A system, method, and program for constructing AI model training datasets through instruction tuning.

Abstract

A system, method, and program for constructing a dataset with improved zero-shot learning performance of an AI model through instruction tuning are disclosed. The above method, The method includes the steps of: extracting instructions from a training task used to train an AI model and a target task which is learned through the training task; comparing the extracted training task instructions with the extracted target task instructions to evaluate their similarity; selecting the training task instructions from which the similarity is greater than or equal to a predetermined value; and outputting the selected training task instructions as a dataset.

Description

Technical Field

[0001] The present invention relates to a system, method, and program for constructing a dataset for AI model learning through instruction tuning, and more particularly, to a system, method, and program for constructing a dataset that improves the zero-shot learning performance of an AI model through instruction tuning.

Background Art

[0002] In recent research, instruction tuning has been established as a major method for improving zero-shot learning performance.

[0003] Zero-shot learning means that when a new task (unseen task) is presented in a situation where an AI model has not learned a specific task, it has a certain level of understanding and execution ability. The main element that enables zero-shot learning is the ability of the model to understand the structure of the data used during learning, abstract it, and generalize it. For example, in the case of a language model, by understanding the instructions and instances included in the tasks used for learning and learning them as abstract concepts, it can interpret and execute a new task even when it is given. In order to improve zero-shot learning ability, it is important to learn various tasks, but it is also important to select and learn training tasks related to each target task. This is because if learning is performed using training tasks that are not related to the target task, negative transfer may occur and the performance of the AI model may decrease.

[0004] A recent method for instruction tuning involves inputting a target task into a model trained through training tasks, evaluating its performance, selecting training tasks that show high performance as having high similarity to the target task, and building them as a dataset for AI learning (the transferability method) (Taskweb: Selecting better 662 source tasks for multi-task NLP, Joongwon Kim et al., 2023). However, this method requires the use of both the instructions (including the task definition) and instances included in the task, and necessitates additional model training for similarity evaluation, resulting in the drawback of requiring excessive computing resources. Another method for instruction tuning has been proposed to construct a dataset by extracting some data from a target task as a sample, measuring the similarity between the extracted target task and the training task, and selecting training tasks with high similarity (Exploring the benefits of training expert language models over instruction tuning, Joel Jang et al., 2023). However, this method has limitations in guaranteeing strict zero-shot performance because it uses the target task extracted as a sample, and it still has limitations in that it can require a lot of computing resources depending on the number of samples, as it still uses both the instructions and instances included in the task. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Taskweb:Selecting better 662 source tasks formulti-task NLP, Joongwon Kim and 3 others, 2023. [Non-Patent Document 2] Exploring the benefits of training expert language models over instruction tuning, Joel Jang et al., 2023. [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] The problem that this invention aims to solve is to provide a system, method, and program for constructing a dataset in which the zero-shot learning performance of an AI model is improved through instruction tuning.

[0007] The problems that this invention aims to solve are not limited to those described above, and other problems not mentioned will be clearly understood by an ordinary person from the following description. [Means for solving the problem]

[0008] The present invention provides a system for constructing a dataset, comprising at least one processor and a memory for storing one or more instructions, wherein the at least one processor performs the following steps by executing the one or more instructions stored in the memory: extracting instructions from a training task used to train an AI model and a target task which is a task to be learned through the training task; comparing the extracted training task instructions with the extracted target task instructions to evaluate their similarity; selecting the extracted training task instructions whose similarity is greater than or equal to a predetermined value; and outputting the selected training task instructions as a dataset.

[0009] In the system described above, the step of evaluating the similarity can be performed by representing the extracted training task instructions and the extracted target task instructions as vectors and comparing the directional similarity of the two vectors using cosine similarity. In the above system, the step of evaluating the similarity is to evaluate the similarity using a model transfer method, the model transfer method trains a model transfer AI model with one of the extracted training task instructions, evaluates the performance of the model transfer AI model by inputting multiple extracted target task instructions into the model transfer AI model, and assigns a high similarity to the extracted target task instructions that the model transfer AI model shows high performance in and to the extracted training task instructions used to train the model transfer AI model.

[0010] In the above system, the step of evaluating the similarity is performed by a task selection model that has undergone pre-tuning for the similarity evaluation, and the pre-tuning for the similarity evaluation may include the steps of selecting one of the multiple target tasks as the first task, extracting the first instruction from the first task and designating it as a positive sample, selecting one of the multiple target tasks other than the first task as the second task, extracting the second instruction from the second task and designating it as a negative sample, assigning similarity scores to the positive and negative samples, and training the task selection model using the positive and negative samples including the similarity scores. In the above system, the step of assigning similarity scores to the positive and negative samples can be such that a similarity score of "1" is assigned to the positive sample and a similarity score of "0" is assigned to the negative sample.

[0011] In the above system, the step of extracting the instructions may further include the step of unifying the placeholders contained in the extracted training task instructions and the extracted target task instructions into specific terminology. A method for constructing a dataset according to another aspect of the present invention is a method for generating predictive data performed by at least one processor, comprising the steps of: extracting instructions from a training task used to train an AI model and a target task which is a task to be learned through the training task; comparing the extracted training task instructions with the extracted target task instructions to evaluate their similarity; selecting the extracted training task instructions whose similarity is greater than or equal to a predetermined value; and outputting the selected training task instructions as a dataset.

[0012] In the above method, the step of evaluating the similarity can be performed by representing the extracted training task instructions and the extracted target task instructions as vectors and comparing the directional similarity of the two vectors using cosine similarity. In the above method, the step of evaluating the similarity is to evaluate the similarity using a model transfer method, the model transfer method trains a model transfer AI model with one of the extracted training task instructions, evaluates the performance of the model transfer AI model by inputting multiple extracted target task instructions into the model transfer AI model, and assigns a high similarity to the extracted target task instructions that the model transfer AI model shows high performance in and to the extracted training task instructions used to train the model transfer AI model.

[0013] In the above method, the step of evaluating the similarity is performed by a task selection model that has undergone pre-tuning for the similarity evaluation, and the pre-tuning for the similarity evaluation may include the steps of selecting one of the multiple target tasks as the first task, extracting the first instruction from the first task and designating it as a positive sample, selecting one of the multiple target tasks other than the first task as the second task, extracting the second instruction from the second task and designating it as a negative sample, assigning similarity scores to the positive and negative samples, and training the task selection model using the positive and negative samples including the similarity scores. In the above method, the step of assigning similarity scores to the positive and negative samples may involve assigning a similarity score of "1" to the positive sample and a similarity score of "0" to the negative sample.

[0014] In the above method, the step of extracting the instructions may further include the step of unifying the placeholders contained in the extracted training task instructions and the extracted target task instructions into specific terminology. A program according to another aspect of the present invention may be a program recorded on a computer-readable recording medium that causes a computer to execute the method for constructing a dataset according to an embodiment of the present invention. [Effects of the Invention]

[0015] According to the present invention, by presenting a simple and effective method for selecting training tasks related to a target task, it is possible to improve the efficiency of computing resource utilization and enhance zero-shot learning capabilities. Further, according to the present invention, by additionally training a task selection model that evaluates similarity, the similarity between the instructions of the target task and the instructions of the training tasks can be evaluated more accurately. Furthermore, by improving the accuracy of training task selection, the overall performance of the model can be improved.

[0016] Also, according to the present invention, by unifying the placeholders included in the extracted instructions into specific terms, it is possible to prevent the placeholders from having a negative impact on the learning of the AI model and improve the zero-shot performance. The effects of the present invention are not limited to the effects described above, and other effects not mentioned can also be clearly understood by those skilled in the art from the following description.

Brief Description of Drawings

[0017] [Figure 1] It is a schematic diagram of a system for implementing a method of constructing a dataset that improves the zero-shot learning performance of an AI model through instruction tuning according to an embodiment of the present invention. [Figure 2] It is a block diagram for explaining the configuration of an apparatus for constructing a dataset that improves the zero-shot learning performance of an AI model through instruction tuning according to an embodiment of the present invention. [Figure 3] It is a block diagram for explaining a method of constructing a dataset that improves the zero-shot learning performance of an AI model through instruction tuning according to an embodiment of the present invention. [Figure 4] It is the result of evaluating the zero-shot learning performance of a dataset constructed according to an embodiment of the present invention. [Figure 5] It is the result of evaluating the zero-shot learning performance of a dataset in which the instructions of training tasks are selected based on those having a similarity ranking of the top n (n is 1, 3, 5, or 10) or higher according to an embodiment of the present invention. [Figure 6]A block diagram for explaining a method of evaluating similarity using a task selector model that has undergone pre-tuning according to an embodiment of the present invention. [Figure 7] Code representing data including placeholders.

Embodiments for Carrying Out the Invention

[0018] The embodiments introduced below are provided as examples so that the idea of the present invention can be sufficiently conveyed to an ordinary technician in the technical field to which the present invention belongs. Therefore, the present invention is not limited to the embodiments described below and may be embodied in other forms. The same reference numerals throughout the present invention refer to the same components. The present invention does not explain all the elements of the embodiments, and the content in the technical field to which the present invention belongs or the overlapping content in the embodiments is omitted. The terms "part, module, member, block" used in the specification can be implemented by software or hardware, and depending on the embodiment, a plurality of "parts, modules, members, blocks" may be implemented as one component, or one "part, module, member, block" may include a plurality of components.

[0019] Throughout the specification, when a part is said to be "connected" to another part, it includes not only the case of being directly connected but also the case of being indirectly connected, and the indirect connection includes the connection by a wireless communication network. Also, when a part is said to "include" a certain component, it means that it does not exclude other components and may further include other components unless otherwise stated to the contrary.

[0020] Throughout the specification, when a member is said to be "on" another member, it includes not only the case where a member is in contact with another member but also the case where another member exists between the two members. Terms such as "first," "second," etc., are used to distinguish one component from another, and the components are not limited by the terms used above.

[0021] A singular expression includes plural forms unless there is a clear exception in the context. The identification codes in each step are used for explanatory purposes only and do not indicate the order of the steps. Each step may be performed in a different order than specified unless the context clearly indicates a specific order.

[0022] The system for constructing a dataset with improved zero-shot learning performance of an AI model through instruction tuning according to the present invention may include devices, and such devices may include all kinds of devices that can perform computational processing and provide results to a user. For example, the system for constructing a dataset with improved zero-shot learning performance of an AI model through instruction tuning according to the present invention may include at least one of a computer, a server device, and a portable terminal, or any one of the same or similar functions. However, the present invention is not limited thereto. Here, the above-mentioned computer may include, for example, a notebook, desktop, laptop, tablet PC, or slate PC equipped with a web browser.

[0023] The above-mentioned server device may include application servers, computing servers, database servers, file servers, game servers, mail servers, proxy servers, and web servers, as servers that communicate with external devices and process information. The above-mentioned portable terminals may include, for example, all types of handheld-based wireless communication devices that guarantee portability and mobility, such as PCS (Personal Communication System), GSM (Global System for Mobile communications), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), WiBro (Wireless Broadband Internet) terminals, and smartphones, as well as wearable devices such as watches, rings, bracelets, anklets, necklaces, glasses, contact lenses, or head-mounted devices (HMDs).

[0024] Embodiments of the present invention will be described in detail below with reference to the attached drawings. The present invention relates to a system, method, and program for constructing a dataset for AI model training through instruction tuning, and more particularly to a system, method, and program for constructing a dataset in which the zero-shot learning performance of an AI model has been improved through instruction tuning.

[0025] Figure 1 is a schematic diagram of a system for constructing an AI model training dataset through instruction tuning according to one embodiment of the present invention. As shown in Figure 1, the system 1000 may include a device 100, a database 200, an AI model 300, a dataset generation module 400, and a task selection model 500.

[0026] The device 100, database 200, AI model 300, dataset generation module 400, and task selection model 500 included in system 1000 can communicate via network W. Here, network W can include wired and wireless networks. For example, the network can include various types of networks such as local area networks (LAN), metropolitan area networks (MAN), and wide area networks (WAN). Furthermore, network W may also include the well-known World Wide Web (WWW). However, network W according to embodiments of the present invention is not limited to the networks listed above, and may also include, at least in part, a well-known wireless data network or a well-known telephone network, or a well-known wired or wireless television network.

[0027] The device 100 can input the dataset generated by the dataset generation module 400 into the AI ​​model 300 or the task selection model 500, and train the AI ​​model 300 or the task selection model 500. The dataset generation module 400 can generate a dataset 410 using data output through instruction tuning. The dataset 410 is identical to, and more preferably, a dataset with improved zero-shot learning performance according to the present invention, which is commonly used to train an AI model or to evaluate its performance. The dataset generated through the dataset generation module 400 consists of instructions extracted from a training task that show a similarity of a predetermined value or higher to the instructions of the target task in the similarity evaluation. Here, an instruction includes instructions, information, or a definition of the task necessary to perform the task (e.g., "Perform a translation"), and an instance includes input data or examples used for model training (e.g., a specific document to be translated).

[0028] The database 200 can store various data (e.g., datasets) for training the AI ​​model 300 or for evaluating its performance. The database 200 may also store training tasks including instructions and instances, target tasks including instructions and instances, and, in various embodiments, calculated datasets output by the AI ​​model 300. However, the system 1000 does not need to include the database 200 once the training of the AI ​​model 300 is complete. Figure 1 shows a case where the database 200 is implemented outside of the device 100. In this case, the database 200 may be connected to the device 100 by wired or wireless connection. However, this is only one embodiment, and the database 200 may also be implemented as a component of the device 100.

[0029] Figure 1 shows the case where the AI ​​model 300 is implemented outside the device 100 (for example, implemented in a cloud-based manner), but is not limited to this, and may be implemented as a component of the device 100. Figure 2 is a block diagram illustrating the configuration of an apparatus for constructing a dataset for AI model training through instruction tuning, according to one embodiment of the present invention.

[0030] As shown in Figure 2, the device 100 may include a memory 110, a communication module 120, a display 130, an input module 140, and a processor 150. However, it is not limited to these, and the software and hardware configuration of the device 100 can be modified, added to, or omitted to the extent that is obvious to the average technician, depending on the required operation. The device 100 may also be replaced by a system, and the device 100 may include multiple devices, in which case each component included in the device 100 may be included in at least one of the multiple devices. Memory 110 can store data supporting various functions of the device 100, programs for the operation of the processor 150, input / output data, and multiple application programs or applications driven by the device, data for the operation of the device 100, instructions, and AI models. At least some of these application programs may be downloaded from an external server via wireless communication.

[0031] Such memory 110 may include at least one type of storage medium from among flash memory type, hard disk type, SSD type (Solid State Disk type), SDD type (Silicon Disk Drive type), multimedia card micro type, card type memory (e.g., SD or XD memory), RAM (random access memory), SRAM (static random access memory), ROM (read-only memory), EEPROM (electrically erasable programmable read-only memory), PROM (programmable read-only memory), magnetic memory, magnetic disk, and optical disk. Furthermore, the memory 110 is detachable from the device and may include a database connected via wired or wireless connection. The database 200 shown in Figure 1 may be implemented as a component of the memory 110.

[0032] The communication module 120 may include one or more components that enable communication with external devices, for example, at least one of a broadcast receiving module, a wired communication module, a wireless communication module, a short-range communication module, and a location information module. Wired communication modules can include not only various wired communication modules such as LAN (Local Area Network) modules, WAN (Wide Area Network) modules, or VAN (Value Added Network) modules, but also various cable communication modules such as USB (Universal Serial Bus), HDMI (High Definition Multimedia Interface), DVI (Digital Visual Interface), RS-232 (recommended standard 232), power line communication, or POTS (plain old telephone service).

[0033] Wireless communication modules can include not only WiFi (Wireless Fidelity) modules and WiBro (Wireless broadband) modules, but also modules that support various wireless communication methods such as GSM (Global System for Mobile Communication), CDMA (Code Division Multiple Access), WCDMA (Wideband Code Division Multiple Access), UMTS (Universal Mobile Telecommunications System), TDMA (Time Division Multiple Access), LTE (Long Term Evolution), 4G, 5G, and 6G. The display 130 displays (outputs) information or data processed by the device 100, data input or output through the AI ​​model 300, etc. The display 130 can also display execution screen information of an application program (e.g., an application) driven by the device 100, or UI (User Interface) or GUI (Graphic User Interface) information corresponding to such execution screen information.

[0034] The input module 140 is for receiving information from the user, and when the user inputs information through the input unit, the processor 150 can control the operation of the device 100 in accordance with the input information. Such an input module 140 may include hardware physical keys (for example, buttons, dome switches, jog wheels, jog switches, etc., located on at least one of the front, rear, and side surfaces of the device) and software touch keys. For example, the touch keys may consist of virtual keys, soft keys, or visual keys displayed on the touchscreen-type display 130 through software processing, or touch keys located in a part other than the touchscreen. On the other hand, the virtual keys or visual keys can be displayed on the touchscreen in various forms, and may consist of graphics, text, icons, videos, or combinations thereof.

[0035] The processor 150 may be embodied as a memory that stores data relating to an algorithm or a program that reproduces an algorithm for controlling the operation of the components within the device 100 (including learning or executing an AI model), and at least one processor (not shown) that performs the above operation using the data stored in the memory. In this case, the memory and the processor may be embodied as separate chips, or as a single chip. In one embodiment, the system 1000 or apparatus 100 according to the present invention may include at least one processor, and if it includes multiple processors, the multiple processors may be contained within different apparatuses 100.

[0036] Furthermore, the processor 150 can control any one or more of the above-mentioned components in combination in order to implement various embodiments of the present invention described below on the device 100. Figure 3 is a block diagram illustrating a method for constructing a dataset in which the zero-shot learning performance of an AI model has been improved through instruction tuning, according to an embodiment of the present invention.

[0037] Referring to Figure 3, instructions are extracted from the training task, which is data used to train the AI ​​model, and the target task, which is information about the tasks that the AI ​​model learns through the training task (S311 and S312). The training task and target task may, but are not limited to, be stored in the memory 110 of the device 100. The training task and target task each include instructions and instances. Here, the instructions include the necessary instructions, information, and task definitions for performing the task, and the instances include input data or examples used for model training. For example, in the case of a task called "Sentiment Analysis," the instructions might be "Determine whether the emotion expressed in the given text is positive, negative, or neutral," and the instance might be "This movie was really fun! (positive emotional expression)." Also, in the case of a task called "Text Classification," the instructions might be "Classify the news article into categories such as politics, sports, entertainment, or technology," and the instance might be "sports article." Furthermore, in the case of a "Question Answering (QA)" task, the instruction is "Provide an informed and correct answer to a given text and a question related to the text," and an instance could be "Who is the author of the Harry Potter series? (Question) It is J.K. Rowling (Answer)." Also, in the case of a "Text Summarization" task, the instruction is "Generate a concise and consistent summary of a long text document while maintaining the main information and subject matter," and an instance could be "An article about today's weather." Also, in the case of a "Named Entity Linking (NEL)" task, the instruction is "Link an entity recognized in the text to the corresponding entry in a knowledge base or database to provide additional context and information," and an instance could be "Jeju Island is an island in South Korea."It could be (Entity: Jeju Island, Link: Island in South Korea). Next, the instructions for the extracted training task are compared with the instructions for the extracted target task to evaluate their similarity (S320).

[0038] One method for evaluating similarity is cosine similarity. Cosine similarity is a method for measuring the similarity between two vectors in a vector space, and it calculates similarity by measuring how similar the directions of the two vectors are. Cosine similarity uses how "similar" the directions of the two vectors are. More specifically, cosine similarity is defined as follows:

[0039]

number

[0040] The cosine similarity value ranges from -1 to 1. A value closer to 1 indicates that the directions of the two vectors are similar, while a value closer to -1 indicates that the directions of the two vectors are opposite. For example, a value of 0 may indicate that the two vectors are orthogonal (at a 90-degree angle) and therefore have no directional similarity. In this invention, the extracted training task and target task instructions are each represented as vectors, and the similarity can be calculated using these two vectors.

[0041] Another method for evaluating similarity is model transfer. This method involves training a model with instructions for an extracted training task, then inputting instructions for an extracted target task into the trained model and evaluating its performance. Specifically, the higher the performance the trained model shows when the extracted target task instructions are input, the higher the similarity between the extracted target task instructions and the extracted training task instructions. In other words, the extracted target task instructions and the extracted training task instructions are assigned higher similarity scores in order of their performance when the trained model is input. Here, metrics such as accuracy and loss function can be used to measure the performance of the model.

[0042] Accuracy represents the proportion of samples that the model correctly classified, and can be calculated as follows:

[0043]

number

[0044] For example, when instructions for a target task extracted from a trained model are input, the accuracy can be calculated by dividing the number of samples for which accurate results are output by the total number of extracted instructions for the target task, and this can be used as the similarity score.

[0045] A loss function is a function that measures the difference between the model's predicted values ​​and the actual values. More specifically, it can use the Mean Squared Error (MSE), Cross-Entropy Loss, etc. For example, in regression problems, the Mean Squared Error is used to measure the error between the values ​​predicted by the model and the actual values, and in classification problems, the Cross-Entropy Loss is used to measure the difference between the probability distribution predicted by the model and the actual distribution. Next, the extracted training task instructions are selected based on their similarity to a predetermined value (S330). Here, the "predetermined value" may be a similarity rank, a specific similarity score, or the like. For example, it may be "the top 5 or higher training task instructions in terms of similarity" (similarity rank), or it may be "training task instructions with a similarity score of 10 or higher" (specific similarity score).

[0046] Next, the training task instructions selected through S330 are output as a dataset (S340). This dataset may be used to train an AI model and may be stored in the database 200 of system 1000 for that purpose. Because instructions contain information that defines the characteristics of a task, a dataset containing only the instructions for selected training tasks (i.e., a dataset without instances) is sufficient to train an AI model. In fact, training with a dataset that includes both instructions and instances can cause negative transfers during training due to the diverse content included in the tasks, potentially degrading zero-shot performance. Therefore, it is necessary to select related tasks from among the tasks.

[0047] This invention can improve the efficiency of computing resource utilization and enhance zero-shot learning capabilities by providing a simple and effective method for selecting training tasks related to a target task. Figure 4 shows the results of evaluating the zero-shot learning performance when training with a dataset containing both "instructions" and "instances" (T5(3b)+DS-BTS) and when training with a dataset containing only "instructions" of training tasks selected according to the present invention (T5(3b)+I-BTS). The dataset used as the task was P3 (Public Pool of Prompts) (Multitask prompted training enables zero-shot task generation, Victor Sanh et al., 2022), which contains a total of 35 tasks across 8 task clusters for training, with each task containing an average of 11.7 instructions. Furthermore, the evaluation results NLI (Natural Language Inference), Sentence Completion, Coreference Resolution, and WSD (Word Sense Disambiguation) represent major categories of the target tasks, while RTE (Recognizing Textual Entailment), CB (Commonsense Based), and COPA (Choice of Plausible Alternatives) represent subcategories of the target tasks.

[0048] T0(3b) represents a model trained on all data without any selection process based on similarity; T5(3b)+DS-BTS represents a model trained using both the "instructions" and "instances" included in the selected tasks, similar to the prior art, by selecting the top 5 tasks with high similarity calculated from a sample of data; and T5(3b)+I-BTS represents a model trained using only the "instructions" included in the selected tasks, as in the present invention. Looking at the scores that evaluated zero-shot learning performance, the T5(3b)+I-BTS according to the present invention showed higher performance scores than the T5(3b)+DS-BTS in the majority of tasks, and the average performance score of the T5(3b)+I-BTS according to the present invention (54.92 points) was 4.3% higher than that of the T5(3b)+DS-BTS (50.62 points).

[0049] These results indicate that instances within a task can cause negative transfers during training, potentially degrading model performance. Therefore, this invention demonstrates that training an AI model solely with training task instructions, as described in this invention, helps improve zero-shot performance. Figure 5 shows the results of evaluating the zero-shot learning performance when selecting instructions for extracted training tasks that have a similarity score of at least a predetermined value (S330), using the top n similarity scores (n is 1, 3, 5, or 10) as the criterion. As can be seen in Figure 5, the zero-shot learning performance improves stepwise from selecting the top 1 task to selecting the top 5 tasks. However, for some tasks, a tendency for zero-shot learning performance to decrease when selecting the top 10 tasks is observed. Therefore, when selecting instructions based on the top n similarity scores (n is 1, 3, 5, or 10), it is preferable to use the top 5 or higher as the criterion, but it is not limited to this.

[0050] Figure 6 is a block diagram illustrating another example of a method for evaluating similarity (step S320 in Figure 3), which uses a pre-tuned task selector model to evaluate similarity. The task selector model may be a separate AI model that has been pre-trained for more accurate similarity evaluation. Referring to Figure 6, first, one of the multiple target tasks is selected as the first task (S411), and the first instruction is extracted from the first task and designated as the positive sample (S421). Then, one of the multiple target tasks other than the first task is selected as the second task (S412), and the second instruction is extracted from the second task and designated as the negative sample (S422).

[0051] Next, positive samples are assigned a high similarity score (S431), and negative samples are assigned a low similarity score (S432). Positive samples are extracted from the first task, so they are considered to be relatively similar to the first task, while negative samples are extracted from tasks other than the first task, so they are considered to be relatively similar to the first task. As a result, positive samples with high similarity are assigned a high similarity score, and negative samples with low similarity are assigned a low similarity score. Similarity scores may be assigned differently by the user. For example, in one embodiment, a similarity score of "1" could be assigned to positive samples and a similarity score of "0" to negative samples. Thus, the similarity scores for positive and negative samples can be set and assigned by the user, but are not limited to this. For example, a method could be used that vectorizes the positive and negative samples and calculates the Euclidean distance between these two vectors, including the similarity evaluation method described above. Alternatively, a natural language processing-based approach could be used that measures the similarity between sentences and assigns scores using techniques such as sentence embedding, tokenization, and morphological analysis. Furthermore, the user can further tune the calculated similarity scores so that positive samples are assigned high similarity scores and negative samples are assigned low similarity scores.

[0052] Next, the positive and negative samples, each assigned a similarity score, are output as tuning data for the task selection model (S440), and the task selection model is trained using this tuning data (S450). Since the positive and negative samples are assigned similarity scores that clearly distinguish them based on the similarity of the first and second instructions, training the task selection model using these scores allows the task selection model to be tuned to more clearly distinguish between the similarity of the instruction for the target task and the instruction for the training task. As a result, the present invention can more accurately evaluate the similarity between the instructions for the target task and the instructions for the training task by further training the task selection model that evaluates similarity, thereby improving the accuracy of training task selection and enhancing the overall performance of the model.

[0053] On the other hand, the step of extracting instructions (S311 and S312 in Figure 3) may further include a step of unifying the placeholders included in the extracted instructions into the same sentence. A placeholder is a code that is temporarily inserted where elements such as characters and images in the data would normally go. For example, as shown in Figure 7, the data "question" included in the code ("Does the word {{word1}} and {{word2}} have the same meaning in these two sentences? Yes, No?\n{{sentence1}}\n{{sentence2}}") receives "word1", "word2", "sentence1", and "sentence2" entered by the user and outputs the completed data. Here, {{word1}}, {{word2}}, {{sentence1}}, and {{sentence2}} are placeholders that are temporarily inserted where user input values ​​would normally go.

[0054] Because placeholders are not only inconsistent in form but can also have unclear meanings, data containing placeholders can negatively impact the training process of AI models. Therefore, unifying the placeholders in the data used to train AI models with specific terminology can eliminate such negative impacts. For example, placeholders can be identified and removed through data preprocessing. In Figure 7, {{word1}} and {{word2}} in the data can be unified into {{text}}, and {{sentence1}} and {{sentence2}} can be unified into {{candidate}}. Methods for identifying and removing placeholders include, but are not limited to, methods where the user directly identifies and removes them, methods where the text is tokenized and the divided placeholders are identified and removed, and methods where placeholder patterns are recognized and identified and removed.

[0055] This allows the present invention to improve zero-shot performance by unifying the placeholders included in the extracted instructions into specific terminology, thereby preventing placeholders from negatively impacting the AI ​​model's learning. Instruction tuning as described in this invention means selecting training tasks relevant to the target task, thereby enhancing the learning of the AI ​​model and enabling it to efficiently perform complex tasks. More specifically, instruction tuning involves adjusting parameters and directives to improve the behavior of the AI ​​algorithm. Instruction tuning has the advantage of being a simple way to increase the efficiency of AI learning and, in particular, being able to generalize to unseen tasks.

[0056] The training or target tasks used in this invention, and the instructions extracted from them, can be configured to be suitable for performing in-context learning (ICL) by classifying those that perform the same task, so that deep learning can proceed in a specific domain. In-context learning refers to a method of learning that extracts only a portion of the data included in a downstream task and uses that context. It is classified into zero-shot learning, one-shot learning, and few-shot learning. As the number of parameters in a language model increases, large-scale language models (LLMs) can learn through the context included in examples, allowing a few examples to be input into a pre-trained LLM to predict results. More specifically, in-context refers to the flow within a prompt, which could be related to arithmetic operations, typographical error correction, language translation, etc. For example, in a task of summarizing dialogue content, in addition to the few-shot learning method that uses prompt examples, zero-shot learning methods that can be controlled through instructional text without prompt examples, and learning methods based on fine-tuning can also be utilized. This allows large-scale language models (e.g., GPT-3, HyperClova, etc.) to solve natural language processing (NLP) problems by understanding context or patterns contained in small amounts of data through prompts, acting as excellent zero-shot or few-shot learners controllable via prompts. Such learning methods have the advantage of not requiring parameter updates and reducing computational loss.

[0057] On the other hand, the method for constructing a dataset for AI model training through instruction tuning according to an embodiment of the present invention may be embodied by the system described with reference to Figure 1. An AI model according to an embodiment of the present invention can be controlled, executed, learned, and driven by a processor, thereby enabling the AI ​​model to perform at least one of the tasks of execution, learning, and driving by at least one processor. Furthermore, the AI ​​model can be stored in memory, and the feature data according to the present invention can also be stored in memory.

[0058] On the other hand, the disclosed embodiments may be embodied in the form of a recording medium for storing computer-executable instructions. The instructions may be stored in the form of program code, which, when executed by a processor, can generate program modules to perform the operations of the disclosed embodiments. The recording medium may be embodied as a computer-readable recording medium. Computer-readable recording media include all types of recording media that store instructions that can be deciphered by a computer. Examples include ROM (Read Only Memory), RAM (Random Access Memory), magnetic tape, magnetic disks, flash memory, and optical data storage devices.

[0059] Embodiments disclosed with reference to the accompanying drawings have been described above. Those with ordinary skill in the art to which the present invention pertains will understand that the present invention may be carried out in forms different from those disclosed without altering the technical spirit or essential features of the invention. The disclosed embodiments are illustrative and should not be construed as restrictive.

Claims

1. A system comprising at least one processor and one or more memory for storing instructions, The at least one processor executes the one or more instructions stored in the memory, The steps include extracting instructions from a training task used to train an AI model and a target task which is learned through the training task, A step of comparing the instructions for the extracted training task with the instructions for the extracted target task to evaluate their similarity, The steps include selecting from the extracted training task instructions those whose similarity is equal to or greater than a predetermined value, The steps include outputting the instructions for the selected training tasks as a dataset, and system.

2. The step of evaluating the similarity is, The extracted training task instructions and the extracted target task instructions are represented as vectors, and the cosine similarity of the two vectors is used for evaluation. The system according to claim 1.

3. The step of evaluating the similarity is to evaluate the similarity using a model transfer method. The aforementioned model transition method is Train the model transfer AI model using one of the instructions for the extracted training tasks. The performance of the model transfer AI model is evaluated by inputting the instructions for multiple extracted target tasks into the model transfer AI model. The model transfer AI model is assigned a high degree of similarity to the extracted target task instructions that demonstrate high performance and to the extracted training task instructions used to train the model transfer AI model. The system according to claim 1.

4. The step of evaluating the similarity is performed by a task selection model that has undergone pre-tuning for the evaluation of the similarity. The pre-tuning for the evaluation of similarity is, A step of selecting one of the multiple target tasks as the first task, The steps include extracting the first instruction from the first task and designating it as a positive sample, A step of selecting one of the multiple target tasks other than the first task as the second task, The steps include extracting the second instruction from the second task and designating it as a negative sample, The steps include assigning similarity scores to the positive and negative samples, The process includes the step of training the task selection model using the positive and negative samples, which include similarity scores. The system according to claim 1.

5. The step of assigning similarity scores to the positive sample and the negative sample is: The positive sample is assigned a similarity score of "1", and the negative sample is assigned a similarity score of "0". The system according to claim 4.

6. The step of extracting the aforementioned instructions is: The process further includes the step of unifying the placeholders contained in the instructions for the extracted training task and the instructions for the extracted target task into specific terminology. The system according to claim 1.

7. A method for generating predictive data performed by at least one processor, The steps include extracting instructions from a training task used to train an AI model and a target task which is learned through the training task, A step of comparing the instructions for the extracted training task with the instructions for the extracted target task to evaluate their similarity, The steps include selecting from the extracted training task instructions those whose similarity is equal to or greater than a predetermined value, The step includes outputting the instructions for the selected training tasks as a dataset, method.

8. The step of evaluating the similarity is, The extracted training task instructions and the extracted target task instructions are represented as vectors, and the cosine similarity of the two vectors is used for evaluation. The method according to claim 7.

9. The step of evaluating the similarity is to evaluate the similarity using a model transfer method. The aforementioned model transition method is Train the model transfer AI model using one of the instructions for the extracted training tasks. The performance of the model transfer AI model is evaluated by inputting the instructions for multiple extracted target tasks into the model transfer AI model. The extracted target task instructions, which demonstrate high performance for the model transfer AI model, are assigned a high degree of similarity to the extracted training task instructions used to train the model transfer AI model. The method according to claim 7.

10. The step of evaluating the similarity is performed by a task selection model that has undergone pre-tuning for the evaluation of the similarity. The pre-tuning for the evaluation of similarity is, A step of selecting one of the multiple target tasks as the first task, The steps include extracting the first instruction from the first task and designating it as a positive sample, A step of selecting one of the multiple target tasks other than the first task as the second task, The steps include extracting the second instruction from the second task and designating it as a negative sample, The steps include assigning similarity scores to the positive and negative samples, The process includes the step of training the task selection model using the positive and negative samples, which include similarity scores. The method according to claim 7.

11. The step of assigning similarity scores to the positive sample and the negative sample is: The positive sample is assigned a similarity score of "1", and the negative sample is assigned a similarity score of "0". The method according to claim 10.

12. The step of extracting the aforementioned instructions is: The process further includes the step of unifying the placeholders contained in the instructions for the extracted training task and the instructions for the extracted target task into specific terminology. The method according to claim 7.

13. A computer-readable recording medium is used to cause a computer to perform the method according to any one of claims 7 to 12, program.

Images