System, method, and program for constructing data set for ai model training through instruction tuning

Abstract

Disclosed are a system, method, and program for constructing a data set for improving zero-shot learning performance of an AI model through instruction tuning. The method comprises the steps of: extracting each instruction 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; evaluating similarity by comparing the extracted instructions of the training task with the extracted instructions of the target task; selecting instructions having a similarity not less than a predetermined value from among the extracted instructions of the training task; and outputting the selected instructions of the training task as a data set.

Description

Systems, methods, and programs for building data sets for AI model training through instruction tuning

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

[0002] Recent research has established instruction tuning as a key method for improving zero-shot learning performance.

[0003] Zero-shot learning refers to an AI model's ability to understand and perform a certain level of new tasks (unseen tasks) when presented with them, even though it has never learned them before. A key element that enables zero-shot learning is the model's ability to understand the structure of the data used during training, abstract it, and generalize it. For example, a language model can understand the instructions or instances contained in the training task and learn them as abstract concepts, enabling it to interpret and perform new tasks even when presented with them.

[0004] To improve zero-shot learning, it's important to train on a variety of tasks, but it's also crucial to select and train training tasks that are relevant to each target task. Training on training tasks unrelated to the target task can result in negative transfer, degrading the AI ​​model's performance.

[0005] Recently, a method of instruction tuning has been proposed, which involves inputting a target task into a model learned through a training task to evaluate its performance, and selecting training tasks with high performance based on their similarity to the target task and constructing a data set for AI learning (model transferability method) (Taskweb: Selecting better 662 source tasks for multi-task NLP, Joongwon Kim et al., 2023). However, this method has the disadvantage of requiring excessive computing resources because it requires all instructions (including task definitions) and instances included in the task and additional model learning for similarity evaluation.

[0006] Another instruction tuning method proposed was to construct a dataset by sampling a portion of the target task data, 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 sampled target tasks, and it still requires significant computing resources depending on the number of samples because it utilizes all instructions and instances included in the task.

[0007] (Non-patent document 1) Taskweb: Selecting better 662 source tasks for multi-task NLP, Joongwon Kim et al., 2023.

[0008] (Non-patent literature 2) Exploring the benefits of training expert language models over instruction tuning, Joel Jang et al., 2023.

[0009] The problem to be solved by the present invention is to provide a system, method, and program for constructing a data set that improves the zero-shot learning performance of an AI model through instruction tuning.

[0010] The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the description below.

[0011] A system for constructing a data set of the present invention comprises a system including at least one processor and one or more instructions and a memory, wherein the at least one processor executes the one or more instructions stored in the memory, thereby extracting instructions from a training task used to train an AI model and a target task which is a task to be trained through the training task, a step of evaluating similarity by comparing the instructions of the extracted training task with the instructions of the extracted target task, a step of selecting instructions of the extracted training task whose similarity is greater than or equal to a predetermined value, and a step of outputting the instructions of the selected training task as a data set.

[0012] In the above system, the step of evaluating the similarity may be performed by expressing the instructions of the extracted training task and the instructions of the extracted target task as vectors and comparing the directional similarity of the two vectors through cosine similarity.

[0013] In the above system, the step of evaluating the similarity may be to evaluate the similarity through a model transfer method, wherein the model transfer method may be to train a model transfer AI model with one of the instructions of the extracted training task, input a plurality of instructions of the extracted target tasks into the model transfer AI model to evaluate the performance of the model transfer AI model, and give a high similarity to the instructions of the extracted target task for which the model transfer AI model shows high performance and the instructions of the extracted training task used for learning the model transfer AI model.

[0014] In the above system, the step of evaluating the similarity may be 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 any one of the plurality of target tasks as a first task; extracting a first instruction from the first task and designating it as a positive sample; selecting any one of the plurality of target tasks other than the first task as a second task; extracting a second instruction from the second task and designating it as a negative sample; assigning a similarity score to the positive sample and the negative sample; and learning the task selection model using the positive sample and the negative sample including the similarity score.

[0015] In the above system, the step of assigning a similarity score to the positive sample and the negative sample may assign a similarity score of “1” to the positive sample and a similarity score of “0” to the negative sample.

[0016] In the above system, the step of extracting the instructions may additionally include a step of unifying placeholders included in the instructions of the extracted training task and the instructions of the extracted target task into specific terms.

[0017] A method for constructing a data set according to another aspect of the present invention is a method for generating prediction data performed by at least one processor, comprising: a step 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; a step of evaluating similarity by comparing the instructions of the extracted training task with the instructions of the extracted target task; a step of selecting instructions of the extracted training task whose similarity is greater than or equal to a predetermined value; and a step of outputting the instructions of the selected training task as a data set.

[0018] In the above method, the step of evaluating the similarity may be performed by expressing the instructions of the extracted training task and the instructions of the extracted target task as vectors and comparing the directional similarity of the two vectors through cosine similarity.

[0019] In the above method, the step of evaluating the similarity may be to evaluate the similarity through a model transfer method, wherein the model transfer method may be to train a model transfer AI model with one of the instructions of the extracted training task, input a plurality of instructions of the extracted target tasks into the model transfer AI model to evaluate the performance of the model transfer AI model, and give a high similarity to the instructions of the extracted target task for which the model transfer AI model shows high performance and the instructions of the extracted training task used for learning the model transfer AI model.

[0020] In the above method, the step of evaluating the similarity may be 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 any one of the plurality of target tasks as a first task; extracting a first instruction from the first task and designating it as a positive sample; selecting any one of the plurality of target tasks other than the first task as a second task; extracting a second instruction from the second task and designating it as a negative sample; assigning a similarity score to the positive sample and the negative sample; and learning the task selection model using the positive sample and the negative sample including the similarity score.

[0021] In the above method, the step of assigning a similarity score to the positive sample and the negative sample may assign a similarity score of “1” to the positive sample and a similarity score of “0” to the negative sample.

[0022] In the above method, the step of extracting the instructions may additionally include a step of unifying placeholders included in the instructions of the extracted training task and the instructions of the extracted target task into specific terms.

[0023] A program according to another aspect of the present invention may be a program stored in a computer-readable recording medium that causes a computer to execute a method for constructing a data set according to embodiments of the present invention.

[0024] According to the present invention, by providing a simple and effective method for selecting a training task related to a target task, the efficiency of computing resource utilization can be increased and zero-shot learning capability can be improved.

[0025] In addition, 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 task can be evaluated more accurately, and further, the accuracy of training task selection can be increased, thereby improving the overall performance of the model.

[0026] In addition, according to the present invention, by unifying placeholders included in extracted instructions into specific terms, zero-shot performance can be improved by preventing placeholders from having a negative impact on AI model learning.

[0027] The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

[0028] FIG. 1 is a schematic diagram of a system for implementing a method for constructing a data set that improves zero-shot learning performance of an AI model through instruction tuning, according to one embodiment of the present disclosure.

[0029] FIG. 2 is a block diagram illustrating a configuration of a device for constructing a data set that improves zero-shot learning performance of an AI model through instruction tuning according to one embodiment of the present disclosure.

[0030] FIG. 3 is a block diagram illustrating a method for constructing a data set that improves the zero-shot learning performance of an AI model through instruction tuning according to embodiments of the present invention.

[0031] Figure 4 shows the results of evaluating the zero-shot learning performance of a data set constructed according to embodiments of the present invention.

[0032] FIG. 5 is a result of evaluating the zero-shot learning performance of a data set in which instructions of a training task are selected based on similarities that are at least the top n (where n is 1, 3, 5, or 10) according to embodiments of the present invention.

[0033] FIG. 6 is a block diagram illustrating a method for evaluating similarity using a pre-tuned task selector model according to embodiments of the present invention.

[0034] Figure 7 is a code that expresses data including placeholders.

[0035] The following examples are provided as examples to ensure that those skilled in the art can fully grasp the spirit of the present invention. Therefore, the present invention is not limited to the embodiments described below and may be embodied in other forms.

[0036] Throughout the present invention, the same reference numerals refer to the same components. The present invention does not describe all elements of the embodiments, and any content that is general in the technical field to which the present invention pertains or that overlaps between the embodiments is omitted. The terms 'part, module, element, block' used in the specification may be implemented in software or hardware, and depending on the embodiments, multiple 'parts, modules, elements, blocks' may be implemented as a single component, or a single 'part, module, element, block' may include multiple components.

[0037] Throughout the specification, when a part is said to be "connected" to another part, this includes not only direct connection but also indirect connection, and indirect connection includes connection via a wireless communication network.

[0038] Additionally, when a part is said to "include" a component, this does not mean that it excludes other components, but rather that it may include other components, unless otherwise specifically stated.

[0039] Throughout the specification, when we say that an element is "on" another element, this includes not only cases where the element is in contact with the other element, but also cases where another element exists between the two elements.

[0040] The terms first, second, etc. are used to distinguish one component from another, and the components are not limited by the aforementioned terms.

[0041] Singular expressions include plural expressions unless the context clearly indicates otherwise.

[0042] The identification codes for each step are used for convenience of explanation and do not describe the order of each step. Each step may be performed in a different order than specified unless the context clearly indicates a specific order.

[0043] A system for constructing a data set that improves the zero-shot learning performance of an AI model through instruction tuning according to the present invention may include a device, and the device may include various devices capable of performing computational processing and providing results to a user. For example, a system for constructing a data set that improves the 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 may be any form having the same or similar functions as these. However, the present invention is not limited thereto.

[0044] Here, the computer may include, for example, a notebook, desktop, laptop, tablet PC, slate PC, etc. equipped with a web browser.

[0045] The above server device is a server that processes information by communicating with an external device, and may include an application server, a computing server, a database server, a file server, a game server, a mail server, a proxy server, and a web server.

[0046] The above portable terminal may include, for example, a wireless communication device that ensures portability and mobility, and may include all kinds of handheld-based wireless communication devices such as a 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) terminal, a smart phone, and a wearable device such as a watch, a ring, a bracelet, an anklet, a necklace, glasses, contact lenses, or a head-mounted device (HMD).

[0047] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0048] The present invention relates to a system, method, and program for constructing a data set for AI model training through instruction tuning, and more particularly, to a system, method, and program for constructing a data set that improves the zero-shot learning performance of an AI model through instruction tuning.

[0049] FIG. 1 is a schematic diagram of a system for constructing a data set for AI model learning through instruction tuning according to one embodiment of the present invention.

[0050] As illustrated in FIG. 1, the system (1000) may include a device (100), a database (200), an AI model (300), a data set generation module (400), and a task selection model (500).

[0051] The device (100), database (200), AI model (300), data set generation module (400), and task selection model (500) included in the system (1000) can communicate via a network (W). Here, the network (W) can include a wired network and a wireless network. For example, the network can include various networks such as a local area network (LAN), a metropolitan area network (MAN), and a wide area network (WAN).

[0052] Additionally, the network (W) may include the well-known World Wide Web (WWW). However, the network (W) according to an embodiment of the present invention is not limited to the networks listed above, and may include at least part of a well-known wireless data network, a well-known telephone network, or a well-known wired / wireless television network.

[0053] The device (100) can input a data set generated by the data set generation module (400) into the AI ​​model (300) or task selection model (500) to perform learning of the AI ​​model (300) or task selection model (500).

[0054] The data set generation module (400) can generate a data set (410) using data output through instruction tuning. The data set (410) is the same as that generally used to train an AI model or evaluate its performance, and more preferably, it may be a data set with improved zero-shot learning performance according to the present invention, but is not limited thereto. The data set generated through the data set generation module (400) is composed of instructions extracted from a training task, which show a similarity higher than a predetermined value in a similarity evaluation with the instructions of the target task. Here, the instruction includes instructions, information, or a definition of the task (e.g., "perform translation") required to perform the task, and the instance includes input data or examples (e.g., a specific document to be translated) used for model training.

[0055] The database (200) may store various data (e.g., data sets) for training or evaluating the performance of the AI ​​model (300). Furthermore, the database (200) may store training tasks including instructions and instances, target tasks including instructions and instances, and, in various embodiments, output data sets output by the AI ​​model (300). However, the system (1000) may not include the database (200) if training of the AI ​​model (300) is complete.

[0056] FIG. 1 illustrates a case where a database (200) is implemented outside of a device (100). In this case, the database (200) may be connected to the device (100) via wired or wireless means. However, this is merely an example, and the database (200) may also be implemented as a component of the device (100).

[0057] FIG. 1 illustrates a case where the AI ​​model (300) is implemented outside the device (100) (e.g., cloud-based), but is not limited thereto, and may be implemented as a component in the device (100).

[0058] FIG. 2 is a block diagram illustrating the configuration of a device for constructing a data set for AI model learning through instruction tuning according to one embodiment of the present invention.

[0059] As illustrated in FIG. 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, the present invention is not limited thereto, and the device (100) may have its software and hardware configurations modified / added / omitted within a range apparent from a perspective of ordinary skill in the art, depending on the required operation. In addition, the device (100) may be replaced with a system, and the device (100) may include a plurality of devices, in which case each component included in the device (100) may be included in at least one of the plurality of devices.

[0060] The memory (110) can store data supporting various functions of the device (100), programs for the operation of the processor (150), input / output data, and a plurality of application programs or applications run on the device, data for the operation of the device (100), commands, and AI models. At least some of these application programs can be downloaded from an external server via wireless communication.

[0061] The memory (110) may include at least one type of storage medium among a flash memory type, a hard disk type, an SSD (Solid State Disk type), an SDD (Silicon Disk Drive type), a multimedia card micro type, a card type memory (e.g., SD or XD memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, and an optical disk.

[0062] Additionally, the memory (110) may be separate from the device and may include a database connected wired or wirelessly. The database (200) illustrated in FIG. 1 may be implemented as a component of the memory (110).

[0063] The communication module (120) may include one or more components that enable communication with an external device, and may include, for example, at least one of a broadcast reception module, a wired communication module, a wireless communication module, a short-range communication module, and a location information module.

[0064] The wired communication module may include various wired communication modules such as a Local Area Network (LAN) module, a Wide Area Network (WAN) module, or a Value Added Network (VAN) module, as well as various cable communication modules such as a Universal Serial Bus (USB), a High Definition Multimedia Interface (HDMI), a Digital Visual Interface (DVI), RS-232 (recommended standard 232), power line communication, or plain old telephone service (POTS).

[0065] The wireless communication module may include a wireless communication module that supports 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, in addition to a WiFi module and a Wireless Broadband module.

[0066] The display (130) displays (outputs) information or data processed in the device (100), data input or output through the AI ​​model (300), etc. In addition, the display (130) can display execution screen information of an application program (e.g., an application) running in the device (100), or UI (User Interface) or GUI (Graphical User Interface) information according to such execution screen information.

[0067] The input module (140) is for receiving information from a user. When a user inputs information through the input unit, the processor (150) can control the operation of the device (100) to correspond to the input information.

[0068] The input module (140) may include hardware physical keys (e.g., buttons located on at least one of the front, rear, and side of the device, dome switches, jog wheels, jog switches, etc.) and software touch keys. For example, the touch keys may be formed as virtual keys, soft keys, or visual keys displayed on a touchscreen type display (130) through software processing, or as touch keys placed on a part other than the touchscreen. Meanwhile, the virtual keys or visual keys may have various forms and be displayed on the touchscreen, and may be formed as, for example, graphics, text, icons, videos, or a combination thereof.

[0069] The processor (150) may be implemented as a memory that stores data on an algorithm for controlling the operation of components within the device (100) (including learning or executing an AI model) or a program that reproduces the algorithm, and at least one processor (not shown) that performs the aforementioned operation using the data stored in the memory. In this case, the memory and the processor may be implemented as separate chips, or may be implemented as a single chip.

[0070] In one embodiment, the system (1000) or device (100) according to the present invention may include at least one processor, and when including multiple processors, the multiple processors may be included in different devices (100).

[0071] In addition, the processor (150) can control any one or a combination of the components described above to implement various embodiments according to the present disclosure described below on the device (100).

[0072] FIG. 3 is a block diagram illustrating a method for constructing a data set that improves the zero-shot learning performance of an AI model through instruction tuning according to an embodiment of the present invention.

[0073] Referring to FIG. 3, instructions are extracted from a training task, which is data used to train an AI model, and a target task, which is information about a task for the AI ​​model to learn through the training task (S311 and S312). The training task and the target task may be stored in the memory (110) of the device (100), but are not limited thereto. The training task and the target task each include instructions and instances. Here, the instructions include instructions, information, task definitions, etc. required to perform 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 instruction may be "Determine whether the emotion expressed in the given text is positive, negative, or neutral", and the instance may be "This movie was really fun! (expressing positive emotion)". Also, for a task called 'Text Classification', the instruction might be "Classify news articles into categories such as politics, sports, entertainment, or technology", and the instance might be a 'sports article'. Also, for a task called 'Question Answering (QA)', the instruction might be "Give the given text and provide the correct answer to a question related to the text based on information", and the instance might be "Who is the author of the Harry Potter series? (Question) It is J.K. Rowling (Answer)". Also, for a task called 'Text Summarization', the instruction might be "Generate a concise and coherent summary of a long text document while preserving the main information and topic", and the instance might be "An article about today's weather".Also, for a task called 'Named Entity Linking (NEL)', the instruction might be "Link the entities recognized in the text to corresponding entries in a knowledge base or database to provide additional context and information", and the instance might be "Jeju Island is an island in Korea" (entity: Jeju Island, link: island in Korea).

[0074] Next, the similarity is evaluated by comparing the instructions of the extracted training task with the instructions of the extracted target task (S320).

[0075] Cosine similarity can be used as an example of a method for assessing similarity. Cosine similarity is a method for measuring the similarity between two vectors in a vector space. It calculates the similarity by measuring how similar the directions of two vectors are. (Cosine similarity uses how "similar" the directions of two vectors are.) More specifically, cosine similarity is defined as follows.

[0076]

[0077] The value of cosine similarity ranges from -1 to 1. The closer the value is to 1, the more similar the directions of the two vectors are, and the closer the value is to -1, the more opposite the directions of the two vectors are. For example, if the value is 0, it can be determined that the two vectors are orthogonal (at a 90 degree angle) and there is no directional similarity. In the present invention, each of the instructions of the extracted training task and the target task is expressed as a vector, and the similarity can be calculated using the two vectors.

[0078] Another example of a method for evaluating similarity is model transfer. Calculating similarity through model transfer involves training a model with the instructions of the extracted training task, inputting the instructions of the extracted target task into the trained model, and evaluating its performance. Specifically, the higher the performance of the extracted target task instructions when inputted into the trained model, the higher the similarity between the extracted target task instructions and the extracted training task. In other words, the instructions of the extracted target task and the extracted training task are given higher similarity scores in the order of their highest performance when inputted into the trained model.

[0079] Here, accuracy, loss function, etc. can be used as indicators to measure the model's performance value.

[0080] Accuracy represents the proportion of samples that the model correctly classified, and can be calculated as follows:

[0081]

[0082] For example, when the instructions of the target task extracted from the learned model are input, the accuracy can be calculated by dividing the number of samples that output the correct result by the total number of instructions of the target task extracted, and this can be used as the similarity.

[0083] The loss function is a function that measures the difference between the model's predicted value and the actual value. More specifically, the mean squared error (MSE) and cross-entropy loss can be used. For example, in regression problems, the mean squared error is used to measure the error between the model's predicted value and the actual value, and in classification problems, the cross-entropy loss is used to measure the difference between the model's predicted probability distribution and the actual distribution.

[0084] Next, among the extracted training task instructions, those with a similarity greater than a predetermined value are selected (S330). Here, the "predetermined value" may be a similarity ranking, a specific similarity score, etc. For example, it may be "instructions in the training task that are ranked 5th or higher in similarity" (similarity ranking), or "instructions in the training task that have a similarity score of 10 or higher" (specific similarity score).

[0085] Next, the instructions for the training task selected through S330 are output as a data set (S340). This data set can be used to train an AI model and, for this purpose, can be stored in the database (200) of the system (1000).

[0086] Because instructions contain information defining the characteristics of a task, a dataset containing only the instructions for a selected training task (i.e., no instances) is sufficient for training an AI model. However, training with a dataset that includes instances in addition to instructions risks negatively impacting learning due to the diverse content within the task, potentially degrading zero-shot performance. Therefore, selecting relevant tasks from among the tasks is crucial.

[0087] Through this, the present invention can increase the efficiency of computing resource utilization and improve zero-shot learning capability by providing a simple and effective method for selecting a training task related to a target task.

[0088] Figure 4 shows the results of evaluating the zero-shot learning performance of the case where learning was performed using a dataset containing both 'instructions' and 'instances' {T5(3B)+DS-BTS}, and the case where learning was performed using a dataset containing only 'instructions' of training tasks selected according to the present invention {T5(3B) + I-BTS}. The dataset used as the task is P3 (Public Pool of Prompts) (Multitask prompted training enables zero-shot task generalization, Victor Sanh et al. 38, 2022.), and P3 contains a total of 35 tasks across 8 task clusters for learning, and each task contains an average of 11.7 instructions. In addition, the evaluation results of NLI (Natural Language Inference), Sentence Completion, and Coreference Resol. (Coreference Resolution), WSD (Word Sense Disambiguation) are major categories of target tasks, while RTE (Recognizing Textual Entailment), CB (Commonsense Based), COPA (Choice of Plausible Alternatives) are minor categories of target tasks.

[0089] T0(3B) represents a model that learns all data without a selection process based on similarity, T5(3B)+DS-BTS represents a model that learns by selecting the top 5 tasks with high similarity calculated by extracting some data as samples like in the prior art, but using both the 'instructions' and 'instances' included in the tasks, and T5(3B)+I-BTS represents a model that learns by selecting the top 5 tasks with high similarity like in the present invention, but using only the 'instructions' included in the tasks.

[0090] Looking at the scores evaluating the zero-shot learning performance, it was confirmed that the performance scores of T5(3B)+I-BTS according to the present invention were higher than those of T5(3B)+DS-BTS in most tasks, and the average performance score was also 4.3% higher for T5(3B)+I-BTS according to the present invention (54.92 points) than for T5(3B)+DS-BTS (50.62 points).

[0091] These results demonstrate that instances within a task can cause negative transfer in learning, which can degrade model performance, and thus, training an AI model using only instructions from a training task, as in the present invention, can help improve zero-shot performance.

[0092] Fig. 5 shows the results of evaluating the zero-shot learning performance when selecting instructions based on the similarity of the top nth (where n is 1, 3, 5, or 10) or higher when selecting instructions of the extracted training task with a similarity of a predetermined value or higher (S330). Looking at Fig. 5, the zero-shot learning performance is shown to gradually improve from selecting the top 1st task to selecting the top 5th task. However, in some tasks, a tendency for the zero-shot learning performance to decrease was also shown when selecting the top 10th task. Accordingly, when selecting instructions based on the similarity of the top nth (where n is 1, 3, 5, or 10) or higher, it is preferable to use the top 5th or higher as the standard, but is not limited thereto.

[0093] Figure 6 is a block diagram illustrating another example of a method for evaluating similarity (step S320 of Figure 3), using a pre-tuned task selector model to evaluate similarity. The task selector model may be a separate, pre-trained AI model for more accurate similarity evaluation.

[0094] Referring to FIG. 6, first, one of the plurality of target tasks is selected as a first task (S411), and a first instruction is extracted from the first task and designated as a positive sample (S421). In addition, among the plurality of target tasks, one task other than the first task is selected as a second task (S412), and a second instruction is extracted from the second task and designated as a negative sample (S422).

[0095] Next, a high similarity score is assigned to the positive sample (S431), and a low similarity score is assigned to the negative sample (S432). Since the positive sample is extracted from the first task, it can be considered to have a relatively high similarity to the first task. Since the negative sample is extracted from a task other than the first task, it can be considered to have a relatively low similarity to the first task. Accordingly, a high similarity score is assigned to the positive sample with a high similarity, and a low similarity score is assigned to the negative sample with a low similarity.

[0096] Similarity scores may be assigned differently depending on the user, but in one example, a positive sample may be assigned a similarity score of "1" and a negative sample may be assigned a similarity score of "0." While similarity scores for positive and negative samples can be set and assigned by the user, the scope of the present invention is not limited thereto. For example, a method of vectorizing positive and negative samples and calculating the Euclidean distance between the two vectors, including the similarity evaluation method described above, may be used. Additionally, a natural language processing-based approach may be used that measures and scores the similarity between sentences using techniques such as sentence embedding, tokenization, and morphological analysis. Furthermore, the user may further tune the calculated similarity scores so that positive samples are assigned high similarity scores and negative samples are assigned low similarity scores.

[0097] Next, the positive and negative samples with assigned similarities are output as data for tuning the task selection model (S440), and the task selection model is trained using the data for tuning the task selection model (S450). Since the positive and negative samples are assigned similarities that are clearly distinguishable based on the similarities of the first and second instructions, by training the task selection model using these, the task selection model can be tuned to more clearly distinguish the similarities between the instructions of the target task and the training task.

[0098] Through this, the present invention can more accurately evaluate the similarity between the instructions of the target task and the instructions of the training task by additionally training a task selection model that evaluates similarity, and further improve the accuracy of training task selection, thereby improving the overall performance of the model.

[0099] Meanwhile, the step of extracting instructions (S311 and S312 of FIG. 3) may additionally include a step of unifying placeholders included in the extracted instructions into the same sentence.

[0100] A placeholder is a code that is temporarily filled in where elements such as characters and images included in the data are to be placed. For example, the data called "question" included in the code in Fig. 7 ("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 "sentence 2" input by the user and outputs the completed data. Here, {{word1}}, {{word2}}, {{sentence1}}, and {{sentence2}} correspond to placeholders that temporarily fill in the positions where the user's input values ​​are to be placed.

[0101] Because placeholders are not only inconsistent in form but can also be ambiguous in meaning, data containing them can negatively impact the training process of AI models. Therefore, unifying the placeholders in AI model training data into specific terms can eliminate this negative impact.

[0102] For example, placeholders can be identified and removed through data preprocessing. In Fig. 7, {{word1}} and {{word2}} included in the data can be unified as {{text}}, and {{sentence1}} and {{sentence2}} can be unified as {{candidate}}. Methods for identifying and removing placeholders include, but are not limited to, a method in which a user directly identifies and removes them, a method in which text is tokenized to identify and remove divided placeholders, and a method in which placeholder patterns are recognized to identify and remove them.

[0103] Through this, the present invention can improve zero-shot performance by unifying placeholders included in extracted instructions into specific terms, thereby preventing placeholders from having a negative impact on AI model learning.

[0104] Instruction tuning, as described in the present invention, refers to selecting training tasks relevant to a target task, thereby enhancing learning for an AI model and enabling efficient execution of complex tasks. More specifically, instruction tuning involves adjusting parameters and directives to improve the operation of an AI algorithm. Instruction tuning can increase AI learning efficiency in a simple manner and has the advantage of generalizing to new tasks (unseen tasks).

[0105] The training or target tasks used in the present invention and the instructions extracted therefrom can be configured to be suitable for performing in-context learning (ICL) by classifying tasks performing the same task together so that in-depth learning can be performed on a specific field.

[0106] In-context learning refers to a method that extracts only a portion of the data included in the downstream task and uses that context to learn. It is categorized into zero-shot learning, one-shot learning, and few-shot learning. As the number of parameters in language models increases, large-scale language models (LLMs) also learn through the context included in examples. This allows the large-scale language model (LLM) to predict the results by inputting a few examples into a pre-trained large-scale language model (LLM). More specifically, in-context refers to the flow within a prompt, and can be related to arithmetic operations, typo correction, language translation, etc. For example, in the task of summarizing conversation content, in addition to the few-shot learning method that utilizes prompt examples, zero-shot learning that can be controlled through instructions without prompt examples, and learning methods based on fine-tuning can be utilized. This allows large-scale language models (e.g., GPT-3, HyperClova, etc.) to be excellent zero-shot or few-shot learners controlled by prompts, solving natural language processing (NLP) problems by understanding context and patterns contained in small amounts of data. This learning method has the advantage of not requiring parameter updates and reducing computational loss.

[0107] Meanwhile, a method for constructing a data set for AI model learning through instruction tuning according to embodiments of the present invention can be implemented by the system described with reference to FIG. 1.

[0108] AI models according to embodiments of the present invention can be controlled, executed, trained, driven, etc. by a processor, and thus, at least one of the tasks of executing, learning, and driving the AI ​​models can be performed by at least one processor. Furthermore, the AI ​​models can be stored in memory, and feature data according to the present invention can also be stored in memory.

[0109] Meanwhile, the disclosed embodiments may be implemented in the form of a recording medium storing computer-executable instructions. The instructions may be stored in the form of program code, and when executed by a processor, may generate program modules to perform the operations of the disclosed embodiments. The recording medium may be implemented as a computer-readable recording medium.

[0110] Computer-readable storage media include all types of storage media that store instructions that can be deciphered by a computer. Examples include read-only memory (ROM), random access memory (RAM), magnetic tape, magnetic disks, flash memory, and optical data storage devices.

[0111] The disclosed embodiments have been described with reference to the attached drawings as described above. Those skilled in the art will understand that the present disclosure can be implemented in forms other than the disclosed embodiments without altering the technical spirit or essential features of the present disclosure. The disclosed embodiments are illustrative and should not be construed as limiting.

Claims

1. In a system including at least one processor and a memory storing one or more instructions, The at least one processor executes the one or more instructions stored in the memory, A step 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; A step of evaluating similarity by comparing the instructions of the extracted training task with the instructions of the extracted target task; A step of selecting instructions of the extracted training task whose similarity is greater than a predetermined value; and A step of outputting the instructions of the above-mentioned selected training task as a data set; System.

2. In claim 1, The step of evaluating the above similarity is: A system that expresses the instructions of the extracted training task and the instructions of the extracted target task as vectors and performs evaluation through cosine similarity that compares the directional similarity of the two vectors.

3. In claim 1, The step of evaluating the above similarity evaluates the similarity through a model transfer method, The above model transfer method is, Train the model transfer AI model with any one of the instructions of the above extracted training task, The performance of the model transfer AI model is evaluated by inputting instructions of a plurality of the extracted target tasks into the model transfer AI model, The above model transfer AI model provides high similarity to the instructions of the extracted target task that exhibits high performance and the instructions of the extracted training task used in learning the above model transfer AI model. System.

4. In claim 1, The step of evaluating the above similarity is performed by a task selection model that has undergone pre-tuning for the above similarity evaluation, Pre-tuning for the above similarity evaluation is A step of selecting one of the plurality of target tasks as a first task; A step of extracting a first instruction from the first task and designating it as a positive sample; A step of selecting one of the plurality of target tasks other than the first task as a second task; A step of extracting a second instruction from the second task and designating it as a voice sample; A step of assigning a similarity score to the positive sample and the negative sample; A step of training the task selection model using the positive sample and the negative sample including the similarity score; System.

5. In claim 4, The step of assigning a similarity score to the positive sample and the negative sample is as follows: A system that assigns a similarity score of "1" to the positive sample and a similarity score of "0" to the negative sample.

6. In claim 1, The step of extracting the above instructions is: A system further comprising a step of unifying placeholders included in the instructions of the extracted training task and the instructions of the extracted target task into specific terms.

7. A method for generating prediction data performed by at least one processor, A step 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; A step of evaluating similarity by comparing the instructions of the extracted training task with the instructions of the extracted target task; A step of selecting instructions of the extracted training task whose similarity is greater than a predetermined value; A method comprising: a step of outputting instructions of the selected training task as a data set; 8. In claim 7, The step of evaluating the above similarity is: A method for expressing the instructions of the extracted training task and the instructions of the extracted target task as vectors, and performing evaluation through cosine similarity that compares the directional similarity of the two vectors.

9. In claim 7, The step of evaluating the above similarity evaluates the similarity through a model transfer method, The above model transfer method is, Train the model transfer AI model with any one of the instructions of the above extracted training task, The performance of the model transfer AI model is evaluated by inputting instructions of multiple extracted target tasks into the model transfer AI model. A method for granting high similarity to the instructions of the extracted target task in which the model transfer AI model exhibits high performance and the instructions of the extracted training task used in learning the model transfer AI model.

10. In claim 7, The step of evaluating the above similarity is performed by a task selection model that has undergone pre-tuning for the above similarity evaluation, Pre-tuning for the above similarity evaluation is A step of selecting one of the plurality of target tasks as a first task; A step of extracting a first instruction from the first task and designating it as a positive sample; A step of selecting one of the plurality of target tasks other than the first task as a second task; A step of extracting a second instruction from the second task and designating it as a voice sample; A step of assigning a similarity score to the positive sample and the negative sample; A method comprising: a step of training the task selection model using the positive sample and the negative sample including a similarity score; 11. In claim 10, The step of assigning a similarity score to the positive sample and the negative sample is as follows: A method in which a similarity score of “1” is given to the positive sample and a similarity score of “0” is given to the negative sample.

12. In claim 7, The step of extracting the above instructions is: A method further comprising a step of unifying placeholders included in the instructions of the extracted training task and the instructions of the extracted target task into specific terms.

13. A program stored on a computer-readable recording medium that causes a computer to execute the method of any one of claims 7 to 12.

Images