big model can cope with different business scenarios and edge cases.

2. The method for constructing a large model data set based on a small number of samples according to claim 1, characterized in that: The key data blocks are steps, regulations or conditions in an independent business processing flow.

3. The method for constructing a large model data set based on a small number of samples according to claim 1, characterized in that: In step S5, the sorted object elements are expanded and augmented based on the operation behaviors of the object at different time points, based on the processing steps of the object in different business scenarios, and the behavior model of the object in a specific situation.

4. The method for constructing a large model data set based on a small number of samples according to claim 1, characterized in that: In step S7, the large model refines the behavior elements to generate specific operation descriptions under different conditions, and generates behavior objects under different situations based on time and condition changes, thereby realizing behavior element augmentation; When expanding behavioral elements, use the Few Shot Prompt method to improve the prompt words based on the basic behaviors previously sorted out.

5. The method for constructing a large model data set based on a small number of samples according to claim 1, characterized in that: In step S9, the big model generates specific operational items based on the sorted time behaviors, generates detailed processing steps for each item, and ensures that all key links in the business process are covered to achieve item element augmentation.

6. A device for constructing a large model data set based on a small number of samples, characterized in that: include: The original sample input module is responsible for providing a channel for business personnel to input original business processing rules or work instructions as samples. The samples include specific processes and regulations of business operations and are the basis for building data sets. The sample preliminary annotation module is responsible for helping business personnel to segment the original sample by chapter or use the existing unstructured text word segmenter, extract the feature information of the original sample, and convert it into several independent data blocks, perform preliminary annotation on the segmented independent data blocks, and select key data blocks from them by definition; The key data blocks are steps, regulations or conditions in an independent business processing flow; The downstream task instruction clarification module is responsible for clarifying the downstream task instructions and customizing the instruction format according to the task type; and after clarifying the relevant downstream task instructions, it specifically builds the alpaca format data set; The object element sorting module is responsible for sorting out the object elements in the business, including the people, time, place and objects involved in the business, and further dividing each data block that has been divided according to the dimensions of people, time, place or objects to achieve the expansion of the number of data blocks; The object element large model augmentation module is responsible for expanding and augmenting the sorted object elements using the large model, generating operation behaviors in different situations for different objects, and expanding the number of data blocks; The behavior factor combing module is responsible for combing the behavior factors after the big model has augmented the object factors. Through the collaboration between business personnel and the big model, the module clarifies the behavior of each object factor one by one, including the behavior of different objects in different situations and their impact. Through the detailed combing of the behavior in the business scenario, the module further expands the coverage of the data set and generates and annotates future data samples based on the behavior. The behavior factor big model augmentation module is responsible for further augmenting the behavior factors based on the triggering conditions and execution sequence of behaviors in custom scenarios after completing the basic sorting of the behavior factors, generating more detailed behaviors and further expanding the number of data blocks; The item element sorting module is responsible for sorting out the item elements through the collaboration between business personnel and the big model after the behavioral element augmentation is completed, and generating a complete list of operational items by describing the processing steps of each item in the business in detail; The item elements refer to specific operational events or process links in the business, which are further refined to clarify the specific items in the business operation; The item element big model augmentation module is responsible for further augmenting the item elements based on the specific operation items and detailed processing steps in the custom scenario using the big model after the item elements are manually sorted out; The positive and negative example generation module is responsible for using the big model to further generate positive and negative examples based on the event elements. Positive examples refer to the standard operating procedures in the business, while negative examples include scenarios where abnormal processing or erroneous operations are discovered. While generating positive and negative examples, the data is annotated, labels are generated for positive and negative examples based on classification tasks, and the big model is used to generate answers for answering tasks to ensure that the big model can cope with different business scenarios and edge cases.

7. The device for constructing a large model data set based on a small number of samples according to claim 6, characterized in that: The object element large model augmentation module expands and augments the sorted object elements based on the object's operational behavior at different time points, the object's processing steps in different business scenarios, and the object's behavior model in a specific situation.

8. The device for constructing a large model data set based on a small number of samples according to claim 6, characterized in that: The behavior factor large model augmentation module uses the large model to refine the behavior factors, generate specific operation descriptions under different conditions, and generate behavior objects under different situations based on time and condition changes to achieve behavior factor augmentation; When expanding the behavioral elements, use the Few Shot Prompt method to improve the prompt words based on the basic behaviors previously sorted out; The item element big model augmentation module uses the big model to generate specific operation items based on the sorted time behavior, generates detailed processing steps for each item, and ensures that all key links in the business process are covered to achieve item element augmentation.

9. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions, which, when executed by a computing device that constructs a large model dataset based on a small number of samples, cause the computing device that constructs a large model dataset based on a small number of samples to perform any one of the methods according to claims 1 to 5.

10. A computing device for constructing a large model data set based on a small number of samples, characterized in that: include: One or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods according to claims 1 to 5.

Images