Information processing device, information processing system, information processing method, and program

Abstract

To enable the generation of the desired model even when vocabulary modifications are made to the base model and further pre-training is performed. [Solution] The information processing device includes an integrated model generation unit that generates an integrated model by integrating a first language model generated by modifying the vocabulary of the input / output layer of a base language model, a second language model generated by performing continuous pre-training on the first language model, and a fourth language model generated by applying the same vocabulary modification used in the generation of the first language model to the input / output layer of a third language model generated by performing instruction learning on the base language model.

Description

[Technical Field]

[0001] The present invention relates to an information processing apparatus, an information processing system, an information processing method, and a program. [Background Art]

[0002] The following Non-Patent Document 1 discloses a technique for generating a target LM model, which is a CP model obtained by performing continued pre-training on a base model and imparted with instruction following performance, by merging a chat vector (a difference between a base model and a model obtained by performing instruction learning on the base model and acquiring instruction following performance) into the CP model obtained by performing continued pre-training on the base model. [Summary of the Invention] [Problem to be Solved by the Invention]

[0003] However, with the technique of Non-Patent Document 1, when vocabulary modification is performed on a base model and then continued pre-training is further performed, even if the chat vector is merged into the base model after the continued pre-training, a model imparted with instruction following performance cannot be generated.

[0004] In order to solve the above-mentioned problem of the conventional technology, an object of the present invention is to enable generation of a desired model even when vocabulary modification is performed on a base model and continued pre-training is further performed. [Means for Solving the Problem]

[0005] To solve the above-mentioned problems, an information processing device according to one embodiment includes an integrated model generation unit that generates an integrated model by integrating a first language model generated by modifying the vocabulary of the input / output layer of a base language model, a second language model generated by performing continuous pre-training on the first language model, and a fourth language model generated by applying the same vocabulary modification as used in the generation of the first language model to the input / output layer of a third language model generated by performing instruction learning on the base language model. [Effects of the Invention]

[0006] According to one embodiment of the information processing device, even when vocabulary changes are made to the base model and further pre-training is performed, the desired model can be generated. [Brief explanation of the drawing]

[0007] [Figure 1] A diagram showing an example of the system configuration of an information processing system according to one embodiment. [Figure 2] Hardware configuration diagram of a server and PC according to one embodiment. [Figure 3] Functional block diagram of a server and PC according to one embodiment. [Figure 4] Conceptual diagram showing the integrated model generation process by a server according to one embodiment. [Figure 5] A diagram showing an example of vocabulary modification by a server according to one embodiment. [Figure 6] Conceptual diagram of server-based model integration according to one implementation. [Figure 7] Conceptual diagram of model integration using chat vectors by a server according to one embodiment. [Figure 8] Conceptual diagram showing a first modified example of the integrated model generation process by a server according to one embodiment. [Figure 9] Conceptual diagram showing a second modified example of the integrated model generation process by the server according to one embodiment. [Figure 10]This diagram illustrates the relationship between each language model and the embedding layer described in this embodiment. [Figure 11] This figure shows an example of the changes in the embedding layer in each language model described in this embodiment. [Figure 12] Figure showing an example of the matrix of the embedding layer of a merged model integrated model generated by a server according to one embodiment. [Figure 13] This figure shows another example of the matrix of the embedding layer of the merge model integration model generated by the server according to one embodiment. [Figure 14] Conceptual diagram showing a third modified example of the integrated model generation process by the server according to one embodiment. [Figure 15] This figure shows an example of a UI screen that accepts input for a dataset displayed by a PC according to one embodiment. [Figure 16] Functional block diagram of a server and PC according to one embodiment. [Figure 17] A flowchart illustrating an example of the procedure for AI chat generation by a server and PC according to one embodiment. [Figure 18] This figure shows an example of the generated AI chat screen displayed on a PC according to one embodiment. [Figure 19] This figure shows an example of the generated AI chat screen displayed on a PC according to one embodiment. [Modes for carrying out the invention]

[0008] One embodiment of the present invention will be described below with reference to the drawings.

[0009] (An example of the system configuration of information processing system 10) Figure 1 shows an example of the system configuration of an information processing system 10 according to one embodiment.

[0010] The information processing system 10 shown in FIG. 1 is an example of an "information processing system". The information processing system 10 includes a server 12 and a PC 14. The server 12 and the PC 14 are communicatively connected to each other via a communication network 16. As the communication network 16, for example, a LAN, the Internet, or the like is used.

[0011] The PC 14 is an example of a terminal device. The PC 14 includes an input device for inputting a character string. Examples of the input device include a keyboard, a touch panel, and the like. The PC 14 can transmit a character string input by a user using the input device to the server 12 via the communication network 16. Note that the PC 14 may have a voice recognition function that recognizes voice input by the user from a microphone, and in this case, the character string recognized by the voice recognition function may be transmitted to the server 12.

[0012] The server 12 is an example of an "information processing apparatus". The server 12 receives a character string transmitted from the PC 14 via the communication network 16, and divides the character string into a plurality of vocabularies by character string division processing.

[0013] Specifically, the server 12 includes a tokenizer. The tokenizer has a vocabulary group. The vocabulary group includes a plurality of vocabularies each associated with an identifier. The vocabulary group is realized, for example, by a JSON file. The tokenizer performs character string division processing (i.e., tokenization processing) using a known technique on the character string acquired from the PC 14 based on the vocabulary group, thereby dividing the character string acquired from the PC 14 into a plurality of vocabularies (i.e., a plurality of morphemes). Then, the tokenizer outputs identifiers corresponding to the divided plurality of vocabularies to an LLM (Large Language Model).

[0014] The PC 14 can receive the plurality of vocabularies transmitted from the server 12 (i.e., the plurality of vocabularies obtained by the character string division processing) via the communication network 16, and perform various processes using the plurality of vocabularies.

[0015] In the example shown in Figure 1, one terminal device (PC 14) is connected to the server 12 via the communication network 16. However, in reality, the information processing system 10 allows multiple terminal devices to connect to the server 12 via the communication network 16, and each of these terminal devices can send a string to the server 12.

[0016] Furthermore, while the example shown in Figure 1 uses a PC14 as an example of a terminal device for inputting text, the terminal device is not limited to this. Any other device capable of inputting and transmitting text (for example, a smartphone, tablet, electronic whiteboard, inkjet printer, MFP, video conferencing terminal, projector, 360-degree camera, etc.) may be used.

[0017] (Hardware configuration of Server 12 and PC 14) Figure 2 is a hardware configuration diagram of server 12 and PC 14 according to one embodiment. Here, we will describe the hardware configuration common to server 12 and PC 14.

[0018] As shown in Figure 2, the server 12 and PC 14 are built by a computer and include a CPU 501, ROM 502, RAM 503, HD 504, HDD (Hard Disk Drive) controller 505, display 506, external device connection I / F (Interface) 508, network I / F 509, data bus 510, keyboard 511, pointing device 512, DVD-RW (Digital Versatile Disk Rewritable) drive 514, and media I / F 516.

[0019] The CPU 501 controls the operation of the entire server 12 or PC 14. The ROM 502 stores programs used to drive the CPU 501, such as IPLs. The RAM 503 is used as the work area for the CPU 501. The HD 504 stores various data, such as programs. The HDD controller 505 controls the reading or writing of various data to the HD 504 according to the control of the CPU 501. The display 506 displays various information such as cursors, menus, windows, characters, or images. The external device connection I / F 508 is an interface for connecting various external devices. In this case, external devices include, for example, USB (Universal Serial Bus) memory and printers. The network I / F 509 is an interface for data communication using the communication network 16. The data bus 510 is an address bus and data bus, etc., for electrically connecting each component such as the CPU 501 shown in Figure 2.

[0020] The keyboard 511 is a type of input means equipped with multiple keys for inputting characters, numbers, and various instructions. The pointing device 512 is a type of input means for selecting and executing various instructions, selecting processing targets, and moving the cursor. The DVD-RW drive 514 controls the reading or writing of various data to the DVD-RW 513, which is an example of a removable recording medium. Note that it is not limited to DVD-RW, but may also be DVD-R, etc. The media interface 516 controls the reading or writing (storage) of data to the recording medium 515, such as flash memory.

[0021] (Functional configuration of Server 12 and PC 14) Figure 3 is a functional block diagram of server 12 and PC 14 according to one embodiment. Figure 3 shows several functions related to the generation of integrated models, among the multiple functions that server 12 and PC 14 have. Note that each model described below refers to an LLM (Large Language Model).

[0022] As shown in Figure 3, the server 12 comprises a base model acquisition unit 101, a first language model generation unit 102, a second language model generation unit 103, a third language model acquisition unit 104, a fourth language model generation unit 105, and an integrated model generation unit 106.

[0023] The base model acquisition unit 101 acquires a base model, which is the underlying language model. For example, the base model acquisition unit 101 acquires an available open model (e.g., Meta-Llama-3) as the base model from an external source (e.g., the internet).

[0024] The first language model generation unit 102 generates a first language model by modifying the vocabulary of the input / output layer of the base model acquired by the base model acquisition unit 101. The input / output layers correspond to the embedding layer (hereinafter sometimes called the input layer or Embedding layer) and the lm_head layer (hereinafter sometimes called the output layer) in LLM. The embedding layer is the layer that converts identifiers corresponding to strings partitioned by the tokenizer into corresponding embedding vectors. The lm_head layer is the layer that outputs a vector sequence (predicted probability of identifiers corresponding to each string as the next token) based on processing by the hidden layer (hereinafter sometimes called the intermediate layer or Hidden layer) of LLM. Modifying the vocabulary of the input / output layers means changing the mapping between identifiers and vectors corresponding to strings in the input / output layers in accordance with changes to the tokenizer's vocabulary. Modifying the tokenizer's vocabulary includes vocabulary expansion (adding new mappings between vocabulary and identifiers), reduction (deleting mappings between vocabulary and identifiers), and replacement (changing a vocabulary mapped to one identifier to another). In other words, vocabulary changes in the input / output layer include vocabulary expansion (adding identifiers added in the tokenizer vocabulary change and the vectors corresponding to those identifiers), reduction (deleting identifiers removed in the tokenizer vocabulary change and the vectors corresponding to those identifiers), and replacement (changing the vectors that are associated with identifiers whose vocabulary has been changed in the tokenizer vocabulary change to other vectors).

[0025] The second language model generation unit 103 generates a second language model by performing continuous pre-training on the first language model generated by the first language model generation unit 102. Here, continuous pre-training means training a model that has already been pre-trained on a large dataset with a new, additional large dataset. For example, continuous pre-training is used for localizing an English model into Japanese, or for specializing it in the financial domain or the medical domain.

[0026] The third language model acquisition unit 104 acquires a third language model generated by performing instructional learning on a base model. For example, the third language model acquisition unit 104 acquires an available open model (e.g., Meta-Llama3-Instruct) from an external source (e.g., the internet) as the third language model. Here, instructional learning means training a pre-trained (or continuously pre-trained) model using a small amount of task-specific dataset. For example, instructional learning is used to enable classification, summarization, and chat-style responses.

[0027] The fourth language model generation unit 105 generates a fourth language model by applying vocabulary modifications to the input / output layers of the third language model generated by the third language model acquisition unit 104, similar to the vocabulary modifications performed by the first language model generation unit 102. Vocabulary modifications similar to those performed by the first language model generation unit 102 involve expanding, reducing, and replacing vocabulary in the same way as the vocabulary modifications performed by the first language model generation unit 102. Specifically, vocabulary reduction is performed by deleting the correspondence between identifiers and vectors that were reduced by the first language model generation unit 102, and vocabulary expansion and replacement are performed by associating vectors generated using the same calculation method as the first language model generation unit 102 with identifiers that represent new or existing vocabulary.

[0028] The integrated model generation unit 106 generates an integrated model by integrating the first language model generated by the first language model generation unit 102, the second language model generated by the second language model generation unit 103, and the fourth language model generated by the fourth language model generation unit 105.

[0029] Furthermore, as shown in Figure 3, the server 12 includes a communication unit 112 and a display control unit 111.

[0030] The communication unit 112 communicates with the PC 14. For example, the communication unit 112 receives user input from the PC 14. Also, for example, the communication unit 112 transmits screen information for displaying the UI screen to the PC 14 under the control of the display control unit 111.

[0031] The display control unit 111 controls the screen display on the PC 14. Specifically, the display control unit 111 transmits screen information for displaying the UI screen to the PC 14 via the communication unit 112, thereby causing the PC 14 to display the UI screen.

[0032] Furthermore, as shown in Figure 3, the PC 14 includes an input receiving unit 121, a communication unit 122, and a display unit 123.

[0033] The input receiving unit 121 receives input from the user on the UI screen displayed on the display unit 123.

[0034] The communication unit 122 communicates with the server 12. For example, the communication unit 122 transmits user input received by the input reception unit 121 to the server 12. Also, for example, the communication unit 122 receives screen information from the server 12 for displaying the UI screen.

[0035] The display unit 123 displays a UI screen based on the screen information received by the communication unit 122.

[0036] Each of the functions of the embodiments described above can be realized by one or more processing circuits. Hereinafter, "processing circuit" as used herein includes processors programmed to execute each function by software, such as processors implemented by electronic circuits, as well as devices such as ASICs (Application Specific Integrated Circuits), DSPs (digital signal processors), FPGAs (field programmable gate arrays), and conventional circuit modules designed to execute each of the functions described above.

[0037] (Integrated model generation process by Server 12) Figure 4 is a conceptual diagram showing the integrated model generation process by server 12 according to one embodiment.

[0038] In the integrated model generation process shown in Figure 4, the base model acquisition unit 101 first acquires the base model θbase.

[0039] Next, the first language model generation unit 102 generates a vocabulary replacement base model g(θbase) (first language model) by applying vocabulary modification g to the input / output layer of the base model θbase acquired by the base model acquisition unit 101.

[0040] Next, the second language model generation unit 103 generates a continuously pre-trained model θ'base (second language model) by performing continuous pre-training on the lexical substitution base model g(θbase) generated by the first language model generation unit 102.

[0041] Next, the third language model acquisition unit 104 acquires the instruction model θinst (third language model) which is generated by performing instruction learning on the base model θbase.

[0042] Next, the fourth language model generation unit 105 generates a vocabulary replacement instruction model g(θinst) (fourth language model) by applying the same vocabulary modification g as performed by the first language model generation unit 102 to the input / output layer of the instruction model θinst acquired by the third language model acquisition unit 104.

[0043] Then, the integrated model generation unit 106 generates a merge model θmerge (integrated model) by integrating the lexical substitution base model g(θbase) (first language model) generated by the first language model generation unit 102, the continuous pre-learning model θ'base (second language model) generated by the second language model generation unit 103, and the lexical substitution instruction model g(θinst) (fourth language model) generated by the fourth language model generation unit 105. Here, the merge model θmerge (integrated model) is a model in which the continuous pre-learning model θ'base (second language model) acquires the same instruction-following ability as the instruction model θinst (third language model).

[0044] The integrated model generation process shown in Figure 4 aims to omit the instruction learning process used when generating the instruction model θinst (third language model) by integrating the vocabulary substitution instruction model g(θinst) (fourth language model) with the continuously pre-trained model θ'base (second language model), which is generated by performing vocabulary modification g on the base model θbase and then performing continuous pre-training.

[0045] Therefore, in the integrated model generation process shown in Figure 4, when the server 12 generates the vocabulary replacement instruction model g(θinst) (fourth language model), it applies the same vocabulary changes g as the base model θbase to the instruction model θinst, thereby aligning the input / output layer of the instruction model θinst with the input layer of the base model θbase.

[0046] Then, in order to extract the effect of instruction learning used when the instruction model θinst (third language model) was generated from the lexical substitution instruction model g(θinst) (fourth language model), server 12 also includes the lexical substitution base model g(θbase) (first language model) in the merge.

[0047] For the reasons stated above, in the integrated model generation process shown in Figure 4, server 12 generates a merge model θmerge (integrated model) by integrating the lexical replacement base model g(θbase) (first language model), the continuous pre-training model θ'base (second language model), and the lexical replacement instruction model g(θinst) (fourth language model).

[0048] In one embodiment, the server 12 performs the integrated model generation process shown in Figure 4, thereby modifying the vocabulary of the base model θbase and, after further continuous pre-training, generating a merge model θmerge (integrated model) that includes the effect of instruction learning used when the instruction model θinst (third language model) was generated. As a result, it can generate a merge model θmerge (integrated model) that has acquired instruction following ability for the continuously pre-trained model θ'base (second language model).

[0049] In particular, the server 12 according to one embodiment can generate a merged model θmerge (integrated model) that includes the effects of instruction learning without having to perform instruction learning on the continuous pre-trained model θ'base (second language model), which is very time-consuming and laborious, by performing the integrated model generation process shown in Figure 4. Thus, it is possible to easily generate a merged model θmerge (integrated model) that has acquired instruction following ability.

[0050] Furthermore, "followability to follow instructions" refers to the ability to respond appropriately to user instructions. In particular, in chat models, followability to follow instructions is often used to describe the ability to appropriately follow user instructions and to converse naturally.

[0051] (An example of vocabulary modification by Server 12) Figure 5 shows an example of vocabulary modification by the server 12 according to one embodiment.

[0052] In the example shown in Figure 5, the old embedding layer contains identifiers for the vocabulary "queen", "trivial", "woman", and "king", and vectors corresponding to each of these identifiers. On the other hand, the new embedding layer contains identifiers for the vocabulary "woman", "king", "queen", and "queen", and vectors corresponding to each of these identifiers.

[0053] Here, for identifiers in the new embedding layer that represent vocabulary common to both the old and new embedding layers, a vector common to both layers is associated. On the other hand, for identifiers in the new embedding layer that represent vocabulary not included in the old embedding layer (i.e., added vocabulary), a vector generated based on vectors of multiple vocabulary words included in the old embedding layer (for example, a vector generated by the average or weighted average of multiple vocabulary vectors) is associated.

[0054] For example, in the example shown in Figure 5, for identifiers representing the vocabulary "queen" included in the new embedding layer, which is also included in the old embedding layer, a vector corresponding to the identifier representing the vocabulary "queen" included in the old embedding layer is associated.

[0055] Furthermore, for example, in the example shown in Figure 5, for the identifier representing the vocabulary word "queen" that is included in the new embedding layer but not in the old embedding layer, the average of the vectors of the vocabulary word "woman" and the vocabulary word "king" that are included in the old embedding layer is associated with it.

[0056] For example, the first language model generation unit 102 of server 12 generates a first language model by modifying the vocabulary of the input / output layer of the base model using the method shown in Figure 5.

[0057] Furthermore, for example, the fourth language model generation unit 105 of server 12 generates a fourth language model by modifying the vocabulary of the input / output layer of the third language model using the method shown in Figure 5.

[0058] In one embodiment, the server 12 can also perform the integrated model generation process shown in Figure 5 to modify the vocabulary of the base model θbase, and after further pre-training, generate a merged model θmerge (integrated model) that includes the effect of instruction learning used when the instruction model θinst (third language model) was generated. Thus, it can generate a merged model θmerge (integrated model) that has acquired instruction following ability.

[0059] (The concept of model integration) Figure 6 is a conceptual diagram of model integration by server 12 according to one embodiment.

[0060] Figure 6 illustrates an example of integrating Model A and Model B to generate an integrated model M. As shown in Figure 6, each of Models A, B, and M has one Embedding layer (input layer), multiple Hidden layers (intermediate layers), and one LM_head layer (output layer). The total number of Hidden layers in the integrated model M may increase or decrease depending on the total number of Hidden layers in Models A and B.

[0061] (Concept of model integration using chat vectors) Figure 7 is a conceptual diagram of the integration of a model using chat vectors by server 12 according to one embodiment.

[0062] Figure 7 illustrates an example of generating an integrated model M by adding the differences between models B and C (called "chat vectors") to model A. As shown in Figure 7, each of models A, B, C, and M has one embedding layer (input layer), multiple hidden layers (intermediate layers), and one LM_head layer (output layer).

[0063] As shown in Figure 7, when adding the chat vector to Model A, it is preferable to multiply the chat vector by an appropriate real coefficient (0.8 in the example shown in Figure 7).

[0064] (First modified example of the integrated model generation process by Server 12) Figure 8 is a conceptual diagram showing a first modified example of the integrated model generation process by the server 12 according to one embodiment.

[0065] In the integrated model generation process shown in Figure 8, the method by which the integrated model generation unit 106 generates the merge model θmerge (integrated model) differs from the integrated model generation process shown in Figure 4.

[0066] Specifically, in the integrated model generation process shown in Figure 8, the integrated model generation unit 106 generates a merge model θmerge (integrated model) by adding to the continuous pre-trained model θ'base (second language model) generated by the second language model generation unit 103 the difference (i.e., chat vector) between the lexical substitution instruction model g(θinst) (fourth language model) generated by the fourth language model generation unit 105 and the lexical substitution base model g(θbase) (first language model) generated by the first language model generation unit 102, multiplied by a predetermined multiple α (where α is a non-zero positive real number).

[0067] According to the integrated model generation process shown in Figure 8, the capabilities (instruction-following ability) of the instruction model θinst can be copied to the continuously pre-trained model θ'base (second language model).

[0068] (Second variation of the integrated model generation process by server 12) Figure 9 is a conceptual diagram showing a second modified example of the integrated model generation process by the server 12 according to one embodiment.

[0069] The integrated model generation process shown in Figure 9 differs from the integrated model generation process shown in Figure 4 in the following respects.

[0070] In the integrated model generation process shown in Figure 9, the server 12 generates an I / O layer-removed base model h(θbase) (an example of a "first intermediate model") by removing the input / output layers from the base model θbase.

[0071] Furthermore, server 12 generates an I / O layer-removed instruction model h(θinst) (an example of a "second intermediate model") by removing the input / output layer of the instruction model θinst (third language model).

[0072] Furthermore, server 12 generates a post-pre-trained instruction model θ'inst (fifth language model) by performing instruction learning on the pre-trained pre-trained model θ'base (second language model).

[0073] The integrated model generation unit 106 then adds to the continuous pre-trained model θ'base (second language model) the difference between the I / O layer removal instruction model h(θinst) and the I / O layer removal base model h(θbase) multiplied by a predetermined multiple α (where α is a non-zero positive real number), and further adds to the difference between the continuous pre-trained instruction model θ'inst (fifth language model) and the continuous pre-trained model θ'base (second language model) multiplied by a predetermined multiple β (where β is a non-zero positive real number), thereby generating the merge model θmerge (integrated model).

[0074] According to the integrated model generation process shown in Figure 9, by including the difference between the I / O layer-removed instruction model h(θinst) and the I / O layer-removed base model h(θbase) in the merge, the capabilities of the instruction model θinst (i.e., instruction-following ability) can be copied to the continuous pre-trained model θ'base (second language model) without applying vocabulary changes g to the input / output layers of the instruction model θinst (third language model) (i.e., without generating a vocabulary replacement instruction model g(θinst) (fourth language model)).

[0075] Furthermore, according to the integrated model generation process shown in Figure 9, by including the difference between the continuously pre-trained instruction model θ'inst (5th language model) and the continuously pre-trained model θ'base (2nd language model) in the merge, the merge model θmerge (integrated model) can correctly output the token that serves as a signal for the end of the LLM dialogue, which is artificially added for instruction following.

[0076] In one embodiment, the server 12 can also perform the integrated model generation process shown in Figure 9 to modify the vocabulary of the base model θbase, and after further continuous pre-training, generate a merge model θmerge (integrated model) that includes the effect of instruction learning used when the instruction model θinst (third language model) was generated. Thus, it can generate a merge model θmerge (integrated model) in which the continuously pre-trained model θ'base (second language model) has acquired instruction following ability.

[0077] Figure 10 is a diagram illustrating the relationship between each language model and the embedding layer described in this embodiment. As shown in Figure 10, each language model (large-scale language model) includes an input layer (W_embedding), a hidden layer (W_hidden), and an output layer (W_output). Changes to the vocabulary of the language model are made to the input layer (W_embedding) and the output layer (W_output). A specific example of a vocabulary change is shown in Figure 5. On the other hand, in instructional learning and continuous pre-training of the language model, changes are made to all of the input layer (W_embedding), the hidden layer (W_hidden), and the output layer (W_output).

[0078] Figure 11 shows an example of how the embedding layer changes in each language model described in this embodiment. In Figure 11, the embedding layer of the base model has vectors corresponding to identifiers that represent the vocabulary "woman", "king", and "kingdom", and the vocabulary replacement base model has vectors corresponding to identifiers that represent the vocabulary "queen", "woman", and "king". However, these vocabulary words are representative examples extracted for the purpose of understanding the explanation, and in reality, both the base model and the vocabulary replacement base model have a vast number of vocabulary words.

[0079] Furthermore, Figure 11 shows an example of generating a merge model θmerge (integrated model) using the integrated model generation process shown in Figure 4. In Figure 11, it is assumed that the vectors associated with each vocabulary are two-dimensional, and therefore the two-dimensional vectors of the three vocabulary words are represented as a 3*2 matrix. In the example shown in Figure 11, the method for changing the vocabulary is the method shown in Figure 5. Therefore, in the example shown in Figure 11, for the vocabulary words "woman" and "king," which are included in both the model before and after the vocabulary change, the vectors associated with the identifiers representing these vocabulary words in the embedding layer are the same before and after the vocabulary change. In addition, in the embedding layer of the model after the vocabulary change, the average of the vectors of the vocabulary word "woman" and the vocabulary word "king" from the base model (model before the vocabulary change) is associated with the identifier representing "queen," which is a vocabulary word newly added to the model after the vocabulary change.

[0080] Figure 12 shows an example of the embedding layer matrix of a merge model integrated model generated by server 12 according to one embodiment. Figure 12 shows the embedding layer matrix of the merge model θmerge (integrated model) generated when the integrated model generation process shown in Figure 8 is used instead of the integrated model generation process shown in Figure 4, in the case of the changes in the embedding layer shown in Figure 11. Here, the value of the predetermined multiple α is set to 0.5.

[0081] Figure 13 shows another example of the embedding layer matrix of the merge model integrated model generated by the server 12 according to one embodiment. Figure 13 shows the embedding layer matrix of the merge model θmerge (integrated model) generated when the integrated model generation process shown in Figure 9 is used instead of the integrated model generation process shown in Figure 4, in the case of the change in the embedding layer shown in Figure 11. Here, the value of the predetermined multiple α is set to 0.4. Also, here, the value of the predetermined multiple β is set to 0.5.

[0082] (Third variation of the integrated model generation process by Server 12) Figure 14 is a conceptual diagram showing a third modified example of the integrated model generation process by the server 12 according to one embodiment.

[0083] The integrated model generation process shown in Figure 14 differs from the integrated model generation process shown in Figure 4 in the following respects.

[0084] In the integrated model generation process shown in Figure 14, server 12 obtains a chat vector τ created from the base model θbase and the instruction model θinst (third language model). The chat vector τ may be an externally provided pre-configured model or it may be generated by server 12. The chat vector τ is the difference between the base model θbase and the instruction model θinst (third language model).

[0085] Then, server 12 generates a lexical replacement chat vector g(τ) by applying a lexical modification g to the chat vector τ.

[0086] Subsequently, server 12 generates a merged model θmerge (integrated model) by merging the continuous pre-trained model θ'base (second language model) and the lexical replacement chat vector g(τ).

[0087] The integrated model generation process shown in Figure 14 can generate a merge model θmerge (integrated model) similar to the integrated model generation process shown in Figure 4. In particular, the integrated model generation process shown in Figure 14 does not require vocabulary changes g to be applied to both the base model and the instruction model, thus reducing the processing load compared to the integrated model generation process shown in Figure 4.

[0088] (An example of a UI screen that accepts input for the dataset used for training) Figure 15 shows an example of a UI screen that accepts input for a dataset displayed by a PC 14 according to one embodiment.

[0089] Server 12 may obtain the dataset used for training the model from the internet, or it may obtain it by accepting input from a terminal device such as PC 14.

[0090] When server 12 obtains a dataset to be used for training a model by receiving input from a terminal device, it may display a UI screen, as illustrated in Figure 15, on the terminal device and receive the dataset and training settings selected by the user on the UI screen from the terminal device. Server 12 may then train the model based on the dataset and training settings received from the terminal device. For example, server 12 may perform continuous pre-training when generating a second language model, or instructional training when generating a fifth language model, based on the dataset and training settings received from the terminal device.

[0091] The UI screen 210 shown in Figure 15 includes a model selection field 211, a task selection field 212, a dataset source setting field 213, an optimizer selection field 214, a batch size setting field 215, and an epoch size setting field 216.

[0092] The model selection field 211 is used by the user to select the model to be trained.

[0093] The task selection field 212 is used by the user to select tasks for the model to perform. Examples of tasks include the following:

[0094] • Text Classification: Classifying which category a text belongs to.

[0095] • Named Entity Recognition (NER): Extracts named entities such as names of people, places, and organizations from text.

[0096] • Question Answering (QA): Generates answers to given questions.

[0097] • Text summarization: Summarizing long texts into shorter ones.

[0098] Machine translation: Translates text from one language into another language.

[0099] The dataset source setting field 213 is used by the user to set the storage location for the dataset. For example, the dataset source setting field 213 may be set to "local" as the upload destination for the dataset. Alternatively, for example, the dataset source setting field 213 may be set to the URL of a website as the storage location for the dataset.

[0100] The optimizer selection field 214 is used by the user to select an optimization algorithm for updating the model weights. Examples of optimization algorithms include SGD (Stochastic Gradient Descent), Adam (Adaptive Moment Estimation), RMSprop (Root Mean Square Propagation), and Adagrad (Adaptive Gradient Algorithm).

[0101] The batch size setting field 215 is used by the user to set the batch size (the number of data samples to train at one time). Any integer value can be set in the batch size setting field 215.

[0102] The epoch size setting field 216 is used by the user to set the epoch size (the number of training iterations for the entire dataset). Any integer value can be set in the epoch size setting field 216.

[0103] (Examples) The following describes an example of realizing a generated AI chat using the LLM (integrated model generated by server 12) described in the above embodiment.

[0104] (Functional configuration of Server 12 and PC 14) Figure 16 is a functional block diagram of a server 12 and a PC 14 according to one embodiment. Figure 16 differs from Figure 3 in that the server 12 includes a response generation unit 113.

[0105] The response generation unit 113 inputs the text entered by the user on the PC 14 into the LLM and generates a response sentence from the output of the LLM. The response generation unit 113 may edit and modify the output of the LLM to generate the response sentence, or it may use the output of the LLM as is as the response sentence.

[0106] (An example of the procedure for AI chat generation by Server 12 and PC 14) Figure 17 is a flowchart showing an example of the procedure for generating AI chat using server 12 and PC 14 according to one embodiment.

[0107] First, in server 12, the display control unit 111 transmits the screen information of the generated AI chat screen to PC 14 via the communication unit 112, thereby displaying the generated AI chat screen on PC 14 (step S301). At this time, in PC 14, the communication unit 122 receives the screen information of the generated AI chat screen, and the display unit 123 displays the generated AI chat screen.

[0108] Next, in PC14, the input reception unit 121 receives input text from the user on the generated AI chat screen (step S302).

[0109] Then, in PC14, the communication unit 122 sends the input message received in step S302 to the server 12 (step S303).

[0110] Next, in server 12, the response generation unit 113 inputs the input sentence received from PC 14 into the LLM, and uses the output from the LLM to generate a response sentence for the input sentence entered by the user (step S304). Here, the response generation unit 113 may use the output from the LLM as the response sentence as is, or it may generate a response sentence based on the output from the LLM.

[0111] Next, in the server 12, the display control unit 111 transmits the screen information of the generated AI chat screen, including the response sentence generated in step S304, to the PC 14 via the communication unit 112, thereby displaying the generated AI chat screen on the PC 14 (step S305). At this time, in the PC 14, the communication unit 122 receives the screen information of the generated AI chat screen, and the display unit 123 displays the generated AI chat screen.

[0112] Subsequently, Server 12 and PC 14 complete the series of processes shown in Figure 17.

[0113] (Example of a generated AI chat screen) Figures 18 and 19 show an example of a generated AI chat screen displayed on a PC14 according to one embodiment.

[0114] The generated AI chat screen 220 shown in Figure 18 is displayed on the PC 14 under the control of the display control unit 111 of the server 12, and represents the state after the image and input text have been displayed by the user.

[0115] In the example shown in Figure 18, a section for selecting a model is displayed at the top of the right-hand display area 222 of the generated AI chat screen 220. The user can select an LMM language model in this section.

[0116] In the example shown in Figure 18, the input field at the bottom of the display area 222 on the right side of the generated AI chat screen 220 displays the input text entered by the user, "I like pancakes. Do you know any good places in Tokyo?". The user can send this input text to the server 12 by pressing the triangular send button displayed in the input field. In addition, buttons for voice input and file attachment are displayed in the input field, and the user can input the text by voice or attach a file by pressing these buttons.

[0117] The generated AI chat screen 230 shown in Figure 19 represents the state after the response to the input sentence has been displayed under the control of the display control unit 111 of the server 12. In the example shown in Figure 19, the input sentence entered by the user and the response to the input sentence generated by the response generation unit 113 of the server 12 are displayed in chronological order in the display area 232 on the right side of the generated AI chat screen 230.

[0118] In addition, in the generated AI chat screens 220 and 230, the left-hand display area 221 and 231 shows the history of generated AI chats. However, the history of generated AI chats may also be displayed in several groups, such as "Recent" and "2024 / 12 / 11" (date).

[0119] Additionally, display areas 221 and 231 show a search field, and by entering a search term into this field, the user can view the history of generated AI chats that match that search term.

[0120] According to one embodiment of the server 12 and PC 14, in the generation AI chat processing, the performance of the model can be improved by using the LLM (integrated model generated by the server 12) described in the above embodiment, thereby improving the accuracy of the response sentence to the input sentence.

[0121] Although preferred embodiments of the present invention have been described in detail above, the present invention is not limited to these embodiments, and various modifications or changes are possible within the scope of the gist of the present invention as described in the claims.

[0122] The apparatus described in the examples represents only one of several computing environments for carrying out the embodiments disclosed herein. In one embodiment, server 12 includes multiple computing devices, such as a server cluster. The multiple computing devices are configured to communicate with each other via any type of communication link, including a network or shared memory, and perform the processing disclosed herein. In another embodiment, this may be carried out by a single device having the functional configuration of both a terminal device and an information processing device. [Explanation of Symbols]

[0123] 10. Information Processing Systems 12 servers 14 PC 16. Communication Networks 101 Base Model Acquisition Section 102 First Language Model Generation Unit 103 Second Language Model Generation Unit 104 Third Language Model Acquisition Unit 105 Fourth Language Model Generation Unit 106 Integrated Model Generation Unit 111 Display Control Unit 112 Communications Department 113 Response generation unit 121 Input Reception Section 122 Communications Department 123 Display section 210 UI screen 211 Model Selection Field 212 Task Selection Field 213 Dataset Source Settings Section 214 Optimizer Selection Field 215 Batch size setting field 216 Epoch Size Setting Field 220,230 Generated AI chat screen 221,222,231,232 display area [Prior art documents] [Non-patent literature]

[0124] [Non-Patent Document 1] https: / / arxiv.org / pdf / 2310.04799

Claims

1. The first language model is generated by modifying the vocabulary of the input / output layer of the base language model, A second language model generated by performing continuous pre-training on the first language model, A fourth language model is generated by applying the same vocabulary changes to the input / output layer of a third language model, which is generated by performing instruction learning on the base language model, as the vocabulary changes used when generating the first language model. The integrated model generation unit generates an integrated model by integrating the components. An information processing device equipped with the following features.

2. The integrated model generation unit is, The integrated model is generated by adding the difference between the second language model and the fourth language model and the first language model. The information processing apparatus according to claim 1.

3. The integrated model generation unit is, The integrated model is generated by adding to the second language model the difference between the fourth language model and the first language model multiplied by a predetermined multiple. The information processing apparatus according to claim 1.

4. In the aforementioned vocabulary modification, For vocabulary common to both the pre- and post-modification language models, a common vector is associated with the post-modification language model. For vocabulary added to the modified language model, the average of the vectors of multiple vocabulary words included in the original language model is associated with the modified language model. The information processing apparatus according to claim 1.

5. The second language model is generated by continuously pre-training the first language model, which is generated by modifying the vocabulary of the input / output layers of the base language model, A fifth language model generated by performing instructional learning on the second language model, A first intermediate model is generated by removing the input / output layer of the base language model, A second intermediate model is generated by removing the input / output layer of the third language model, which was generated by performing instruction learning on the aforementioned base language model. The integrated model generation unit generates an integrated model by integrating the components. An information processing device equipped with the following features.

6. The second language model is generated by continuously pre-training the first language model, which is generated by modifying the vocabulary of the input / output layers of the base language model, The input / output layer of the chat vector, which is the difference between the base language model and the third language model generated by performing instruction learning on the base language model, is configured to apply a vocabulary replacement chat vector generated by applying the same vocabulary modification as the vocabulary modification used when generating the first language model. The integrated model generation unit generates an integrated model by integrating the components. An information processing device equipped with the following features.

7. The first language model is generated by modifying the vocabulary of the input / output layer of the base language model, A second language model generated by performing continuous pre-training on the first language model, A fourth language model is generated by applying the same vocabulary changes to the input / output layer of a third language model, which is generated by performing instruction learning on the base language model, as the vocabulary changes used when generating the first language model. The integrated model generation unit generates an integrated model by integrating the components. An information processing system equipped with the following features.

8. The second language model is generated by continuously pre-training the first language model, which is generated by modifying the vocabulary of the input / output layers of the base language model, A fifth language model generated by performing instructional learning on the second language model, A first intermediate model is generated by removing the input / output layer of the base language model, A second intermediate model is generated by removing the input / output layer of the third language model, which was generated by performing instruction learning on the aforementioned base language model. The integrated model generation unit generates an integrated model by integrating the components. An information processing system equipped with the following features.

9. The second language model is generated by continuously pre-training the first language model, which is generated by modifying the vocabulary of the input / output layers of the base language model, The input / output layer of the chat vector, which is the difference between the base language model and the third language model generated by performing instruction learning on the base language model, is configured to apply a vocabulary replacement chat vector generated by applying the same vocabulary modification as the vocabulary modification used when generating the first language model. The integrated model generation unit generates an integrated model by integrating the components. An information processing system equipped with the following features.

10. The first language model is generated by modifying the vocabulary of the input / output layer of the base language model, A second language model generated by performing continuous pre-training on the first language model, A fourth language model is generated by applying the same vocabulary changes to the input / output layer of a third language model, which is generated by performing instruction learning on the base language model, as the vocabulary changes used when generating the first language model. The integrated model generation process generates an integrated model by integrating the elements. Information processing methods including

11. The second language model is generated by continuously pre-training the first language model, which is generated by modifying the vocabulary of the input / output layers of the base language model, A fifth language model generated by performing instructional learning on the second language model, A first intermediate model is generated by removing the input / output layer of the base language model, A second intermediate model is generated by removing the input / output layer of the third language model, which was generated by performing instruction learning on the aforementioned base language model. The integrated model generation process generates an integrated model by integrating the elements. Information processing methods including

12. The second language model is generated by continuously pre-training the first language model, which is generated by modifying the vocabulary of the input / output layers of the base language model, The input / output layer of the chat vector, which is the difference between the base language model and the third language model generated by performing instruction learning on the base language model, is configured to apply a vocabulary replacement chat vector generated by applying the same vocabulary modification as the vocabulary modification used when generating the first language model. The integrated model generation process generates an integrated model by integrating the elements. Information processing methods including

13. Computers, The first language model is generated by modifying the vocabulary of the input / output layer of the base language model, A second language model generated by performing continuous pre-training on the first language model, A fourth language model is generated by applying the same vocabulary changes to the input / output layer of a third language model, which is generated by performing instruction learning on the base language model, as the vocabulary changes used when generating the first language model. The integrated model generation unit generates an integrated model by integrating the components. A program that makes it function as such.

14. Computers, The second language model is generated by continuously pre-training the first language model, which is generated by modifying the vocabulary of the input / output layers of the base language model, A fifth language model generated by performing instructional learning on the second language model, A first intermediate model is generated by removing the input / output layer of the base language model, A second intermediate model is generated by removing the input / output layer of the third language model, which was generated by performing instruction learning on the aforementioned base language model. The integrated model generation unit generates an integrated model by integrating the components. A program that makes it function as such.

15. Computers, The second language model is generated by continuously pre-training the first language model, which is generated by modifying the vocabulary of the input / output layers of the base language model, The input / output layer of the chat vector, which is the difference between the base language model and the third language model generated by performing instruction learning on the base language model, is configured to apply a vocabulary replacement chat vector generated by applying the same vocabulary modification as the vocabulary modification used when generating the first language model. The integrated model generation unit generates an integrated model by integrating the components. A program that makes it function as such.

Images