Training method

Abstract

A training method includes: an input step of inputting data provenance-specific text data to a large language model in pre-training of the large language model; an acquisition step of acquiring a value of an intermediate layer of the large language model at a time when the large language model is pre-trained with the text data as input; and a training step of generating a training model by supervised learning with the acquired value of the intermediate layer as input data and with provenance information representing provenance of the text data as correct answer data.

Description

Technical Field

[0001] This invention relates to the technical field of training methods. Background Technology

[0002] As an example of this approach, the following method has been proposed: Large Language Models (LLMs) generate document-based query data, and a retrieval model for a chatbot is trained using pairs of document and query data (see Japanese Patent Application Publication No. 2023-076413). Summary of the Invention

[0003] For example, in services using large language models, the output of the large language model may contain all or part of the work of others. In this case, copyright-related issues may arise. Note that a large language model refers to a language model built using very large datasets and deep learning techniques.

[0004] The present invention was made in view of the above-mentioned problems, and its objective is to provide a training method that can reduce the risk of infringement.

[0005] One aspect of the training method of the present invention includes: an input step, in which source-clear text data is input into the large language model during pre-training; an acquisition step, in which the values ​​of the intermediate layers of the large language model are acquired during pre-training with the text data as input; and a training step, in which a training model is generated through supervised learning, wherein the supervised learning takes the acquired values ​​of the intermediate layers as input data and source information indicating the source of the text data as positive solution data. Attached Figure Description

[0006] The features, advantages, and technical and industrial significance of exemplary embodiments of the present invention will now be described with reference to the accompanying drawings, in which similar symbols denote similar elements, wherein:

[0007] Figure 1 It is a concept map illustrating the concepts of a large language model; and

[0008] Figure 2 This is a flowchart illustrating the training method of the implementation method. Detailed Implementation

[0009] Reference Figure 1 and Figure 2 Explain the implementation method of the training method. Figure 1 In a large language model (LLM), there is an input layer, an output layer, and multiple intermediate layers. Note that the "intermediate layers" can also be called "hidden layers".

[0010] Training a large language model can include pre-training and post-training. For example, in pre-training, a pre-training corpus (i.e., a large amount of text data) is used to train the large language model. Specific examples of pre-training include "next token prediction" and "masked token prediction." However, since this implementation is unrelated to post-training, its detailed description is omitted.

[0011] Pre-training corpora may contain all or part of a work. The output of a large language model trained using such a pre-training corpus may contain all or part of the work. In this case, copyright-related issues may arise.

[0012] In this embodiment, the training method of the model (hereinafter referred to as model M) is described, wherein the model infers the data upon which the large language model's response (i.e., output) is based. It is assumed that the pre-training corpus data contains text data with clearly identifiable sources. It is also assumed that the text data with clearly identifiable sources is assigned metadata containing source information (see [reference]). Figure 1 The symbol MD).

[0013] During pre-training, when text data is input into a large language model, the values ​​of the intermediate layers (e.g., intermediate layer MLx) of the large language model are influenced by the input text data. Furthermore, during pre-training, the data output from the large language model is influenced by the text data input into it. Therefore, it can be said that the values ​​of the intermediate layers (e.g., intermediate layer MLx) of the large language model are influenced by the data upon which the output of the large language model is based. Therefore, in this embodiment, a model M is constructed by using metadata assigned to the text data and the values ​​of the intermediate layers to infer the data upon which the response (i.e., output) of the large language model is based. Furthermore, a server (e.g., a cloud server) can be used to construct model M.

[0014] Reference Figure 2 The flowchart is explained in detail below. In the pre-training of the large language model, text data with metadata including source information is input into the large language model (step S101). The values ​​of the intermediate layers (e.g., intermediate layer MLx) of the large language model are obtained when the large language model is pre-trained with the input text data (step S102). Then, the source information contained in the metadata assigned to the input text data in step S101 and the values ​​of the intermediate layers obtained in step S102 can be correlated. The source information and the values ​​of the intermediate layers become the training data used to train model M.

[0015] Repeat steps S101 and S102 until sufficient training data for training model M is collected. After sufficient training data (i.e., source information and intermediate layer values) has been collected, model M is trained through supervised learning using the intermediate layer values ​​as input data and the source information as the positive solution data (step S103). Such model M can be a training model related to a multi-label classifier.

[0016] The model M constructed as described above can obtain the values ​​of the intermediate layers of the large language model when the constructed large language model generates an answer to an input (e.g., a question). Model M uses the obtained intermediate layer values ​​as input to infer the source of the text data used by the large language model when generating the answer.

[0017] Technical effect

[0018] The model M, constructed using the training method of this embodiment, infers the source of the text data used by the large language model when generating responses. For example, by referring to the source inferred by model M (in other words, the data upon which the large language model bases its responses), it is relatively easy to determine whether copyright-related issues will arise. Therefore, the training method of this embodiment can reduce the risk of infringement.

[0019] Various aspects of the invention derived from the above embodiments will be described below.

[0020] One aspect of the training method of the present invention includes: an input step, in which source-clear text data is input into the large language model during pre-training; an acquisition step, in which the values ​​of the intermediate layers of the large language model are acquired during pre-training with the text data as input; and a training step, in which a training model is generated by supervised learning using the acquired intermediate layer values ​​as input data and source information representing the source of the text data as positive solution data.

[0021] In the training method described above, the training model can be a model that infers the source of the text data used by the large language model when generating responses. In the training method described above, the training model can be a training model related to a multi-label classifier.

[0022] This invention is not limited to the embodiments described above. Modifications can be made as appropriate without departing from the spirit or idea of ​​the invention as understood in its entirety by the claims and description. Training methods that accompany such modifications are also included within the technical scope of this invention.

Claims

1. A training method, comprising: In the input step, during the pre-training of the large language model, text data with a clear source is input into the large language model; The acquisition step involves acquiring the values ​​of the intermediate layers of the large language model during pre-training with the text data as input. as well as The training steps involve generating a training model through supervised learning, where the values ​​of the intermediate layers are used as input data, and source information representing the origin of the text data is used as the correct solution data.

2. The training method according to claim 1, wherein The training model is a training model that infers the source of the text data used by the large language model when generating answers.

3. The training method according to claim 1, wherein The training model is a training model related to a multi-label classifier.

Images