Information processing device and information processing method

Abstract

An information processing device (1) comprises a first acquisition unit (20), a second acquisition unit (21), and a combining unit (22). The first acquisition unit (20) acquires information about a first large language model that has been fine-tuned specifically for risk countermeasures. The second acquisition unit (21) acquires information about a second large language model that has been fine-tuned specifically for a particular task. The combining unit (22) combines the first large language model and the second large language model on the basis of the information about the first large language model acquired by the first acquisition unit (20) and the information about the second large language model acquired by the second acquisition unit (21).

Description

Information Processing Apparatus and Information Processing Method

[0001] The present invention relates to an information processing apparatus and an information processing method.

[0002] In recent years, large language models (LLMs) have made remarkable progress in the field of natural language processing and shown performance close to human capabilities in various tasks. However, on the other hand, in large language models, issues related to risk factors such as hallucination and jailbreak have become more serious (see, for example, Non-Patent Documents 1 and 2).

[0003] For example, regarding hallucination, it has been pointed out that models published by major technology companies and the like have insufficient accuracy for fact-based questions. Also, regarding jailbreak attacks, in investigations by attack tests, many examples have been confirmed where a large number of large language models easily avoid restrictions.

[0004] Huang, Lei, et al. "A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions". ACM Transactions on Information Systems (2023), [online], [searched on December 29, 2024], Internet <URL: https: / / arxiv.org / pdf / 2311.05232> Yi, Sibo, et al. "Jailbreak Attacks and Defenses Against Large Language Models: A Survey", arXiv preprint arXiv:2407.04295 (2024), [online], [searched on December 29, 2024], Internet <URL: https: / / arxiv.org / abs / 2407.04295>

[0005] Fine-tuning is a common practice as a countermeasure against hallucination and jailbreaking. Fine-tuning is a learning process that involves further training on a pre-trained model to adapt it to specific tasks or datasets, and is expected to have some effect in suppressing hallucination and improving resistance to jailbreak attacks. However, fine-tuning has challenges such as requiring large amounts of data and computationally intensive resources.

[0006] The present invention has been made in view of the above, and aims to provide an information processing device and an information processing method that can easily take measures against the risks of large-scale language models.

[0007] To solve the above-mentioned problems and achieve the objective, the information processing device according to the present invention comprises a first acquisition unit, a second acquisition unit, and a synthesis unit. The first acquisition unit acquires information of a first large-scale language model that has been fine-tuned specifically for risk mitigation. The second acquisition unit acquires information of a second large-scale language model that has been fine-tuned specifically for a specific task. The synthesis unit synthesizes the first large-scale language model and the second large-scale language model based on the information of the first large-scale language model acquired by the first acquisition unit and the information of the second large-scale language model acquired by the second acquisition unit.

[0008] According to the present invention, it is possible to easily take measures against the risks of large-scale language models.

[0009] Figure 1 is a diagram illustrating the overview of processing in the information processing apparatus according to the embodiment. Figure 2 is a diagram showing an example of the configuration of the information processing apparatus according to the embodiment. Figure 3 is a diagram illustrating further processing in the information processing apparatus according to the embodiment. Figure 4 is a flowchart showing an example of information processing by the processing unit of the information processing apparatus according to the embodiment. Figure 5 is a diagram showing an example of the results of a simple experiment to evaluate a new model obtained using the processing method according to the embodiment. Figure 6 is a diagram showing an example of a computer that executes an information processing program.

[0010] Hereinafter, one embodiment of the present invention will be described in detail with reference to the drawings. However, the present invention is not limited to this embodiment. Furthermore, in the drawings, the same parts are denoted by the same reference numerals.

[0011] [Overview of Processing in the Information Processing Device] Figure 1 is a diagram illustrating the overview of processing in the information processing device according to the embodiment. As shown in Figure 1, the information processing device 1 according to the embodiment generates a third large-scale language model M3 by combining a first large-scale language model M1 and a second large-scale language model M2.

[0012] The first large-scale language model M1 is a large-scale language model fine-tuned specifically for risk mitigation, and is adjusted to be resilient to risks such as hallucination and jailbreaking. Risk mitigation includes, for example, measures against jailbreaking and measures against hallucination, at least one of these. The first large-scale language model M1 is, for example, a large-scale language model obtained by fine-tuning a reference large-scale language model MM, which is a standard large-scale language model, using training data for risk mitigation.

[0013] The second large-scale language model M2 is a large-scale language model that has been fine-tuned specifically for a particular task. The second large-scale language model M2 is, for example, a large-scale language model obtained by fine-tuning a reference large-scale language model using training data for the specific task. The specific task is different from the risk mitigation task in the first large-scale language model M1 described above.

[0014] The information processing device 1, for example, uses TIES-MERGEING (TRIM, ELECT SIGN, MERGE), a model synthesis technique, to synthesize a first large-scale language model M1 and a second large-scale language model M2 to generate a third large-scale language model M3. This allows the information processing device 1 to generate a third large-scale language model M3 that is specialized for a specific task while also incorporating risk mitigation measures.

[0015] Model synthesis using TIES is performed by generating the parameters of a third large-scale language model M3 using the parameters of a first large-scale language model M1 and the parameters of a second large-scale language model M2.

[0016] When fine-tuning a second large-scale language model M2 using risk mitigation training data without using the first large-scale language model M1, a large amount of data and computationally intensive resources are required, posing a cost challenge. On the other hand, model synthesis performed by the information processing device 1 involves the synthesis of information (e.g., parameters) from multiple models, thus significantly reducing costs. Therefore, the information processing device 1 can easily implement risk mitigation measures for large-scale language models.

[0017] [Configuration and Processing of Information Processing Device 1] Next, the configuration and processing of the information processing device 1 will be described in detail with reference to Figure 2, etc. Figure 2 is a diagram showing an example of the configuration of the information processing device 1 according to the embodiment. As shown in Figure 2, the information processing device 1 comprises a communication unit 10, a storage unit 11, and a processing unit 12.

[0018] The communication unit 10 is connected to a communication network (not shown), and can send and receive information to and from external devices (not shown), such as a server or terminal device, via this communication network. The communication network (not shown) is, for example, a LAN (Local Area Network) or an internet or other telecommunications line, but is not limited to these examples.

[0019] The storage unit 11 stores data and programs used for various processes performed by the processing unit 12. For example, the storage unit 11 stores information such as the first large-scale language model M1 and the second large-scale language model M2 acquired by the processing unit 12 so that it can be read by the processing unit 12.

[0020] The processing unit 12 comprises a first acquisition unit 20, a second acquisition unit 21, a synthesis unit 22, and an output unit 23. The first acquisition unit 20 acquires information of a first large-scale language model M1 that has been fine-tuned specifically for risk mitigation. The information of the first large-scale language model M1 acquired by the first acquisition unit 20 is part or all of the first large-scale language model M1, for example, the parameters of the first large-scale language model M1.

[0021] The first acquisition unit 20 acquires information about the first large-scale language model M1 from an external device, such as a server or terminal device, via the communication unit 10, for example. As described above, the first large-scale language model M1 is a large-scale language model obtained by fine-tuning a reference large-scale language model using risk mitigation training data. The reference large-scale language model is a transformer-based model, but is not limited to this example.

[0022] Risk mitigation training data includes, for example, at least one of jailbreak mitigation training data and hallucination mitigation training data. Jailbreaking is an attack method that intentionally causes a large-scale language model to output answers that circumvent the restrictions set on the model for ethical, legal, or security reasons. Hallucination is a phenomenon in which a large-scale language model incorrectly outputs patterns or information that do not exist.

[0023] The training data for jailbreak prevention includes, for example, jailbreak attack prompts and refusal responses for each prompt. Jailbreak attack prompts include, but are not limited to, prompts that elicit answers in a hypothetical context or assumption, prompts that extract information by dividing it into multiple questions, and prompts that have the model play a specific role and elicit responses based on that role. Refusal responses include, but are not limited to, the string "I'm sorry, I cannot answer that request."

[0024] The training data for hallucination correction includes, for example, data containing multiple pairs of hallucination content and its modified content as identified in a standard large-scale language model, but is not limited to such examples.

[0025] The first large-scale language model M1 is, for example, a large-scale language model finely tuned specifically for jailbreak countermeasures as a risk mitigation measure, a large-scale language model finely tuned specifically for hallucination countermeasures as a risk mitigation measure, or a large-scale language model finely tuned specifically for both jailbreak countermeasures and hallucination countermeasures as a risk mitigation measure.

[0026] The first large-scale language model M1 may be a large-scale language model that has been overfitted by fine-tuning for risk management. In other words, the first large-scale language model M1 may be a large-scale language model whose performance is lower than that of the baseline large-scale language model in tasks other than risk management due to fine-tuning for risk management.

[0027] The second acquisition unit 21 acquires information about the second large-scale language model M2, which has been fine-tuned specifically for a particular task. The information about the second large-scale language model M2 acquired by the second acquisition unit 21 is part or all of the second large-scale language model M2, for example, the parameters of the second large-scale language model M2.

[0028] The second acquisition unit 21 acquires information about the second large-scale language model M2 from an external device such as a server or terminal device via, for example, the communication unit 10. The specific task is a task different from the risk management task in the first large-scale language model M1, as described above, and is, but is not limited to, text summarization, translation, sentiment analysis, topic classification, database search, customer support, and personal assistant. The second large-scale language model M2 is a large-scale language model obtained by fine-tuning a reference large-scale language model using training data for the specific task.

[0029] The synthesis unit 22 synthesizes the first large-scale language model M1 and the second large-scale language model M2 based on the information of the first large-scale language model M1 acquired by the first acquisition unit 20 and the information of the second large-scale language model M2 acquired by the second acquisition unit 21 to generate a third large-scale language model M3. Hereinafter, the first large-scale language model M1 may be simply referred to as model M1, and the second large-scale language model M2 may be simply referred to as model M2.

[0030] The combination of model M1 and model M2 is performed, for example, by combining parameters. Parameter combination may include, but is not limited to, the TIES method, simple averaging, or Fisher weighted averaging methods mentioned above.

[0031] Here, we assume that the synthesis unit 22 performs parameter synthesis using TIES. The synthesis unit 22 calculates a task vector for each parameter based on the difference between the parameters before and after fine tuning. The parameters before fine tuning are the parameters of the standard large-scale language model. The parameters after fine tuning are the parameters of model M1 or model M2.

[0032] The synthesis unit 22 identifies the task vectors that have a significant impact on each of the models M1 and M2. If the signs of the task vectors with a significant impact do not match between models M1 and M2, the signs are unified, and then the task vectors with a significant impact are averaged between models M1 and M2. The synthesis unit 22 performs parameter synthesis by adding the averaged task vectors to the parameters before fine tuning for each task vector with a significant impact. The synthesis unit 22 generates a third large-scale language model M3 by applying the synthesized parameters to a base large-scale language model.

[0033] Thus, the synthesis of models M1 and M2 by the synthesis unit 22 is a synthesis process of the parameters of model M1 and the parameters of model M2, and therefore does not require large amounts of data or computational resources that require high computational costs, as is the case with fine tuning. As a result, costs can be significantly reduced, and measures against the risks of large-scale language models can be easily implemented.

[0034] The output unit 23 outputs information about the third large-scale language model M3 obtained by synthesis by the synthesis unit 22. For example, the output unit 23 outputs information about the third large-scale language model M3 by transmitting it to a terminal device (not shown) via the communication unit 10. The information about the first large-scale language model M1 and the second large-scale language model M2 acquired by the processing unit 12, and the information about the third large-scale language model M3 generated by the processing unit 12 are stored in the storage unit 11.

[0035] The processing unit 12 of the information processing device 1 can create a new large-scale language model with further enhanced risk countermeasures for the third large-scale language model M3 using the model synthesis described above. Figure 3 is a diagram illustrating further processing in the information processing device 1 according to the embodiment. As shown in Figure 3, the processing unit 12 synthesizes the third large-scale language model M3 and the fourth large-scale language model M4, which is specialized for risk countermeasures, to generate a fifth large-scale language model M5. The generation process of the fifth large-scale language model M5 will be described in detail below.

[0036] The first acquisition unit 20 acquires information about a fourth large-scale language model M4 that is different from the first large-scale language model M1. The information about the fourth large-scale language model M4 acquired by the first acquisition unit 20 is part or all of the fourth large-scale language model M4, for example, the parameters of the fourth large-scale language model M4.

[0037] The fourth large-scale language model M4 is a large-scale language model specifically designed for risk management, which has been fine-tuned against the baseline large-scale language model using new risk management training data. The new risk management training data is partially or entirely different from the risk management training data used to fine-tune the first large-scale language model M1.

[0038] For example, the new risk mitigation training data includes training data for mitigating risks newly discovered after the generation of the first large-scale language model M1, and training data containing prompts that have come to be used in risk assessment and other applications after the generation of the first large-scale language model M1.

[0039] The second acquisition unit 21 acquires information about the third large-scale language model M3 from the storage unit 11 and the like. The information about the third large-scale language model M3 acquired by the second acquisition unit 21 is part or all of the third large-scale language model M3, for example, the parameters of the third large-scale language model M3.

[0040] The synthesis unit 22 synthesizes the third large-scale language model M3 and the fourth large-scale language model M4 based on the information of the fourth large-scale language model M4 acquired by the first acquisition unit 20 and the information of the third large-scale language model M3 acquired by the second acquisition unit 21. The synthesis of the third large-scale language model M3 and the fourth large-scale language model M4 generates the fifth large-scale language model M5.

[0041] As a result, the information processing device 1 can easily take measures against risks in the large-scale language model, even if there are newly discovered risks or prompts that are newly used in risk assessment after the generation of the first large-scale language model M1.

[0042] Furthermore, the synthesis unit 22 can also generate a fifth large-scale language model M5 by synthesizing the first large-scale language model M1, the second large-scale language model M2, and the fourth large-scale language model M4.

[0043] In the above example, the model synthesis by the synthesis unit 22 is performed by synthesizing parameters, but it is not limited to such an example. For example, the model synthesis by the synthesis unit 22 may be layer synthesis. For example, the synthesis unit 22 can also use a model configured to ensemble the outputs of each layer of a plurality of models to be synthesized as the synthesized model. Further, the model synthesis by the synthesis unit 22 may be a partial integration of a plurality of models to be synthesized.

[0044] Also, in the above example, the first large language model M1 and the second large language model M2 are described as having the same model structure, but it is not limited to such an example. For example, when the model synthesis by the synthesis unit 22 is layer synthesis, the plurality of models to be synthesized may have different model structures from each other. Also, in the above example, an example of synthesizing two models is described, but the number of models synthesized by the synthesis unit 22 may be three or more.

[0045] Further, the processing unit 12 of the information processing apparatus 1 may include a learning processing unit that causes an external information processing apparatus to execute fine-tuning. In this case, the learning processing unit can use the risk mitigation learning data to cause an external information processing apparatus to execute fine-tuning and generate the second large language model M2 or the fourth large language model M4.

[0046] [Example of Processing Procedure] Next, an example of the processing procedure executed by the information processing apparatus 1 will be described using FIG. 4. FIG. 4 is a flowchart showing an example of information processing by the processing unit 12 of the information processing apparatus 1 according to the embodiment. As shown in FIG. 4, the processing unit 12 of the information processing apparatus 1 acquires information on a plurality of large language models (step S10). For example, the processing unit 12 acquires information on the first large language model M1 and information on the second large language model M2.

[0047] Subsequently, the processing unit 12 synthesizes the plurality of large language models for which information was acquired in step S10 (step S11). For example, when the processing unit 12 has acquired information on the first large language model M1 and information on the second large language model M2 in step S10, the processing unit 12 synthesizes the first large language model M1 and the second large language model M2 to generate a third large language model M3.

[0048] Subsequently, the processing unit 12 outputs information on the synthesized large language model (step S12) and ends the processing of FIG. 4. For example, when the processing unit 12 has generated the third large language model M3 in step S11, the processing unit 12 outputs information on the third large language model M3 and ends the processing of FIG. 4.

[0049] [Experiment] A simple experiment was conducted to evaluate the new model obtained using the processing method according to the embodiment. In the experiment, a T5-based model was used as the reference large language model, and the language used was English. Also, as the first large language model M1, "fine_tuned_t5_jailbreak" was generated, and as the second large language model M2, "fine_tuned_t5_winograd_wsc" was generated.

[0050] "fine_tuned_t5_jailbreak" was generated by fine-tuning a T5-based model using, as risk mitigation learning data, a dataset combining each input of the "Jailbreak dataset" with a specific response. The "Jailbreak dataset" is a dataset used in the benchmark of TrustLLM. The specific response was "I'm sorry, but I cannot assist with that request."

[0051] Furthermore, "fine_tuned_t5_winograd_wsc" was generated by fine-tuning a T5-based model using "ErnestSDavis / winograd_wsc" as training data for a specific task. "ErnestSDavis / winograd_wsc" is a dataset related to the Winograd Schema Challenge (WSC), and is, for example, a dataset for tasks that require accurately identifying what a pronoun refers to.

[0052] Then, using TIES, "fine_tuned_t5_jailbreak" and "fine_tuned_t5_winograd_wsc" were combined to generate a new large-scale language model, "ties_merged_t5". "ties_merged_t5" is the third large-scale language model M3 mentioned above.

[0053] Figure 5 shows an example of the results of a simple experiment to evaluate a new model obtained using the processing method according to the embodiment. The experimental results shown in Figure 5 evaluate the percentage of "ties_merged_t5" that responds to a specific message ("I'm sorry, but I cannot assist with that request.") after inputting 420 jailbreak prompts. As shown in Figure 5, the second large-scale language model M2, "fine_tuned_t5_winograd_wsc", has a correct response rate of 0%, while the third large-scale language model M3, "ties_merged_t5", has a correct response rate of 69.52%, indicating that the measures against jailbreaking have been significantly strengthened.

[0054] [System Configuration, etc.] Furthermore, the components of each part shown in the diagram are functional concepts and do not necessarily have to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown in the diagram, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions. In addition, all or any part of the processing functions performed by each device can be realized by a CPU and the program executed on that CPU, or by hardware using wired logic.

[0055] Furthermore, among the processes described in the embodiments above, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[0056] [Program] A program can also be created that describes the processing performed by the information processing device 1 according to the above embodiment in a language that can be executed by a computer. In one embodiment, the information processing device 1 can be implemented by installing an information processing program that performs the above information processing as packaged software or online software on a desired computer. For example, by having the information processing device execute the above information processing program, the information processing device can be made to function as the information processing device 1. The information processing device referred to here is a general-purpose computer, but is not limited to this example, and may be a general-purpose server (for example, an on-premise server or a cloud server), or any other computer.

[0057] Figure 6 shows an example of a computer that executes an information processing program. Computer 1000 includes, for example, memory 1010, CPU 1020, hard disk drive interface 1030, disk drive interface 1040, serial port interface 1050, video adapter 1060, and network interface 1070. These components are connected by a bus 1080.

[0058] Memory 1010 includes ROM (Read Only Memory) 1011 and RAM 1012. ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1031. The disk drive interface 1040 is connected to the disk drive 1041. A removable storage medium, such as a magnetic disk or optical disk, is inserted into the disk drive 1041. A serial port interface 1050 is connected to, for example, a mouse 1051 and a keyboard 1052. A video adapter 1060 is connected to, for example, a display 1061.

[0059] Here, the hard disk drive 1031 stores, for example, the OS 1091, the application program 1092, the program module 1093, and the program data 1094. The information described in the above embodiment is stored, for example, in the hard disk drive 1031 or the memory 1010.

[0060] Furthermore, the information processing program is stored in the hard disk drive 1031 as a program module 1093 containing instructions to be executed by the computer 1000, for example. Specifically, the program module 1093 containing instructions for each process executed by the information processing device 1 described in the above embodiment is stored in the hard disk drive 1031.

[0061] Furthermore, the data used for information processing by the information processing program is stored as program data 1094, for example, in the hard disk drive 1031. The CPU 1020 then reads the program module 1093 and program data 1094 stored in the hard disk drive 1031 into the RAM 1012 as needed and executes the procedures described above.

[0062] Furthermore, the program module 1093 and program data 1094 related to the information processing program are not limited to being stored in the hard disk drive 1031; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1041 or the like. Alternatively, the program module 1093 and program data 1094 related to the information processing program may be stored in another computer connected via a network such as a LAN or WAN and read by the CPU 1020 via a network interface 1070.

[0063] [Effects of the Embodiment] As described above, the information processing device 1 comprises a first acquisition unit 20, a second acquisition unit 21, and a synthesis unit 22. The first acquisition unit 20 acquires information of a first large-scale language model M1 that has been fine-tuned specifically for risk countermeasures. The second acquisition unit 21 acquires information of a second large-scale language model M2 that has been fine-tuned specifically for a specific task. The synthesis unit 22 synthesizes the first large-scale language model M1 and the second large-scale language model M2 based on the information of the first large-scale language model M1 acquired by the first acquisition unit 20 and the information of the second large-scale language model M2 acquired by the second acquisition unit 21. As a result, the information processing device 1 can easily take measures against risks associated with large-scale language models.

[0064] Furthermore, the first large-scale language model M1 is a large-scale language model that has been fine-tuned specifically for jailbreak countermeasures as a risk mitigation measure. This allows the information processing device 1 to easily implement countermeasures against jailbreak attacks in the large-scale language model.

[0065] Furthermore, the second acquisition unit 21 acquires information on the third large-scale language model M3, which is obtained by combining the first large-scale language model M1 and the second large-scale language model M2. The first acquisition unit 20 acquires information on the fourth large-scale language model M4, which is fine-tuned specifically for risk mitigation using training data different from the training data used for fine-tuning the first large-scale language model M1. The synthesis unit 22 synthesizes the third large-scale language model M3 and the fourth large-scale language model M4 based on the information on the third large-scale language model M3 acquired by the second acquisition unit 21 and the information on the fourth large-scale language model M4 acquired by the first acquisition unit 20. As a result, the information processing device 1 can easily take measures against risks to the large-scale language model, even if, for example, a new risk is discovered after the third large-scale language model M3 has been generated.

[0066] Although embodiments applying the invention made by the present inventors have been described above, the present invention is not limited by the descriptions and drawings that constitute part of the disclosure of the present invention in this embodiment. That is, all other embodiments, examples, and operational techniques made by those skilled in the art based on this embodiment are included in the scope of the present invention.

[0067] 1 Information Processing Device 10 Communication Unit 11 Storage Unit 12 Processing Unit 20 First Acquisition Unit 21 Second Acquisition Unit 22 Synthesis Unit 23 Output Unit M1 First Large-Scale Language Model M2 Second Large-Scale Language Model M3 Third Large-Scale Language Model M4 Fourth Large-Scale Language Model M5 Fifth Large-Scale Language Model

Claims

1. An information processing device comprising: a first acquisition unit for acquiring information on a first large-scale language model finely tuned specifically for risk mitigation; a second acquisition unit for acquiring information on a second large-scale language model finely tuned specifically for a particular task; and a synthesis unit for synthesizing the first large-scale language model and the second large-scale language model based on the information on the first large-scale language model acquired by the first acquisition unit and the information on the second large-scale language model acquired by the second acquisition unit.

2. The information processing device according to claim 1, characterized in that the first large-scale language model is a large-scale language model that has been fine-tuned specifically for jailbreak countermeasures as a risk countermeasure.

3. The information processing apparatus according to claim 1 or 2, characterized in that the second acquisition unit acquires information of a third large-scale language model obtained by combining the first large-scale language model and the second large-scale language model; the first acquisition unit acquires information of a fourth large-scale language model that has been fine-tuned specifically for risk countermeasures using training data different from the training data used for fine-tuning the first large-scale language model; and the synthesis unit synthesizes the third large-scale language model and the fourth large-scale language model based on the information of the third large-scale language model acquired by the second acquisition unit and the information of the fourth large-scale language model acquired by the first acquisition unit.

4. An information processing method performed by a computer, comprising: a first acquisition step of acquiring information of a first large-scale language model fine-tuned specifically for risk mitigation; a second acquisition step of acquiring information of a second large-scale language model fine-tuned specifically for a specific task; and a synthesis step of synthesizing the first large-scale language model and the second large-scale language model based on the information of the first large-scale language model acquired by the first acquisition step and the information of the second large-scale language model acquired by the second acquisition step.

Images