Method and system for performing instruction tuning using heterogeneous languages, and instruction tuning-based inference method.

Abstract

The method and system for performing instruction tuning of a language model using heterogeneous languages ​​according to embodiments of the present invention relates to a method and system for performing instruction tuning of a language model using heterogeneous languages ​​that can improve the zero-shot performance and generalization performance of a language model by performing instruction tuning using different heterogeneous languages.

Description

Technical Field

[0001] This Disclosure relates to a method and system for performing instruction tuning using different languages. More specifically, it relates to a method and system for performing instruction tuning using different languages that can improve instruction tuning performance for different languages.

Background Art

[0002] Recently, pre-trained language models (LLMs) for large-scale general domain data have emerged, and various tasks that were previously processed manually have been replaced by artificial intelligence-based methods.

[0003] As the development of extremely large language models such as Chat GPT, Google's Gemini, Naver's HyperClover, KakaoBrain's KoGPT, and LG's EXAONE increases, various methods for increasing the zero-shot performance of extremely large language models have been studied.

[0004] In particular, among the methods for increasing the zero-shot performance of extremely large language models, research on instruction tuning techniques has been actively conducted. Instruction tuning is a learning method first published in Google's FLAN (finetuned Language Models are Zero-Shot Learners) paper, which means a technique for fine-tuning a large language model (LLM) using an instruction tuning dataset to increase zero-shot performance.

[0005] Instruction tuning is a technique in which a massive language model learns multiple tasks along with instructions (descriptions about the tasks), and then, when given instructions for a new task, it understands the task solely from the instructions without any further learning.

[0006] However, instruction tuning requires diverse task data, and along with that, diverse instructions must be built, which presents a problem in that building the necessary resources takes a lot of time and money. [Overview of the Initiative] [Problems that the invention aims to solve]

[0007] Book Disclosure The embodiment aims to provide a method and system for performing instruction tuning using heterogeneous languages ​​in order to reduce the cost and time required to build resources for instruction tuning.

[0008] Also, book Disclosure The embodiment aims to provide a method and system for performing heterogeneous language instruction tuning to increase the zero-shot performance of a super-large language model by utilizing heterogeneous language instruction tuning.

[0009] However, this Disclosure The technical challenges that this embodiment aims to address are not limited to those described above, and other technical challenges may exist. [Means for solving the problem]

[0010] Book DisclosureThe method for performing instruction tuning using heterogeneous languages ​​according to this embodiment involves a computing system including memory and a processor using heterogeneous languages ​​to perform instruction tuning of a large language model (LLM). A method for performing instruction tuning, comprising the steps of: setting up a first instruction tuning dataset containing one or more first tasks formed in a first language using metadata; setting up a second instruction tuning dataset containing one or more second tasks formed in a second language; generating cross-language instructions; and performing instruction tuning on the first instruction tuning dataset and the second instruction tuning dataset using the cross-language instructions, wherein the step of generating cross-language instructions includes the steps of generating one or more first instructions created in the first language and generating one or more second instructions created in the second language; and the step of performing second instruction tuning includes the steps of performing instruction tuning on the second instruction tuning dataset using the first instructions and performing instruction tuning on the first instruction tuning dataset using the second instructions, wherein the first instructions and the second instructions are set to have the same format.

[0011] In other respects, the step of setting up the first instruction tuning dataset includes the step of generating N or more pre-configured first tasks, where N is a natural number of 1 or more.

[0012] In other respects, the step of generating the first instruction includes generating 3N first instructions and applying 3 instructions to each task.

[0013] In other respects, the step of setting up the first instruction tuning dataset includes generating 34 NLU (Natural Language Understanding) tasks and 17 NLG (Natural Language Generation) tasks based on data collected from one or more of the following: AIHub1, KorPora2, GIthub, Huggingface, KLUE3, Korquad4, ETRI5, Modu's Corpus, and KoBest.

[0014] In other respects, the language model utilizes a multilingual model.

[0015] In other respects, the step of generating the cross-language instructions may include the step of applying one or more first instructions to the second instruction tuning dataset, and the step of applying one or more second instructions to the first instruction tuning dataset.

[0016] On the other hand, book DisclosureThe method for performing instruction tuning using heterogeneous languages ​​according to the embodiment includes the steps of: setting up a first instruction tuning dataset containing one or more first tasks formed in a first language using metadata; setting up a second instruction tuning dataset containing one or more second tasks formed in a second language; generating cross-language instructions; and applying the cross-language instructions to the first instruction tuning dataset and the second instruction tuning dataset to perform inference on the data input to the language model, wherein the step of generating cross-language instructions includes the steps of generating one or more first instructions created in the first language and generating one or more second instructions created in the second language, and the step of performing inference includes applying the first instructions to the second instruction tuning dataset and applying the second instructions to the first instruction tuning dataset to perform inference, wherein the first instructions and the second instructions are set to have the same format.

[0017] On the other hand, book DisclosureThe method for performing instruction tuning using heterogeneous languages ​​according to the embodiment includes the steps of: setting up a first instruction tuning dataset containing one or more first tasks formed in a first language using metadata; setting up a second instruction tuning dataset containing one or more second tasks formed in a second language; generating cross-language instructions; performing instruction tuning on the first instruction tuning dataset and the second instruction tuning dataset using the cross-language instructions; and applying the cross-language instructions to the first instruction tuning dataset and the second instruction tuning dataset to perform inference on the data input to the language model, the step of generating the cross-language instructions. The process includes the steps of generating one or more first instructions created in the first language, and generating one or more second instructions created in the second language, wherein the second instruction tuning step includes the steps of using the first instructions to perform instruction tuning on the second instruction tuning dataset and using the second instructions to perform instruction tuning on the first instruction tuning dataset, wherein the inference step includes the steps of applying the first instructions to the second instruction tuning dataset and applying the second instructions to the first instruction tuning dataset to perform inference, and the first instructions and second instructions are set to have the same format.

[0018] On the other hand, book DisclosureAn instruction tuning execution system using heterogeneous languages ​​according to an embodiment includes at least one memory and at least one processor that reads at least one application stored in the memory and performs instruction tuning of a language model using heterogeneous languages, wherein the processor sets a first instruction tuning dataset containing one or more first tasks formed in the first language using metadata, sets a second instruction tuning dataset containing one or more second tasks formed in the second language, generates cross-language instructions containing one or more first instructions created in the first language and one or more second instructions created in the second language, and uses the cross-language instructions to set the first instruction tuning dataset and the second instruction tuning The instructions include a step of performing instruction tuning on a learning dataset, applying the cross-language instructions to the first instruction tuning dataset and the second instruction tuning dataset to perform inference on the data input to the language model, using the first instruction to perform instruction tuning on the second instruction tuning dataset, using the second instruction to perform instruction tuning on the first instruction tuning dataset, applying the first instruction to the second instruction tuning dataset, applying the second instruction to the first instruction tuning dataset to perform inference, and the first and second installations are set to have the same format.

[0019] On the other hand, book DisclosureThe computing device according to the embodiment includes at least one memory and at least one processor that reads at least one application stored in the memory and performs instruction tuning of a language model using heterogeneous languages, wherein the processor's instructions include the steps of: setting up a first instruction tuning dataset containing one or more first tasks formed in a first language using metadata; setting up a second instruction tuning dataset containing one or more second tasks formed in a second language; generating cross-language instructions; performing instruction tuning on the first instruction tuning dataset and the second instruction tuning dataset using the cross-language instructions; and applying the cross-language instructions to the first instruction tuning dataset and the second instruction tuning dataset and inputting them into the language model. The step of generating the cross-language instruction includes an instruction for performing an inference on the data, the step of generating the cross-language instruction includes a step of generating one or more first instructions created in the first language, and a step of generating one or more second instructions created in the second step, the step of performing instruction tuning includes a step of performing instruction tuning on the second instruction tuning dataset using the first instruction, and a step of performing instruction tuning on the first instruction tuning dataset using the second instruction, the step of performing inference includes a step of applying the first instruction to the second instruction tuning dataset, and applying the second instruction to the first instruction tuning dataset to perform inference, and the first instruction and the second instruction are set to have the same format. [Effects of the Invention]

[0020] Book Disclosure The method and system for executing instruction tuning using different languages according to an embodiment can reduce the costs and time involved in constructing resources for instruction tuning.

[0021] Also, the Disclosure method and system for executing instruction tuning using different languages according to an embodiment can increase the zero-shot performance of a large language model by using instruction tuning with different languages.

[0022] However, the effects obtained herein are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood from the following description. Disclosure

Brief Description of the Drawings

[0023] [Figure 1] An example of a block diagram of a computing system for implementing instruction tuning of a language model using different languages according to an embodiment of the present disclosure is shown. [Figure 2] An example of a block diagram of a computing device for implementing instruction tuning of a language model using different languages according to an embodiment of the present disclosure is shown. [Figure 3] An example of a block diagram from another aspect of a computing device for implementing instruction tuning of a language model using different languages according to an embodiment of the present disclosure is shown. [Figure 4] It is an exemplary block diagram for explaining an instruction tuning system of a language model using different languages according to an embodiment of the present disclosure. [Figure 5] It is an exemplary diagram of a Korean instruction tuning dataset according to an embodiment of the present disclosure. [Figure 6] It is an exemplary diagram of tasks and instructions according to an embodiment of the present disclosure. [Figure 7]This is an illustrative diagram of an English instruction tuning dataset according to an embodiment of the present disclosure. [Figure 8] This is an illustrative diagram of cross-language instruction according to an embodiment of the present disclosure. [Figure 9] This is an illustrative diagram of a language model that has undergone instruction tuning using heterogeneous languages ​​according to an embodiment of the present disclosure. [Figure 10] This is a performance results table for the language model according to the embodiments of this disclosure. [Figure 11] This is a performance graph of a language model according to an embodiment of the disclosure. [Figure 12] This is a flowchart illustrating a method for tuning a language model using heterogeneous languages ​​according to an embodiment of the disclosure. [Figure 13] This is a flowchart illustrating a method for tuning a language model using heterogeneous languages ​​according to an embodiment of the disclosure. [Figure 14] This is a flowchart illustrating a method for tuning a language model using heterogeneous languages ​​according to an embodiment of the disclosure. [Modes for carrying out the invention]

[0024] Book Disclosure Since it can undergo various transformations and have a variety of embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. Disclosure The effects and characteristics, and the methods for achieving them, will become clear when you refer to the embodiments described in detail later with the drawings. However, this DisclosureThe embodiments disclosed below are not limited to those disclosed below and can be realized in a variety of forms. In the embodiments below, terms such as "first," "second," etc., are not restrictive and are used to distinguish one component from another. Also, singular expressions include plural expressions unless they have a clearly different meaning in context. Also, terms such as "includes" or "has" mean that the features or components described in the specification exist, and do not preclude the possibility that one or more other features or components may be added. Also, the sizes of components may be exaggerated or reduced in the drawings for illustrative purposes. For example, the sizes and thicknesses of each component shown in the drawings are arbitrarily shown for illustrative purposes, so Disclosure This is not necessarily limited to what is shown in the illustration.

[0025] Please refer to the attached drawings below. Disclosure When describing embodiments in detail and referring to the drawings, identical or corresponding components will be given the same reference numerals, and redundant descriptions thereof will be omitted.

[0026] The following describes in detail an example system for achieving instruction tuning of a large language pretraining model (LLM) using different heterogeneous languages, with reference to the attached diagrams.

[0027] Figure 1 is the book Disclosure An example block diagram of a computing system that enables instruction tuning of language models using heterogeneous languages ​​according to an embodiment is shown.

[0028] As shown in Figure 1, Disclosure The computing system 1000, which enables instruction tuning of language models using heterogeneous languages, includes a user computing device 110, a server computing system 130, and a training computing system 150, and the devices are able to communicate via a network 170.

[0029] Book Disclosure The instruction tuning method for a language model using heterogeneous languages ​​according to this embodiment can be implemented and provided locally by the user computing device 110, implemented and provided in the form of a web service by a server computing system 130 that communicates with the user computing device 110, or implemented and provided by the user computing device 110 and the server computing system 130 working together.

[0030] In this embodiment, the user computing device 110 and / or the server computing system 130 can train language models (120 and / or 140, machine learning models) through interaction with a training computing system 150 that is communicatively connected via a network 170. The training computing system 150 may be separate from the server computing system 130 or may be part of the server computing system 130.

[0031] In this configuration, the artificial intelligence model (such as a language model in some embodiments) can be trained locally by the user computing device 110, trained through interaction between the server computing system 130 and the user computing device 110 via the network 170, or trained by a separate training computing system 150 using a variety of training and learning techniques. The training computing system 150 can also transmit and provide / update the trained artificial intelligence model to the user computing device 110 and / or the server computing system 130 via the network 170.

[0032] In one embodiment, the training computing system 150 may be part of the server computing system 130 or part of the user computing device 110.

[0033] The user computing device 110 may include all other types of computing devices, such as smartphones, mobile phones, digital broadcasting devices, PDAs (personal digital assistants), PMPs (portable multimedia players), desktops, wearable devices, embedded computing devices, and / or tablet PCs.

[0034] Such a user computing device 110 includes at least one processor 111 and memory 112. Here, the processor 111 may consist of at least one or more electrically connected processors from among a central processing unit (CPU), graphics processing unit (GPU), ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), controllers, microcontrollers, microprocessors, and / or other electrical units for performing functions.

[0035] The memory 112 may include one or more non-temporary / temporary computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, and combinations thereof, and may also include web storage of a server that performs memory storage functions over the internet. Such memory 112 can store data 113 and instruction words 114 necessary for at least one of the processors 111 to train an artificial intelligence model or to perform functional operations such as performing vision checks using the artificial intelligence model.

[0036] In one embodiment, the user computing device 110 can store at least one or more machine learning models 120.

[0037] The machine learning model 120 may be a variety of machine learning models, such as multiple neural networks (e.g., deep neural networks), or other types of machine learning models, including nonlinear and / or linear models, or may be composed of combinations thereof.

[0038] In this case, the neural network may include at least one of the following: feed-forward neural networks, recurrent neural networks (e.g., long-short-term memory recurrent neural networks), convolutional neural networks, and / or other forms of neural networks.

[0039] In this embodiment, the user computing device 110 receives at least one machine learning model 120 from the server computing system 130 via the network 170, stores it in the memory 112, and then executes the stored machine learning model 120 using the processor 111 to perform anomaly detection and the like.

[0040] In this embodiment, the server computing system 130 includes at least one machine learning model 140 and operates using the machine learning model 140, and can provide the user with a language model that has been instruction-tuned using heterogeneous languages ​​in conjunction with the user computing device 110 by communicating with the user computing device 110 and related data.

[0041] For example, the user computing device 110 can provide a language model that has undergone instruction tuning, in which the server computing system 130 uses a machine learning model 140 via the web to provide output in response to user input.

[0042] Furthermore, the artificial intelligence model can also be implemented in a manner in which at least a portion of the machine learning models 120 and / or 140 are executed on the user computing device 110, and the remainder is executed on the server computing system 130.

[0043] Furthermore, the user computing device 110 may include at least one input component 121 that senses user input. For example, the user input component 121 may include a touch sensor (e.g., a touchscreen and / or touchpad) that senses touch from the user's input medium (e.g., a finger or stylus), an image sensor that senses the user's motion input, a microphone that senses the user's voice input, buttons, a mouse and / or keyboard, etc. Also, if the user input component 121 receives input to an external controller (e.g., a mouse and / or keyboard) via an interface, it may include an interface and an external controller.

[0044] The server computing system 130 includes at least one processor 131 and memory 132. Here, the processor 131 may consist of at least one or more electrically connected processors from among central processing units (CPUs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), digital signal processors (DSSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, and / or other electrical units for performing functions.

[0045] The memory 132 may include one or more non-temporary / temporary computer-readable storage media, such as RAM, ROM, EEPROM, EPROM, flash memory devices, and magnetic disks, or combinations thereof. Such a memory 132 can store data 133 and instruction words 134 necessary for the processor 131 to train an artificial intelligence model or to perform functional operations such as anomaly detection by the artificial intelligence model.

[0046] In some embodiments, the server computing system 130 may be implemented to include at least one computing device. For example, the server computing system 130 may be implemented to operate multiple computing devices according to a sequential computing architecture, a parallel computing architecture, or a combination thereof. The server computing system 130 may also include multiple computing devices connected by a network 170.

[0047] Furthermore, the server computing system 130 can store at least one or more machine learning models 140. For example, the server computing system 130 may include neural networks and / or other multilayer nonlinear models as machine learning models 140. Exemplary neural networks may include feedforward neural networks, deep neural networks, recurrent neural networks, and convolutional neural networks.

[0048] The training computing system 150 includes at least one processor 151 and memory 152. Here, the processor 151 may consist of at least one or more electrically connected processors from among central processing units (CPUs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), digital signal processors (DSSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, and / or other electrical units for performing functions.

[0049] The memory 152 may also include one or more non-temporary / temporary computer-readable storage media, such as RAM, ROM, EEPROM, EPROM, flash memory devices, and magnetic disks, or combinations thereof. Such a memory 152 can store data 153 and instruction words 154 necessary for the processor 151 to perform tasks such as training an artificial intelligence model.

[0050] For example, the training computing system 150 may include a model trainer 160 that trains machine learning models 120 and / or 140 stored in the user computing device 110 and / or server computing system 130 using a variety of training or learning techniques, such as back-propagation of errors (according to the framework shown in Figure 3).

[0051] For example, such a model trainer 160 can perform backpropagation-based updates to one or more parameters of the machine learning models 120 and / or 140 based on a defined loss function.

[0052] In some implementations, error backpropagation may include truncated backpropagation through time. The model trainer 160 can perform a number of generalization techniques (e.g., weight reduction, dropout, and / or knowledge distillation) to improve the generalization capabilities of the machine learning models 120 and / or 140 being trained.

[0053] In particular, the model trainer 160 can train the machine learning models 120 and / or 140 based on a series of training data 161. Here, the training data 161 may include data of different forms, such as images, audio samples and / or text. Examples of image types that can be used may include video frames, LiDAR point clouds, X-ray images, computed tomography scans, superspectroscopic images and / or various other forms of images.

[0054] Such training data 161 can be provided by the user computing device 110 and / or the server computing system 130. When the training computing device trains the machine learning models 120 and / or 140 on specific data from the user computing device 110, the machine learning models 120 and / or 140 can be characterized as personalized models.

[0055] Furthermore, the Model Trainer 160 includes computer logic that is used to provide the desired functionality.

[0056] The model trainer 160 may also be implemented by hardware, firmware, and / or software that control a general-purpose processor. In one implementation, the model trainer 160 includes a program file stored in a storage device, which is loaded into memory 152 and can be executed by one or more processors 151. In another implementation, the model trainer 160 includes one or more sets of computer-executable data 153 and instruction words 154 stored in a typical computer-readable storage medium such as a RAM hard disk or an optical or magnetic medium.

[0057] Network 170 includes, but is not limited to, 3GPP® (3rd Generation Partnership Project) networks, LTE (Long Term Evolution) networks, WiMAX (World Interoperability for Microwave Access) networks, the Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), Bluetooth® (Bluetooth) networks, satellite broadcasting networks, analog broadcasting networks, and / or DMB (Digital Multimedia Broadcasting) networks.

[0058] Generally, communication over network 170 can be conducted using any type of wired and / or wireless connection, and can be carried out using various communication protocols (e.g., TCP / IP, HTTP, SMTP, and / or FTP), encoding or formatting (e.g., HTML and / or XML), and / or protection schemes (e.g., VPN, secure HTTP, and / or SSL).

[0059] Figure 2 shows this book Disclosure An example block diagram of a computing device that enables instruction tuning of a language model using heterogeneous languages ​​according to an embodiment is shown.

[0060] Including Figure 2, the computing device 100 included in the user computing device 110, the server computing system 130, and the training computing system 150 contains numerous applications (e.g., Application 1 to Application N). Each application may include a machine learning library and one or more machine learning models. For example, applications may include a language processing application, a text messaging application, an email application, a writing application, a virtual keyboard application, a browser application, and / or a chatbot application.

[0061] In one embodiment, the computing device 100 may include a model trainer 160 for training an artificial intelligence model, and by storing and operating the trained artificial intelligence model, it can provide output data based on predetermined input data (such as image data in one embodiment).

[0062] Each application of computing device 100 can communicate with numerous other components of computing device 100, such as at least one sensor, a context manager, a device status component, and / or additional components. Each application can communicate with each device component using an API (e.g., a public API). The API used by each application may also be specific to that application.

[0063] Figure 3 is the book Disclosure An example of a block diagram from another aspect of a computing device 100 that realizes instruction tuning of a language model using heterogeneous languages ​​according to the embodiment is shown.

[0064] As shown in Figure 3, the computing device 300 contains numerous applications (e.g., Application 1 to Application N). Each application can communicate with the central intelligence layer. For example, applications may include language processing applications, text messaging applications, email applications, writing applications, virtual keyboard applications, and / or browser applications. Each application can communicate with the central intelligence layer (and the models stored within it) using an API (e.g., a common API across all applications).

[0065] The central intelligence layer may include multiple machine learning models. For example, as shown in Figure 3, at least a portion of each machine learning model may be provided to each application and managed by the central intelligence layer. In other implementations, two or more applications may share a single machine learning model. For example, in some implementations, the central intelligence layer may provide a single model to all applications. In some implementations, the central intelligence layer may be included in or implemented separately from the operational structure of the computing device 300.

[0066] The central intelligence layer can communicate with the central device data layer. The central device data layer may be a centralized data storage for the computing device 300. As shown in Figure 3, the central device data layer can communicate with many other components of the computing device 300, such as one or more sensors, a context manager, a device status component, and / or additional components. In some implementations, the central device data layer can communicate with each device component using an API (e.g., a private API).

[0067] The technologies described herein may refer to servers, databases, software applications, and other computer-based systems, as well as the actions performed and the information sent to or from said systems. The inherent flexibility of computer-based systems will be recognized as allowing for a wide range of possible configurations, combinations, and divisions of work, and functionality between and from components. For example, the processes described herein can be implemented using a single device or component, or multiple devices or components operating in combination. Databases and applications can be implemented in a single system or in a distributed system across multiple systems. Distributed components can operate sequentially or in parallel.

[0068] See Figure 4 below. Disclosure This document describes an instruction tuning system for language models using heterogeneous languages ​​according to an embodiment of this system.

[0069] Figure 4 is the book Disclosure This is an illustrative block diagram illustrating an instruction tuning system for language models using heterogeneous languages ​​according to an embodiment.

[0070] Book Disclosure The instruction tuning system 1000 using heterologous languages ​​according to this embodiment may refer to a language model (Large Language Pretraining model, LLM) that performs instruction tuning using two different languages.

[0071] The heterologous language-based instruction tuning system 1000 includes memory and a processor. The memory stores at least one application, and the processor reads the application stored in memory and uses the heterologous language to perform instruction tuning on the learning model.

[0072] The functionality for performing instruction tuning using different languages ​​can be illustrated using a block diagram, as shown in Figure 4. Referring to Figure 4, the processor can perform the functions of the instruction tuning dataset generation module 1100, the cross-language instruction generation module 1200, the instruction tuning module 1300, and the inference module 1400, which will be described later.

[0073] The instruction tuning dataset generation module 1100 generates instruction tuning datasets using data created in different languages ​​in order to perform instruction tuning for each language. In this case, the instruction tuning dataset generation module 1100 can generate a first instruction tuning dataset for the first language and a second instruction tuning dataset for the second language in a different manner.

[0074] The instruction tuning dataset generation module 1100 generates and secures a variety of tasks using metadata and open source to generate the first instruction tuning dataset. The instruction tuning dataset generation module 1100 then sets up M (where M is a natural number greater than or equal to 1) pre-configured instructions (templates) for each task.

[0075] In this case, M may be set to have a value of 10, but Disclosure The embodiments are not limited to these.

[0076] Specifically, the instruction tuning dataset generation module 1100 can generate and classify datasets including language understanding and language generation tasks with a variety of open-source software such as AIHub1, Korpora2, Github, Huggingface, KLUE3, Korquad4, and ETRI5. Furthermore, the instruction tuning dataset generation module 1100 can construct dataset clusters using heuristic rules.

[0077] For example, as shown in Figure 5, the instruction tuning dataset generation module 1100 can generate a first instruction tuning dataset such that it includes 17 NLG (Natural Language Generation) datasets 1110 and 34 NLU (Natural Language Understanding) datasets 1120, and is classified into a total of 17 task clusters.

[0078] Specifically, the summaries, closed-book QA, paraphasing, structure-to-text, dialogues, translation, sentiment, hate speech, extractive QA, word sense disambiguation, coreference resolution, topic classification, natural language inference, intent, paraphase identification, sentence completion, and multiple choice QA shown in Figure 5 represent 17 task cluster classifications.

[0079] Also, Book, Dacon News, Report, Document News, Document Editorial, ETRI QA, Similar Corpus, Com Gen, AIHub Daily Dial, AIHub Emo Dial, AIHub TOD, AIHub Minwon, AIHub Korean Dialog, Twitter, Ko-En Parallel, Ko-En Social, Ko-En Technology show 17 NLG datasets, NSMC, Naver Shopping, Kobest Sentineg, Sosang Sentiment, AIHub Emo, Apeach, BEEP!, Curse Detection, UnSmile, Kobest BooIQ, AIHub MRC, Book MRC, KLUE MRC, KorQuAD1, News QA, NIA QA, Kobeast WIC, NIKL Coref, Callcenter, Ko Conversation, KLUE TC, KLUE NLI, KorNLI, Sae4k, StyleKQC, Daily Chat, KLUE STS, KorSTS, KorSS, Question Pair, ParaKQC, Kobest COPA, Kobest Hellaswag, and Document QA refer to 34 NLU datasets.

[0080] In this way, the instruction tuning dataset generation module 1100 generates a variety of NLP (Natural Language Processing) tasks to generate the first instruction tuning dataset, and sets M instructions for each task.

[0081] At this time, the instruction tuning dataset generation module 1100 generates instructions for the first instruction tuning dataset by using all or part of the data labels included in the first instruction tuning dataset, or by adding new data labels.

[0082] Specifically, as shown in Figure 6, task 1101 included in the first instruction tuning dataset contains multiple data labels and values ​​corresponding to those data labels. The instruction tuning dataset generation module 1100 generates instructions 1150 for task 1101 by using all or part of the data labels included in task 1101, or by adding new data labels. In this case, multiple instructions 1150 can be generated for task 1101, and more than 10 instructions 1150 can be generated and set for each task 1101.

[0083] Furthermore, the instruction tuning dataset generation module 1100 can generate a second instruction tuning dataset created in a second language. The instruction tuning dataset generation module 1100 can generate a second instruction tuning dataset using the P3 data of T0.

[0084] For example, as shown in Figure 7, the second instruction tuning dataset can be generated such that it is classified into a total of 12 task clusters, including 11 NLG (Natural Language Generation) datasets 1111 and 51 NLU (Natural Language Understanding) datasets 1121.

[0085] In other words, the tasks shown in Figure 7—Summarization, Closed-Book QA, Structure-to-Text, Sentiment, Word Sense Disambiguation, Extractive QA, Coreference Resolution, Multiple Choice QA, Paraphrase Identification, Sentence Completion, Natural Language Inference, and Topic Classification—correspond to 12 task clusters.

[0086] Additionally, CNN Daily Mail, Gigaword, MutiNews, SamSum, XSum, Hotpot QA, TriviaQA, WebQuestions, Wiki QA, Common Gen, and Wiki Bio represent 11 NLG datasets, including Amazon, App Reviews, Emo, Emotion, IMDB, Rotten Tomatoes, Yelp, WIC, Adversarial QA, BooIQ, DuoRC, DROP, Quoref, ReCoRD, ROPES, SQuAD(V1), PubMedQA, Winogrande, WSC, ARC, Art, Cbt, CoS-E, Cosmos QA, DREAM, MultiRC, OpenBookQA, PiQA, QASC, QuAIL, QuaRel, QuaRTz, RACE, SciQ, Social IQA, and Wiki Hop, WiQA, MRPC, PAWS, QQP, COPA, StoryCloze, Helaswag, ANLI (R1-3), CB, RTE, EsNLI, AG News, DBPedia, TREC, and Yahho Answers Topic refer to 51 NLU datasets, 1121 in total.

[0087] Furthermore, the instruction tuning dataset generation module 1100 can generate instructions for the second instruction tuning dataset by using all or part of the data labels included in the second instruction tuning dataset, or by adding new data labels.

[0088] The cross-language instruction generation module 1200 generates cross-language instructions for tuning the language model's instructions using different languages ​​(first language and second language).

[0089] The cross-language instruction generation module 1200 can increase the effectiveness of instruction tuning using two languages ​​by generating and setting N cross-language instructions (where N is a natural number greater than or equal to 1) per task included in the first instruction tuning dataset and the second instruction tuning dataset.

[0090] In this case, N may be set to have a value of 3, but Disclosure The embodiments are not limited to these.

[0091] Specifically, the cross-language instruction generation module 1200 generates a first instruction created in the first language and a second instruction created in the second language. The first and second instructions are configured to have the same format (data labels). Furthermore, the first instruction is applied to the second instruction tuning dataset, and the second instruction is applied to the first instruction tuning dataset.

[0092] In other words, the dataset and instructions are configured so that the language of the instruction tuning dataset and the language of the instructions intersect.

[0093] The cross-language instruction generation module 1200 can generate first instructions by machine translating the instructions for the second instruction tuning dataset into the first language, and generate second instructions by machine translating the instructions for the first instruction tuning dataset into the second language.

[0094] Furthermore, the cross-language instruction generation module 1200 can generate cross-language instructions by preferentially selecting data labels that are commonly included in the basic instructions for the first instruction tuning dataset and the data labels included in the second instruction tuning dataset.

[0095] Furthermore, the cross-language instruction generation module 1200 can generate cross-language instructions by setting weight values ​​for each data label based on their frequency of use among the basic instructions for the first instruction tuning dataset and the data labels included in the second instruction tuning dataset, and by adding data labels whose weight values ​​are equal to or greater than the baseline value.

[0096] Furthermore, the cross-language instruction generation module 1200 adds, deletes, and modifies the data labels of the first instruction and the labels of the second instruction to make the format of the first instruction and the format of the second instruction identical.

[0097] Figure 8 below shows an example of cross-linguistic instruction.

[0098] Table 1 in Figure 8 shows an example of cross-language instructions for Xsum:Summarization, Table 2 shows an example of cross-language instructions for WSC:Coreference Resolution, and Table 3 shows an example of cross-language instructions for Emotion:Sentiment.

[0099] In Tables 1 through 3, P3 Template refers to the instructions for the second instruction tuning dataset, Translated Template refers to the instructions that have been simply translated into the first language, and Cross-Lingual Templates refers to cross-linguistic instructions in which the order, position, or presence or absence of representation of data labels has been altered.

[0100] The instruction tuning module 1300 uses the cross-language instructions generated by the cross-language instruction generation module 1200 to perform instruction tuning on the first instruction tuning dataset and the second instruction tuning dataset.

[0101] In other words, language models (LLMs) can be trained using cross-language instructions, the first instruction tuning dataset, and the second instruction tuning dataset.

[0102] Furthermore, by using the inference module 1400, inference using cross-language instructions can be performed in the inference stage, either separately from the learning stage or in addition to the learning stage.

[0103] In other words, the language model can be trained using cross-language instructions and the first instruction tuning dataset and the second instruction tuning dataset, or it can perform inference on newly input data using cross-language instructions and the first instruction tuning dataset and the second instruction tuning dataset, or it can be trained and perform inference on new data using cross-language instructions and the first instruction tuning dataset and the second string tuning dataset.

[0104] In this case, a multilingual model can be used to utilize both the first and second languages.

[0105] The language models that utilize cross-linguistic instruction during the learning and inference phases, as well as the performance of each model, are explained below with reference to Figure 9.

[0106] Figure 9 shows an example of a language model that has been trained and inferred from cross-language instructions.

[0107] In Figure 9, the dotted line separates training and evaluation, with the area above the dotted line representing training and the area below the dotted line representing evaluation. Also in Figure 9, the solid line separates monolingual and heterolingual cases, with the area to the left of the solid line representing monolingual cases and the area to the right representing heterolingual cases.

[0108] Furthermore, in Figure 9, En-mT0 refers to a language model trained using the aforementioned second instruction tuning dataset. En-mT0-CT refers to a language model that is trained on the second instruction tuning dataset using the aforementioned cross-language instructions only during training, and performs inference using the original instructions. En-mT0(CI) refers to a model that is trained on the second instruction tuning dataset using the original instructions and performs inference using cross-language instructions.

[0109] Furthermore, Ko-mT0 refers to a language model trained using the aforementioned first instruction tuning dataset. Ko-mT0-CT refers to a language model that is trained on the first instruction tuning dataset using the aforementioned cross-language instructions only during training, and performs inference using the original instructions. Ko-mT0(CI) refers to a model that is trained on the first instruction tuning dataset using the original instructions and performs inference using cross-language instructions.

[0110] In this case, the examples mentioned above and below were explained assuming that the first language is Korean and the second language is English, but Disclosure The embodiments are not limited to these.

[0111] To evaluate the zero-shot performance of each model for each task, evaluations were conducted using two separate holdout settings. The first group included four tasks: natural language inference, sentence completion, cross-reference resolution, and word semantic clarity. The second group consisted of three tasks: sentiment analysis, summarization, and objective QA.

[0112] Figure 10 shows the zero-shot performance and language-generalized performance scores for each model.

[0113] As shown in Figure 10, performance improvements were observed even when instruction tuning was applied to two different languages. Specifically, tasks such as objective QA, summarization, and sentence completion in Korean evaluation showed similar performance between the En-mT0 and Ko-mT0 models. In English evaluation, Ko-mT0 showed similar performance to En-mT0 in sentiment analysis and summarization tasks.

[0114] Furthermore, CT and CI models utilizing cross-language instructions showed improved performance across most metrics compared to models using original instructions. Specifically, in evaluations of Korean, En-mT0-CT and En-mT0(CI) showed significant performance improvements compared to En-mT0. Similarly, in evaluations of English, Ko-mT0-CT and Ko-mT0(CI) showed improved performance compared to Ko-mT0. Therefore, it was confirmed that applying cross-language instructions to train and infer language models improves performance compared to instruction tuning using a single language.

[0115] Furthermore, Figure 11 is a graph showing the average performance for both Korean and English. As shown in Figure 11, the performance of all instruction-tuned models improves as the model size increases. It can also be seen that models utilizing cross-language instructions (En-mT0-CT, En-mT0-CI, Ko-mT0-CT, Ko-mT0-CI) show an even greater degree of performance improvement compared to general models (En-mT0, Ko-mT0) across a variety of model sizes.

[0116] The following refers to Figures 12 to 14. Disclosure This document provides a detailed explanation of how to perform instruction tuning of a language model using heterogeneous languages ​​according to the embodiment described.

[0117] Figures 12 to 14 are shown in this book. Disclosure This is a flowchart illustrating how to perform instruction tuning of a language model using heterogeneous languages ​​according to an embodiment.

[0118] As shown in Figure 12, Disclosure The method for performing instruction tuning of a language model using heterogeneous languages ​​according to the embodiment may include a first language instruction dataset generation stage (S100), a second language instruction dataset generation stage (S200), a cross-language instruction generation stage (S300), and an instruction tuning stage (S400).

[0119] In the first language instruction dataset generation stage (S100), the instruction tuning system for language models using heterogeneous languages ​​can generate and classify datasets, including language understanding and language generation tasks, using a variety of open-source software such as AIHub1, Korpora2, Github, Huggingface, KLUE3, Korquad4, and ETRI5.

[0120] Furthermore, in the first language instruction dataset generation stage (S100), the instruction tuning system for the language model using heterogeneous languages ​​can construct clusters of datasets using heuristic rules.

[0121] Specifically, in the first language instruction dataset generation stage (S100), the instruction tuning system for a language model using heterogeneous languages ​​can generate the first instruction tuning dataset so that it is classified into a total of 17 task clusters, including 17 NLG (Natural Language Generation) datasets 1110 and 34 NLU (Natural Language Understanding) datasets 1120, as shown in Figure 5.

[0122] Furthermore, during the first language instruction dataset generation stage (S100), the instruction tuning system for a language model using heterogeneous languages ​​can set basic instructions for the first instruction tuning dataset. In this case, the basic instructions are set using the same first language as the first instruction tuning dataset.

[0123] Also , different The instruction tuning system for language models using species languages ​​utilizes all or part of the data labels included in the first instruction tuning dataset, or adds new data labels. For the first instruction tuning dataset Generate basic instructions.

[0124] In the second language instruction dataset generation stage (S200), the instruction tuning system for a language model using heterogeneous languages ​​can generate and configure the second instruction tuning dataset and basic instructions using the P3 data.

[0125] In the cross-language instruction generation stage (S300), the instruction tuning system for a language model using heterologous languages ​​generates and sets N cross-language instructions (where N is a natural number greater than or equal to 1) per task included in the first instruction tuning dataset and the second instruction tuning dataset, thereby increasing the effectiveness of instruction tuning using two languages.

[0126] In this case, N may be set to have a value of 3, but Disclosure The embodiments are not limited to these.

[0127] Specifically, in the cross-language instruction generation stage (S300), the instruction tuning system for a language model using heterogeneous languages ​​generates a first instruction created in the first language and a second instruction created in the second language. At this time, the first and second instructions are set to have the same format (data labels). Furthermore, the first instruction is applied to the second instruction tuning dataset, and the second instruction is applied to the first instruction tuning dataset.

[0128] In other words, the dataset and instructions are configured so that the language of the instruction tuning dataset and the language of the instructions intersect.

[0129] A language model instruction tuning system using heterogeneous languages ​​can generate the first instruction by machine translating the instructions for the second instruction tuning dataset into the first language, and generate the second instruction by machine translating the instructions for the first instruction tuning dataset into the second language.

[0130] Furthermore, the instruction tuning system for language models using heterogeneous languages ​​can generate cross-language instructions by preferentially selecting data labels that are commonly included in the basic instructions for the first instruction tuning dataset and the data labels included in the second instruction tuning dataset.

[0131] Furthermore, the instruction tuning system for language models using heterogeneous languages ​​can generate cross-language instructions by setting weight values ​​for each data label based on its frequency of use, from the basic instructions for the first instruction tuning dataset and the data labels included in the second instruction tuning dataset, and adding data labels whose weight values ​​are equal to or greater than the baseline value.

[0132] Furthermore, in the cross-language instruction generation stage (S300), the instruction tuning system for the language model using different languages ​​adds, deletes, and modifies the data labels of the first instruction and the labels of the second instruction to make the format of the first instruction and the format of the second instruction identical.

[0133] In the instruction tuning phase (S400), the instruction tuning system for a language model using heterogeneous languages ​​applies cross-language instructions to the first instruction tuning dataset and / or the second instruction tuning dataset to train (tune) the language model (LLM, mT0).

[0134] Furthermore, as shown in Figure 13, Disclosure The method for performing instruction tuning of a language model using heterogeneous languages ​​according to the embodiment may include a first language instruction dataset generation stage (S100), a second language instruction dataset generation stage (S200), a cross-language instruction generation stage (S300), and an inference stage (S500).

[0135] In the inference stage (S500), the language model is trained (tuned) by applying the basic instructions from the first instruction tuning dataset and the second instruction tuning dataset. Then, in the inference stage of the trained model on the input data, cross-language instructions are applied to the first instruction tuning dataset and / or the second instruction tuning dataset to perform inference.

[0136] Also, as shown in Figure 14, Disclosure The method for performing instruction tuning of a language model using heterogeneous languages ​​according to the embodiment may include a first language instruction dataset generation stage (S100), a second language instruction dataset generation stage (S200), a cross-language instruction generation stage (S300), an instruction tuning stage (S400), and an inference stage (S500).

[0137] The method for performing instruction tuning of a language model using heterologous languages, as shown in Figure 14, involves applying cross-language instructions to the first instruction tuning dataset and / or the second instruction tuning dataset in both the instruction tuning (learning) stage (S400) and the inference stage (S500).

[0138] Therefore, we not only apply cross-language instructions to the first instruction tuning dataset and / or the second instruction tuning dataset to train (tune) the language model, but we also apply cross-language instructions to the first instruction tuning dataset and / or the second instruction tuning dataset to perform inference.

[0139] The books described above Disclosure The embodiments thereof can be implemented in the form of program instructions that can be executed by various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc., individually or in combination. The program instructions recorded on the computer-readable recording medium are as follows: Disclosure These may be specifically designed and configured for this purpose, or they may be publicly known and available to those skilled in the field of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include not only machine code produced by compilers, but also high-level language code that can be executed by a computer using an interpreter, etc. Hardware devices are... Disclosure It may be modified into one or more software modules to perform the processing, and vice versa.

[0140] Book Disclosure The specific implementation described herein is one embodiment, and this may not be applicable in any way. DisclosureThis does not limit the scope. For the sake of brevity of the specification, descriptions of conventional electronic configurations, control systems, software, and other functional aspects of the system may be omitted. Also, the connections of lines or connecting members between components shown in the drawings are illustrative representations of functional and / or physical or circuit connections and may be replaced or represented as various additional functional, physical, or circuit connections in actual devices. Furthermore, unless specifically mentioned, such as "essential" or "important," this specification is not applicable. Disclosure It may not be a necessary component for its application.

[0141] Also, the book that explained Disclosure For a detailed explanation, see this Disclosure Although described with reference to preferred embodiments, a person skilled in the art or with ordinary knowledge in the art would understand the claims as described above. Disclosure Within the scope of the ideas and technologies of this book Disclosure It will be understood that this can be modified and changed in various ways. Therefore, this Disclosure The technical scope of the patent must be defined by the claims, rather than being limited to the detailed description in the specification.

[0142] The mode for carrying out the invention is the same as the best mode for carrying out the invention. [Industrial applicability]

[0143] This disclosure relates to a method and system for performing instruction tuning using heterogeneous languages, and is therefore applicable to the artificial intelligence industry and thus has industrial applicability.

Claims

1. A method for a computing system including memory and a processor to perform instruction tuning of a large language model (LLM) using heterogeneous languages, The first step is to set up a first instruction tuning dataset that includes one or more first tasks formed in the first language using metadata. The step of setting up a second instruction tuning dataset that includes one or more second tasks formed in a second language. The stage of generating cross-language instructions, and The step includes performing instruction tuning on the first instruction tuning dataset and the second instruction tuning dataset using the cross-language instruction, The step of generating the aforementioned cross-language instructions is: A step of generating one or more first instructions created in the first language, and The process includes the step of generating one or more second instructions written in the second language, The step of performing the aforementioned instruction tuning is, This includes the steps of performing instruction tuning on the second instruction tuning dataset using the first instruction, and performing instruction tuning on the first instruction tuning dataset using the second instruction, A method for performing instruction tuning using different languages, wherein the first instruction and the second instruction are set to have the same format.

2. The step of setting the first instruction tuning dataset is: The step includes generating N or more pre-set first tasks, The method for performing instruction tuning using heterogeneous languages ​​according to claim 1, wherein N is a natural number of 1 or more.

3. The step of generating the first instruction is: A method for performing heterogeneous language-based instruction tuning according to claim 2, comprising the step of generating 3N first instructions and applying 3 instructions per task.

4. The step of setting the first instruction tuning dataset is: A method for performing heterogeneous language-based instruction tuning according to claim 3, comprising the step of generating 34 NLU (natural Language Understanding) tasks and 17 NLG (Natural Language Generation) tasks based on data collected from one or more of the following: AIHub1, KorPora2, GitHub, Huggingface, KLUE3, Korquad4, ETRI5, Modu's Corpus, and KoBest.

5. The method for performing instruction tuning using heterogeneous languages ​​according to claim 4, wherein the language model utilizes a multi-lingual model.

6. A method for a computing system including memory and a processor to perform instruction tuning of a large language model (LLM) using heterogeneous languages, The first step is to set up a first instruction tuning dataset that includes one or more first tasks formed in the first language using metadata. The step of setting up a second instruction tuning dataset that includes one or more second tasks formed in a second language. The stage of generating cross-language instructions, and The step includes applying the cross-language instruction to the first instruction tuning dataset and the second instruction tuning dataset to perform inference on the data input to the language model, The step of generating the aforementioned cross-language instructions is: A step of generating one or more first instructions (instruction tuning) created in the first language, and The process includes the step of generating one or more second instructions written in the second language, The step of performing the aforementioned reasoning is, The process includes the steps of applying the first instruction to the second instruction tuning dataset and applying the second instruction to the first instruction tuning dataset to perform inference, A method for performing instruction tuning using different languages, wherein the first instruction and the second instruction are set to have the same format.

7. The step of setting the first instruction tuning dataset is: A method for performing instruction tuning using heterogeneous languages ​​according to claim 6, comprising the step of generating N or more pre-set first tasks, wherein N is a natural number of 1 or more.

8. The step of setting the first instruction is, A method for performing heterogeneous language-based instruction tuning according to claim 7, comprising the step of generating 3N first instructions and applying 3 instructions per task.

9. The step of setting the first instruction tuning dataset is: A method for performing heterogeneous language-based instruction tuning according to claim 8, comprising the step of generating 34 NLU (natural Language Understanding) tasks and 17 NLG (Natural Language Generation) tasks based on data collected from one or more of the following: AIHub1, KorPora2, GitHub, Huggingface, KLUE3, Korquad4, ETRI5, Modu's Corpus, and KoBest.

10. The method for performing instruction tuning using heterogeneous languages ​​according to claim 9, wherein the language model utilizes a multi-lingual model.

11. A method for a computing system including memory and a processor to perform instruction tuning of a large language model (LLM) using heterogeneous languages, The first step is to set up a first instruction tuning dataset that includes one or more first tasks formed in the first language using metadata. The step of setting up a second instruction tuning dataset that includes one or more second tasks formed in a second language. The stage of generating cross-language instructions, The steps include: performing instruction tuning on the first instruction tuning dataset and the second instruction tuning dataset using the cross-language instruction; and The step includes applying the cross-language instruction to the first instruction tuning dataset and the second instruction tuning dataset to perform inference on the data input to the language model, The step of generating the aforementioned cross-language instructions is: A step of generating one or more first instructions created in the first language, and The process includes the step of generating one or more second instructions written in the second language, The step of performing the aforementioned instruction tuning is, This includes the steps of performing instruction tuning on the second instruction tuning dataset using the first instruction, and performing instruction tuning on the first instruction tuning dataset using the second instruction, The step of performing the aforementioned reasoning is, A method for performing instruction tuning using heterogeneous languages, comprising the steps of applying the first instruction to the second instruction tuning dataset and applying the second instruction to the first instruction tuning dataset to perform inference, wherein the first instruction and the second instruction are set to have the same format.

12. The step of setting the first instruction tuning dataset is: A method for performing instruction tuning using heterogeneous languages ​​according to claim 11, comprising the step of generating N or more pre-set first tasks, wherein N is a natural number of 1 or more.

13. The step of generating the first instruction is: A method for performing heterogeneous language-based instruction tuning according to claim 12, comprising the step of generating 3N first instructions and applying 3 instructions per task.

14. The step of setting the first instruction tuning dataset is: A method for performing heterogeneous language instruction tuning according to claim 13, comprising the step of generating 34 NLU (natural Language Understanding) tasks and 17 NLG (Natural Language Generation) tasks based on data collected from one or more of the following: AIHub1, KorPora2, GitHub, Huggingface, KLUE3, Korquad4, ETRI5, Modu's Corpus, and KoBest.

15. The method for performing instruction tuning using heterogeneous languages ​​according to claim 14, wherein the language model utilizes a multi-lingual model.

16. At least one memory, and An execution system including at least one processor that reads at least one application stored in the memory and performs instruction tuning of a language model using heterogeneous languages, The aforementioned processor, A first instruction tuning dataset is set up that includes one or more first tasks formed in the first language using metadata. A second instruction tuning dataset is set up that includes one or more second tasks formed in a second language. A cross-language instruction is generated which includes one or more first instructions created in the first language and one or more second instructions created in the second language. Using the aforementioned cross-language instructions, instruction tuning is performed on the first instruction tuning dataset and the second instruction tuning dataset. The command includes a step of applying the cross-language instruction to the first instruction tuning dataset and the second instruction tuning dataset to perform inference on the data input to the language model, Using the first instruction, instruction tuning is performed on the second instruction tuning dataset, and using the second instruction, instruction tuning is performed on the first instruction tuning dataset. The first instruction is applied to the second instruction tuning dataset, and the second instruction is applied to the first instruction tuning dataset to perform inference. An instruction tuning execution system using heterogeneous languages, wherein the first instruction and the second instruction are configured to have the same format.

17. At least one memory, and A computing device comprising at least one processor that reads at least one application stored in the memory and performs instruction tuning of a language model using heterogeneous languages, The instruction words of the aforementioned processor are: The first step is to set up a first instruction tuning dataset that includes one or more first tasks formed in the first language using metadata. The step of setting up a second instruction tuning dataset that includes one or more second tasks formed in a second language. The stage of generating cross-language instructions, The steps include: performing instruction tuning on the first instruction tuning dataset and the second instruction tuning dataset using the cross-language instruction; and The command includes a step of applying the cross-language instruction to the first instruction tuning dataset and the second instruction tuning dataset to perform inference on the data input to the language model, The step of generating the aforementioned cross-language instructions is: A step of generating one or more first instructions created in the first language, and The process includes the step of generating one or more second instructions written in the second language, The step of performing the aforementioned instruction tuning is, This includes the steps of performing instruction tuning on the second instruction tuning dataset using the first instruction, and performing instruction tuning on the first instruction tuning dataset using the second instruction, The step of performing the aforementioned reasoning is, A computing device comprising the steps of applying the first instruction to the second instruction tuning dataset and applying the second instruction to the first instruction tuning dataset to perform inference, wherein the first instruction and the second instruction are configured to have the same format.

Images