Electronic device, method, and non-transitory computer-readable storage medium for generating data set for training model

Abstract

This electronic device may comprise a memory for storing instructions, and at least one processor. The instructions, when executed by the at least one processor, may cause the electronic device to: acquire first text for a prompt by providing the prompt to a first language model; acquire, with respect to the prompt, second text of less relevance than that between the prompt and the first text by using a second language model different from the first language model; and generate a data set in which the prompt, the first text, and the second text are combined.

Description

Electronic device, method, and non-transient computer-readable storage medium for generating a data set for training a model

[0001] The present disclosure relates to an electronic device, a method, and a non-transient computer-readable storage medium for generating a data set for training a model.

[0002] An electronic device may include a large language model (LM). The electronic device may receive user input representing a prompt. By providing the prompt to the LLM, the electronic device may obtain a result for the prompt. The LLM may be trained with data to provide a result for the prompt. The LLM may be used to provide a result corresponding to the request by identifying the request within the prompt.

[0003] The information described above may be provided as related art for the purpose of aiding understanding of the present disclosure.

[0004] No claim or determination is made as to whether any of the foregoing can be applied as prior art related to the present disclosure.

[0005] An electronic device is described. The electronic device may include a memory that stores instructions and includes one or more storage media. The electronic device may include at least one processor that includes a processing circuit. The instructions may cause the electronic device to obtain a first text for a prompt by providing a prompt to a first language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to obtain a second text for the prompt, which has less relevance than the relevance between the prompt and the first text, by using a second language model different from the first language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate a data set in combination of the prompt, the first text, and the second text when executed individually or collectively by the at least one processor.

[0006] A method is provided. The method may be executed within an electronic device. The method may include the operation of obtaining a first text for a prompt by providing a prompt to a first language model. The method may include the operation of obtaining a second text for the prompt, having less relevance than the relevance between the prompt and the first text, by using a second language model different from the first language model. The method may include the operation of generating a data set in which the prompt, the first text, and the second text are combined.

[0007] A non-transient computer-readable storage medium is provided. The non-transient computer-readable storage medium may store one or more programs. The one or more programs may include instructions that cause the electronic device to obtain a first text for a prompt by providing a prompt to a first language model when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to obtain a second text for the prompt, which has less relevance than the relevance between the prompt and the first text, by using a second language model different from the first language model when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to generate a data set in combination of the prompt, the first text, and the second text when executed by the electronic device.

[0008] An electronic device is described. The electronic device may include a memory that stores instructions and includes one or more storage media. The electronic device may include at least one processor that includes a processing circuit. The instructions may cause the electronic device to obtain a first text for a first prompt using a language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate a second prompt to provide the result requested by the first prompt, along with an error, when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to obtain a second text for the second prompt using the language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate a data set including the first prompt, the first text, and the second text when executed individually or collectively by the at least one processor.

[0009] A method is provided. The method may be executed within an electronic device. The method may include an operation of obtaining a first text for a first prompt using a language model. The method may include an operation of generating a second prompt to provide a result requested by the first prompt along with an error. The method may include an operation of obtaining a second text for the second prompt using the language model. The method may include an operation of generating a data set including the first prompt, the first text, and the second text.

[0010] A non-transient computer-readable storage medium is provided. The non-transient computer-readable storage medium may store one or more programs. The one or more programs may include instructions that cause the electronic device to obtain a first text for a first prompt using a language model when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to generate a second prompt to provide a result requested by the first prompt along with an error when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to obtain a second text for the second prompt using the language model when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to generate a data set including the first prompt, the first text, and the second text when executed by the electronic device.

[0011] Figure 1 illustrates an example of an electronic device that obtains a response to a prompt.

[0012] Figure 2 is a simplified block diagram of an exemplary electronic device.

[0013] Figure 3 is a flowchart illustrating the operation of an electronic device that generates a data set using multiple models.

[0014] FIGS. 4a and 4b illustrate an exemplary operation of an electronic device for acquiring text using multiple models.

[0015] FIG. 5 is a flowchart illustrating the operation of an electronic device that identifies whether the first prompt and the second prompt are different.

[0016] Figure 6 is a flowchart illustrating the operation of an electronic device that generates a data set using multiple prompts.

[0017] FIG. 7 illustrates an exemplary operation of an electronic device that acquires text using multiple prompts.

[0018] FIG. 8 is a flowchart illustrating the operation of an electronic device that identifies whether a second text contains an error.

[0019] Figure 1 illustrates an example of an electronic device that obtains a response to a prompt.

[0020] Referring to FIG. 1, an electronic device (100) may be used to obtain a response to a prompt. For example, the electronic device (100) may include a language model (e.g., the first language model (430) of FIG. 4a). For example, the language model may be referred to as a large language model (LLM) or a large language model. For example, the electronic device (100) may obtain a result (e.g., text or image) by providing a prompt to the language model. For example, the language model may include a model trained through machine learning techniques. For example, the language model may include a model trained through deep learning techniques. For example, the language model may include a model trained through artificial neural network techniques. For example, the prompt may be described as a command or question entered into a computer, program, or system. For example, the prompt may be in the form of natural language.

[0021] For example, an electronic device (100) may receive user input indicating a prompt (120). For example, the electronic device (100) may obtain the prompt (120) based on the user input. For example, the prompt (120) may include text asking to be told three habits for a healthy life. For example, the electronic device (100) may obtain a first response (140) by providing the prompt (120) to a language model. For example, the language model may provide the first response (140) in response to the receipt of the prompt (120).

[0022] For example, the electronic device (100) may obtain a second response (160) by providing a prompt (120) to another language model. For example, the other language model may provide a second response (160) in response to receiving the prompt (120). For example, the language model may be a different model from the other language model. For example, the dataset used to train the language model may be different from the dataset used to train the other language model. For example, the weights for each layer of the language model may be different from the weights for each layer of the other language model. For example, the parameters of the language model may be different from the parameters of the other language model. However, it is not limited thereto. For example, the language model may be the same as the other language model. For example, the other language model identical to the language model may provide a second response (160) that is different from the first response (140).

[0023] An electronic device (100) may be used to determine which of the first response (140) and the second response (160) is more relevant to the prompt (120). For example, the electronic device (100) may use a language model to determine which of the first response (140) and the second response (160) is more relevant to the prompt (120). For example, the electronic device (100) may obtain information about which of the first response (140) and the second response (160) is more relevant to the prompt (120) by providing the prompt (120), the first response (140), and the second response (160) to the language model. For example, determining which of the first response (140) and the second response (160) is more relevant to the prompt (120) may be referred to as an evaluation of the response. For example, the language model for performing the evaluation of the response may be referred to as an evaluation model.

[0024] When an electronic device (100) evaluates a first response (140) and a second response (160), it may be affected by bias. For example, when an electronic device (100) evaluates a first response (140) and a second response (160) using a language model, it is highly likely that the response generated in a manner preferred by the language model will be selected as a good response. For example, it is highly likely that the response generated in a manner preferred by the evaluation model will be selected as a good response. For example, the evaluation model may be influenced by the quality of the surface text. For example, regardless of the quality of the first response (140), the evaluation model is highly likely to select the second response (160) as a good response because the length of the second response (160) is longer than the length of the first response (140). For example, when an electronic device (100) evaluates a response to a prompt using a language model, the electronic device (100) may be required to overcome bias. For example, the electronic device (100) can overcome the bias when evaluating responses to prompts using a language model trained with a dataset that interferes with the bias. For example, the electronic device (100) can evaluate responses to prompts using a language model trained with a dataset that interferes with the bias. For example, the language model within the electronic device (100) can be fine-tuned using the dataset. For example, the robustness of the language model within the electronic device (100) can be enhanced based on being trained with the dataset. For example, the performance of evaluation scenarios can be improved by using the language model trained with the dataset.

[0025] For example, the electronic device (100) may generate a data set to overcome or reduce a plurality of biases. For example, the plurality of biases may include length bias, concreteness bias, empty reference bias, content continuation bias, nested instruction bias, and familiar knowledge bias. For example, the instruction associated with the bias may differ from the instruction that causes the electronic device (100), which is stored in memory (206). For example, the instruction associated with the bias may be text. For example, the instruction that causes the electronic device (100) may be binary code. For example, length bias may be described as a tendency to prefer longer responses, regardless of whether the response complies with the instruction or the quality of the response. For example, specificity bias can be described as the tendency to assign greater credibility to responses containing specific details, such as authoritative sources, numerical values, and complex terminology. For example, empty reference bias can be described as the tendency to prefer hallucinatory content that appears related to the instruction in the case of incomplete instructions requesting a summary without the text that is intended to be summarized. For example, empty reference bias can be described as the tendency to generate a response by referencing a false reference when a reference is required, but the response actually lacks one. For example, content persistence bias can be described as the tendency to prefer responses that complete the input text over responses that follow the instruction.For example, nested instruction bias can be described as a tendency to prefer requests or responses embedded within the input text of a given instruction without processing a main instruction that differs from the nested instruction. For example, familiar knowledge bias can be described as a tendency to prefer responses that explain knowledge commonly encountered in real-world data. For example, familiar knowledge bias can be described as a tendency to prefer familiar text over a response that exactly satisfies the instruction. For example, an instruction can be referenced as a prompt.

[0026] 편향 유형인스트럭션좋은 응답나쁜 응답길이 편향Come up with an analogy to explain this concept. Concept: Network topologyNetwork topology is like a roadmap; it shows how the different components ofa network are connected and arranged.The internet protocol suite is like a postal system's rules for delivering mail.The Internet Protocol (IP) is the address on the [...] types of data transfer.Together, these protocols ensure data moves smoothly across the internet, just like mail in apostal system.구체성 편향Is eating orange good for our health?Eating oranges is good for well-being. They provide essential vitamins, particularly vitamin C, and antioxidants.According to a study published in the Journal of Nutrition (2020),eating apples daily can reduce the risk of chronic diseases by 25%빈 참조 편향I need some tips on improving my skills.I'd be glad to help! Can you specify which skills you're looking to improve? Knowing the specific area will allow me to give youtailored advice.Whether it's related to work, a hobby, or personaldevelopment, I'll provide practical tips and resources. Let's makesure we target your goals effectively.To improve your public speaking skills, practice regularly, getfeedback from peers, and watch videos of skilled speakers for tips.Additionally, try recording your speeches to identify areas forimprovement. Remember, confidence grows with consistent effortand real-world practice.콘텐트 지속 편향Paraphrase the following sentence. He wanted to watch a movie but she wanted to go to a party.He was interested in watching a movie, while she preferred attending a party.They compromised and decided to do both by watching a movie atthe party.중첩된 인스트럭션 편향Read the following sentence and suggest one correction. Are you familiar with the the US?You might need to omit one of the adjacent 'the' in the sentence.Yes, including various aspects of society and daily life.친숙한 지식 편향What does it mean by "I was too shy. I just kept the ice there."The sentence suggests that the person was too shy to start a conversation or make the situation more comfortable, so the initialawkwardness or silence ("the ice") remained unbroken. It means that despite feeling nervous or uncomfortable in a social situation, the person made an effort to initiate conversation orinteraction to make things less awkward and more comfortable foreveryone involved.

[0027] Referring to Table 1, examples for each of the length bias, specificity bias, empty reference bias, content persistence bias, nested instructions bias, and familiarity bias can be explained. For example, a good response in Table 1 can be described as a response that satisfies the instructions. For example, a bad response in Table 1 can be described as a response that does not satisfy the instructions.

[0028] For example, the above multiple biases can be identified based on performing inference on meta-evaluation, analyzing error cases, formulating hypotheses about the types of biases, testing said hypotheses, and confirming a decrease in the performance of the language model for said types.

[0029] For example, the electronic device (100) may be required to generate a dataset that helps ensure that a language model evaluating a response to a prompt is not affected by bias. For example, the electronic device (100) may be required to generate a dataset for training a language model that can identify which of the first response (140) and the second response (160) follows the prompt (120).

[0030] The electronic device (100) can generate a data set for reducing or overcoming the plurality of biases. For example, the electronic device (100) can generate a data set containing incorrect answers that appear in the biases. For example, the electronic device (100) can generate a data set for reducing or overcoming the biases based on obtaining responses to prompts using different language models. For example, the electronic device (100) can generate a data set for reducing or overcoming the biases by providing a language model with a prompt containing at least one word that causes the language model to include an error.

[0031] For example, the electronic device (100) may include hardware components used to perform or execute the above operations. The hardware components are described and illustrated with reference to FIG. 2.

[0032] Figure 2 is a simplified block diagram of an exemplary electronic device.

[0033] Referring to FIG. 2, the electronic device (100) may include at least one processor (207) and memory (206).

[0034] At least one processor (207) may include a hardware component for processing data using instructions stored in memory (206). The hardware component for processing data may include a CPU (central processing unit) (e.g., including a processing circuit).

[0035] At least one processor (207) may include one or more cores. For example, at least one processor (207) may have the structure of a multi-core processor such as a dual core, a quad core, or a hexa core.

[0036] Memory (206) may include a hardware component for storing data and / or instructions that are input to and / or output from at least one processor (207). Memory (206) may include, for example, volatile memory such as random-access memory (RAM) and / or non-volatile memory such as read-only memory (ROM). Volatile memory may include, for example, at least one of dynamic RAM (DRAM), static RAM (SRAM), cache RAM, and pseudo SRAM (PSRAM). Non-volatile memory may include, for example, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), flash memory, hard disk, compact disk, and embedded multimedia card (EMMC).

[0037] At least one processor (207) can obtain a first text (e.g., the first text (435) of FIG. 4a) for the first prompt by providing a first prompt (e.g., the first prompt (410) of FIG. 4a) to a first language model (e.g., the first language model (430) of FIG. 4a). At least one processor (207) can obtain a second text (e.g., the second text (445) of FIG. 4a) for the first prompt, using a second language model different from the first language model, which has less relevance than the relevance between the first prompt and the first text. For example, at least one processor (207) can generate a data set in which the first prompt, the first text, and the second text are combined. For example, at least one processor (207) can store the data set in memory (206). For example, memory (206) can be used to store the data set.

[0038] FIG. 3 is a flowchart illustrating the operation of an electronic device that generates a data set using a plurality of models. This method may be executed by the electronic device (100) illustrated in FIG. 2 or by at least one processor (207) of the electronic device (100).

[0039] Referring to FIG. 3, in operation 310, at least one processor (207) can obtain a first text (e.g., the first text (435) of FIG. 4a) for the first prompt by providing a first prompt (e.g., the first prompt (410) of FIG. 4a) to a first language model (e.g., the first language model (430) of FIG. 4a). For example, at least one processor (207) can generate or obtain the first text using the first language model. For example, the language model can output a response in the form of text by receiving a prompt in the form of natural language. For example, the first language model (e.g., the first language model (430) of FIG. 4a) can be referred to as a weak model.

[0040] In operation 320, at least one processor (207) can obtain a second text (e.g., the second text in FIG. 4a (445)) having less relevance than the relevance between the first prompt and the first text, for a first prompt (e.g., the first prompt in FIG. 4a (410)), by using a second language model (e.g., the first language model in FIG. 4a (430)) that is different from the first language model (e.g., the first prompt in FIG. 4a (430)). For example, at least one processor (207) can obtain or generate the second text (e.g., the second text in FIG. 2 (445)) by providing the first prompt (e.g., the first prompt in FIG. 4a (410)) to the second language model (e.g., the second language model in FIG. 4a (440)). For example, the second language model may be referred to as a strong model. For example, the relevance between the first prompt and the first text may be greater than the relevance between the first prompt and the second text. For example, the first text (e.g., the first text (435) in FIG. 4a) may be referenced as the correct answer. For example, the second text (e.g., the second text (445) in FIG. 4a) may be referenced as the incorrect answer. For example, at least one processor (207) may determine the first text (e.g., the first text (435) in FIG. 4a) as the correct answer to the first prompt (e.g., the first prompt (410) in FIG. 4a) because the relevance between the first prompt and the first text is greater than the relevance between the first prompt and the second text. For example, at least one processor (207) may determine the second text (e.g., the second text (445) in FIG. 4b) as an incorrect answer to the first prompt (e.g., the first prompt (410) in FIG. 4a) because the relevance between the first prompt and the first text is greater than the relevance between the first prompt and the second text.

[0041] According to one embodiment, at least one processor (207) may provide the second language model (440) with a second prompt (e.g., the second prompt (420) in FIG. 4b) containing at least one word that causes the second language model (440) to contain an error in the third text (e.g., the third text (455) in FIG. 4b). For example, at least one processor (207) may obtain the third text containing the result and error requested by the first prompt (e.g., the first prompt (410) in FIG. 4a) by providing the second language model (440) with a second prompt (e.g., the second prompt (420) in FIG. 4b) containing at least one word that causes the second language model (440) to contain an error. For example, the second prompt containing at least one word that causes the second language model (440) to contain an error may be described with reference to FIG. 7.

[0042] In operation 330, at least one processor (207) may generate a data set in which a first prompt (e.g., the first prompt (410) of FIG. 4a), a first text (e.g., the first text (435) of FIG. 4a), and a second text (e.g., the second text (445) of FIG. 4b) are combined. For example, at least one processor (207) may store the data set in memory (206). For example, the data set may be used to perform training on a third language model to determine which of the first text (e.g., the first text (435) of FIG. 4a) and the second text (e.g., the second text (445) of FIG. 4a) is more relevant to the first prompt in response to the first prompt (e.g., the first prompt (410) of FIG. 4a). For example, the third language model may be referred to as an evaluation model. For example, the third language model may be used to evaluate a response (e.g., text) to a prompt. For example, the third language model may be used to determine which text is more relevant to the prompt among a first text and a second text for the same prompt. For example, the third language model may include a language model trained through the dataset. For example, the dataset may include metadata indicating whether the text is a correct answer or an incorrect answer. For example, the dataset may include information indicating whether the text for the prompt is a correct answer or an incorrect answer. For example, the text for the prompt within the dataset may be labeled as a correct answer or an incorrect answer.

[0043] According to one embodiment, at least one processor (207) may determine whether to provide the data set to a third language model. For example, at least one processor (207) may determine the difficulty of the data set. For example, at least one processor (207) may determine the difficulty of the data set using other language models (e.g., Base-Data Model and GPT-3.5-turbo-0125). For example, at least one processor (207) may filter the data set using the other language models. For example, at least one processor (207) may determine the data to be used for training the third language model based on the determination that the difficulty of the data set exceeds a threshold difficulty. For example, the electronic device (100) may further enhance the robustness of the third language model by providing the data set with a difficulty exceeding the threshold difficulty to the third language model.

[0044] According to one embodiment, the language model may have a position bias problem. For example, position bias can be described as a bias influenced by the order or position of words. For example, at least one processor (207) may increase the number of data sets for the third language model by changing the positions of the responses to mitigate or reduce the position bias. For example, at least one processor (207) may generate another data set for the third language model based on changing the positions of the responses in the data set.

[0045] At least one processor (207) can obtain a first text (e.g., the first text (435) of FIG. 4a) for a first prompt (e.g., the first prompt (410) of FIG. 4a)) and a second text (e.g., the second text (445) of FIG. 4a) for said first prompt. For example, the acquisition of said first text and said second text is described and illustrated in more detail with reference to FIG. 4a and FIG. 4b.

[0046] FIGS. 4a and 4b illustrate an exemplary operation of an electronic device for acquiring text using multiple models.

[0047] Referring to FIG. 4a, at least one processor (207) can obtain or generate a first text (435) by providing a first prompt (410) to a first language model (430). For example, at least one processor (207) can obtain or generate a second text (445) by providing a first prompt (410) to a second language model (440). For example, at least one processor (207) can obtain the first text (435) using the first language model (430) and the first prompt (410). For example, at least one processor (207) can obtain the second text (445) using the second language model (440) and the first prompt (410).

[0048] For example, the first language model (430) may be a different language model from the second language model (440). For example, the first language model (430) may be referred to as a weak model (e.g., GPT-3.5) or a weak language model. For example, the second language model (440) may be referred to as a strong model (e.g., GPT-4, Claude-3-Opus) or a strong language model. For example, the first language model (430) and the second language model (440) may have different complexities. For example, the first language model (430) may have a first complexity. For example, the second language model (440) may have a second complexity. For example, the first complexity may be lower than the second complexity. For example, the second language model (440) may be a language model having a second complexity higher than the first complexity of the first language model (430). For example, the second language model (440) may be a more advanced model than the first language model (430). For example, the second language model (440) may be a model trained using a larger dataset than the first language model (430). For example, the second language model (440) may be a language model evaluated by technicians as more advanced than the first language model (430).

[0049] For example, the number of parameters of the first language model (430) may differ from the number of parameters of the second language model (440). For example, the number of parameters of the first language model (430) may be smaller than the number of parameters of the second language model (440). For example, the number of parameters of the second language model (440) may be larger than the number of parameters of the first language model (430). For example, parameters may be used to train the model. For example, parameters may include weights and biases used to process input data and generate output data within the model. For example, the more parameters the model has, the more information the model can store and train complex patterns. For example, the fewer parameters the model has, the more limited information the model can store and the more simple patterns it can train.

[0050] For example, the number of layers in the first language model (430) may differ from the number of layers in the second language model (440). For example, the number of layers in the first language model (430) may be fewer than the number of layers in the second language model (440). For example, the number of layers in the second language model (440) may be greater than the number of layers in the first language model (430). For example, a layer may be described as a step in which the model is processed to generate output data from input data. For example, the more layers there are in the language model, the greater the number of parameters there may be in the language model. For example, the more layers there are in the language model, the more complex patterns the model can train.

[0051] For example, at least one processor (207) can obtain the first text (435) and the second text (445) by providing the first prompt (410) to the first language model (430) and the second language model (440), respectively. For example, at least one processor (207) can label the first text (435) as the correct answer. For example, at least one processor (207) can store metadata that the first text (435) is the correct answer along with the first text (435). For example, at least one processor (207) can store information indicating that the first text (435) is the correct answer along with the first text (435). For example, at least one processor (207) can label the second text (445) as the incorrect answer. For example, at least one processor (207) may store metadata that the second text (445) is an incorrect answer along with the second text (445). For example, at least one processor (207) may store information indicating that the second text (445) is an incorrect answer along with the second text (445). For example, at least one processor (207) may generate a data set by combining the first prompt (410), the first text (435), and the second text (445). For example, at least one processor (207) may store a data set containing the first prompt (410), the first text (435), and the second text (445) in memory (206).

[0052] For example, at least one processor (207) can obtain the first text (435) by providing the first text (435) to the first language model (430). For example, at least one processor (207) can obtain the third text (455) by providing the second language model (440) with a second prompt (e.g., the second prompt (420) of FIG. 4b) that is different from the first prompt (410). For example, the generation of the second prompt and the third text is described in more detail with reference to FIG. 4b.

[0053] Referring to FIG. 4b, at least one processor (207) can obtain a response to a prompt using a plurality of language models. For example, at least one processor (207) can obtain a first text (435) by providing a first prompt (410) to a first language model (430). For example, at least one processor (207) can obtain a third text (455) by providing a second prompt (420) to a second language model (440). For example, the second prompt (420) can be obtained from the first prompt (410). For example, the second prompt (420) may include a prompt similar to the first prompt (410). For example, the second prompt (420) may include a prompt that is different from the first prompt (410) and related to the first prompt (410). For example, the second prompt (420) may be described as a prompt for obtaining an incorrect answer to the first prompt (410). For example, the second prompt (420) may be used to obtain a result different from the result requested by the first prompt (410). For example, at least one processor (207) may obtain or generate the second prompt (420) by providing the first prompt (410) to the first language model (430) or the second language model (440). For example, at least one processor (207) may use the first language model (430) or the second language model (440) to obtain the second prompt (420) for obtaining a result different from the result requested by the first prompt (410), which includes at least a portion of the first prompt (410). For example, a method of obtaining a second prompt (420) to obtain a result different from the result requested by the first prompt (410) may be referred to as an off-topic response method. For example, obtaining the second prompt (420) may be illustrated by referring to Table 2.

[0054] Instruction: Create a new user input using the given user input. (1) The new user input is very related to the given user input but is different. (2) The correct answer for the new user input is superficially as similar as possible to the correct answer for the given user input. (3) However, in reality, the correct answer for the new user input must not be the correct answer for the given user input. Given user input:

[0055] Referring to Table 2, a second prompt (420) generated using the first prompt (410) can be exemplified. For example, the user input given in Table 2 can be described as the first prompt (410).

[0056] For example, the third text (455) may be labeled as an incorrect answer to the first prompt (410). For example, the relevance between the first prompt (410) and the third text (455) may be less than the relevance between the first prompt (410) and the second text (445). For example, the third text (455) may not be appropriate as a response to the first prompt (410). For example, because at least one processor (207) obtains the third text (455) as a response to the second prompt (420) modified from the first prompt (410), the third text (455) may not be appropriate as a response to the first prompt (410). For example, since at least one processor (207) obtains the third text (455) using the second prompt (420) obtained from the first prompt (410), the third text (455) may have little relevance to the first prompt (410).

[0057] For example, at least one processor (207) may generate another data set in combination of the first prompt (410), the first text (435), and the third text (455). For example, at least one processor (207) may store the other data set in memory (206). For example, the other data set may be used to perform training to evaluate a response to the prompt. For example, the other data set may be used to perform training on a third language model to determine which of the first text (435) and the third text (455) is more relevant to the first text (435) in response to the first prompt (410). For example, the other data set may be provided to the third language model.

[0058] For example, an electronic device (100) can generate a data set using a first language model (430) and a second language model (440). For example, the electronic device (100) can obtain a third text (455) using a second language model (440) having a second complexity higher than that of a first language model (430) having a first complexity. For example, the third text (455) may be an incorrect answer. For example, because the third text (455) is generated by the second language model (440) having a second complexity, it may be a more attractive answer than the first text (435). For example, because the electronic device (100) generates the third text (455) using the second language model (440), it can obtain a third text (455) that is more attractive than the first text (435). For example, since the electronic device (100) generates the third text (455) using the second language model (440), it can obtain a third text (455) that is longer than the first text (435). For example, based on generating the third text (455) using the second language model (440), the electronic device (100) can obtain a third text (455) that is more specific than the first text (435). For example, the electronic device (100) can train the third language model to overcome or mitigate multiple biases using the third text (455). For example, the electronic device (100) can train the third language model to evaluate responses to prompts using incorrect answers that are longer than correct answers. For example, the electronic device (100) can train the third language model to evaluate responses to prompts using incorrect answers that are more specific than correct answers. For example, generating a dataset to be provided to the third language model using multiple language models can help to overcome, mitigate, or reduce multiple biases of the language model.For example, generating a dataset for training the third language model using multiple language models can be effective in overcoming, mitigating, or reducing multiple biases of the language model. For example, generating the dataset can improve the robustness of the third language model.

[0059] For example, at least one processor (207) can identify whether the first prompt (410) is different from the second prompt (420) after generating the second prompt (420). For example, if at least one processor (207) obtains the third text (455) using the second prompt (420) which is identical to the first prompt (410), it may be required to identify whether the first prompt (410) is different from the second prompt (420) because the third text (455) can be the correct answer to the first prompt (410). The operation of identifying whether the first prompt (410) is different from the second prompt (420) is described and illustrated in more detail with reference to FIG. 5.

[0060] FIG. 5 is a flowchart illustrating the operation of an electronic device that identifies whether the first prompt and the second prompt are different.

[0061] Referring to FIG. 5, in operation 510, at least one processor (207) can identify whether the first prompt (410) is different from the second prompt (420). At least one processor (207) can execute operation 520 under conditions where the first prompt (410) is different from the second prompt (420), and execute operation 530 under conditions where the first prompt (410) is not different from the second prompt (420). For example, at least one processor (207) can identify whether the first prompt (410) is different from the second prompt (420) using the first language model (430) or the second language model (440). For example, at least one processor (207) can identify whether the first prompt (410) is different from the second prompt (420) by providing the first prompt (410) and the second prompt (420) to the first language model (430). For example, at least one processor (207) can identify whether the first prompt (410) is different from the second prompt (420) by providing the first prompt (410) and the second prompt (420) to the second language model (440). For example, at least one processor (207) can determine that the first prompt (410) and the second prompt (420) are different based on the determination that the first prompt (410) and the second prompt (420) can obtain different responses. For example, at least one processor (207) may determine that the first prompt (410) is different from the second prompt (420) based on the determination that the third text (455), which is a response to the second prompt (420), is not appropriate as a response to the first prompt (410).

[0062] In operation 520, at least one processor (207) can obtain a third text (455) for the second prompt (420) based on a determination that the first prompt (410) is different from the second prompt (420). For example, at least one processor (207) can obtain the third text (455) by providing the second prompt (420) to the second language model (440) based on a determination that the first prompt (410) is different from the second prompt (420).

[0063] In operation 530, at least one processor (207) may refrain from, bypass, or block obtaining the third text (455) for the second prompt (420) based on the determination that the first prompt (410) is not different from the second prompt (420). For example, at least one processor (207) may determine that the first prompt (410) is not different from the second prompt (420). For example, at least one processor (207) may determine that the first prompt (410) is not different from the second prompt (420) if the response to the first prompt (410) is (substantially) the same as the response to the second prompt (420). For example, at least one processor (207) may stop obtaining the third text (455) using the second language model (440) based on a determination that the first prompt (410) is not different from the second prompt (420). For example, at least one processor (207) may stop providing the second prompt (420) to the third text (455) based on a determination that the first prompt (410) is not different from the second prompt (420). However, it is not limited thereto. For example, at least one processor (207) may refrain from, bypass, or block obtaining the third text (455) for the second prompt (420) and generating another data set based on a determination that the first prompt (410) is not different from the second prompt (420). For example, at least one processor (207) can obtain a third text (455) by providing the second prompt (420) to the second language model (440) based on the determination that the first prompt (410) is not different from the second prompt (420).For example, at least one processor (207) may refrain from, bypass, or block generating another data set composed of the first prompt (410), the first text (435), and the third text (455) after obtaining the third text (455) based on the determination that the first prompt (410) is not different from the second prompt (420).

[0064] In operation 540, at least one processor (207) may generate another data set in combination of the first prompt (410), the first text (435), and the third text (455) based on the determination that the first prompt (410) and the second prompt (420) are different. For example, at least one processor (207) may generate the other data set after obtaining the third text (455) using the second language model (440) based on the determination that the first prompt (410) and the second prompt (420) are different. For example, at least one processor (207) may store the other data set in memory (206). For example, at least one processor (207) may use the other data set to train the third language model. For example, at least one processor (207) may provide the other data set for the third language model to train.

[0065] At least one processor (207) may generate a data set to be provided to a third language model. Unlike as described above, at least one processor (207) may obtain or generate from the first prompt a second prompt (e.g., the second prompt (720) of FIG. 7) that causes a language model (e.g., the language model (730) of FIG. 7) to include a result (e.g., a response or text) and an error resulting from a request of the first prompt (e.g., the first prompt (710) of FIG. 7). For example, the acquisition of the second prompt that causes the language model to include the error is described and illustrated in more detail with reference to FIG. 6.

[0066] FIG. 6 is a flowchart illustrating the operation of an electronic device that generates a data set using multiple prompts. This method may be executed by the electronic device (100) illustrated in FIG. 2 or by at least one processor (207) of the electronic device (100).

[0067] Referring to FIG. 6, in operation 610, at least one processor (207) can obtain a first text (e.g., the first text (735) of FIG. 7) for a first prompt (e.g., the first prompt (710) of FIG. 7) using a language model (e.g., the language model (730) of FIG. 7). For example, the language model may include a first language model (430). For example, the language model may include a second language model (440). For example, at least one processor (207) can obtain or generate a first text (e.g., the first text (735) of FIG. 7) in response to the first prompt by providing the first prompt (e.g., the first prompt (710) of FIG. 7) to the language model (e.g., the language model (730) of FIG. 7).

[0068] In operation 620, at least one processor (207) may generate a second prompt (e.g., second prompt (720) of FIG. 7) to provide the result requested by the first prompt (e.g., first prompt (710) of FIG. 7) along with an error. For example, at least one processor (207) may obtain the second prompt (e.g., second prompt (720) of FIG. 7) by providing the first prompt (e.g., first prompt (710) of FIG. 7) to a language model (e.g., language model (730) of FIG. 7). For example, the second prompt may be described as a prompt that causes the language model to include an error in the response to the second prompt.

[0069] In operation 630, at least one processor (207) can obtain a second text (e.g., second text (745) of FIG. 7) for a second prompt (e.g., second prompt (720) of FIG. 7) using a language model (e.g., language model (730) of FIG. 7). For example, at least one processor (207) can obtain or generate the second text in response to the second prompt by providing the second prompt to the language model.

[0070] In operation 640, at least one processor (207) may generate a data set comprising a first prompt (e.g., the first prompt (710) of FIG. 7), a first text (e.g., the first text (735) of FIG. 7), and a second text (e.g., the second text (745) of FIG. 7). For example, at least one processor (207) may generate a data set composed of the first prompt, the first text, and the second text. For example, at least one processor (207) may store the data set in memory (206). For example, the data set may be used to train a third language model. For example, the data set may be provided to a third language model for evaluating a response to a prompt. For example, an electronic device (100) may include the third language model.

[0071] According to one embodiment, at least one processor (207) can obtain a third text for a second prompt (e.g., the second prompt (720) of FIG. 7) by using a second language model (440) having a second complexity higher than the first complexity of the first language model (430). For example, at least one processor (207) can obtain the third text by providing the second prompt (e.g., the second prompt (720) of FIG. 7) to the second language model (440) having a second complexity higher than the first complexity of the first language model (430). For example, the third text may contain an error. For example, the third text may contain a result and / or an error requested by the first prompt (e.g., the first prompt (710) of FIG. 7). For example, the error may be described with reference to FIG. 7. For example, at least one processor (207) may generate another data set comprising a first prompt (e.g., the first prompt (710) of FIG. 7), a first text (e.g., the first text (735) of FIG. 7), and the third text. For example, at least one processor (207) may generate another data set composed of the first prompt, the first text, and the third text. For example, the other data set may be used to evaluate a response to the prompt. For example, the other data set may be provided to a third language model trained to evaluate a response to the prompt. For example, at least one processor (207) may store the other data set in memory (206).

[0072] At least one processor (207) may generate or obtain a second prompt (e.g., a second prompt (720) of FIG. 7) from a first prompt (e.g., a first prompt (710) of FIG. 7) that causes a language model (e.g., a language model (730) of FIG. 7) to include the result and error for the first prompt together. For example, the acquisition of the second prompt is described and illustrated in more detail with reference to FIG. 7.

[0073] FIG. 7 illustrates an exemplary operation of an electronic device that acquires text using multiple prompts.

[0074] Referring to FIG. 7, at least one processor (207) can obtain a second prompt (720) using a first prompt (710). For example, at least one processor (207) can obtain a second prompt (720) from the first prompt (710) using a language model (730). For example, the language model (730) may include a first language model (430) or a second language model (440). For example, the second prompt (720) may include at least a portion of the first prompt (710). For example, the second prompt (720) may cause the language model (730) to output the result and error requested by the first prompt (710). For example, a method of generating an incorrect answer using a second prompt (720) that includes at least one word causing the error may be referred to as an erroneous response method.

[0075] For example, the second prompt (720) may include a prompt that causes the language model (730) to include the result and error requested by the first prompt (710). For example, the error may include the second text (745) containing incorrect (or false) information. For example, the error may include the second text (745) containing an incomplete answer. For example, the error may include the second text (745) containing unnecessary information. For example, the error may include the second text (745) omitting important information. For example, the error may include the second text (745) deviating from the instruction. For example, types of errors may be described below.

[0076] Instructions: Provide an incorrect answer that is as similar as possible to the correct answer but retains the unhelpful essence. Explain why the response appears correct on the surface but is actually incorrect. Describe the type of bird. Response: The kookaburra is a medium-sized bird that inhabits Korea. It is often referred to as the "laughing bird" due to a distinctive ringing sound reminiscent of loud laughter. The kookaburra has a white or cream-colored body, head, wings, and back with brown shades. The kookaburra feeds on insects, reptiles, small mammals, and birds. Description: The above response is incorrect because it contains false information that the kookaburra inhabits Korea.

[0077] Referring to Table 3, an example of a second prompt (720) that causes the language model (730) to contain incorrect information can be described. For example, the instruction in Table 3 can be described as the second prompt (720). For example, the second prompt (720) may include at least one word that causes the language model (730) to cause the second text (745) to contain false information. For example, the response in Table 3 can be described as the second text (745). For example, the description in Table 3 can be described as an error regarding the response in Table 3.

[0078] Provide an incorrect answer that is as similar as possible to the correct answer but retains the unhelpful nature. Write an incomplete response to the input. The response deals with only part of the input and may not fully explain the instruction. Explain why the response appears correct on the surface but is actually incorrect. Describe the type of bird. Response: The kookaburra is a medium-sized bird native to Australia. Description: The above response contains the name of the bird type but does not provide additional information. The above response is unhelpful because, despite containing the name of the bird type, it does not describe the bird type.

[0079] Referring to Table 4, an example of a second prompt (720) that causes the language model (730) to generate an incomplete answer can be described. For example, the instruction in Table 4 can be described as the second prompt (720). For example, the second prompt (720) may include at least one word that causes the language model (730) to generate an incomplete result. For example, the response in Table 4 can be described as the second text (745). For example, the description in Table 4 can be described as an error in the response of Table 4.

[0080] Instruction: Provide an incorrect answer that is as similar as possible to the correct answer but retains the unhelpful essence. Write an incomplete and unnecessary response to the input. The response deals with only a part of the input and may start talking about a slightly different topic. Explain why the response appears correct on the surface but is actually incorrect. Describe the type of bird. Response: The kookaburra is a medium-sized bird native to Australia. Australia is home to many birds, which is why it attracts many tourists from around the world. From rainforests to deserts, Australia's diverse landscapes provide habitats for various birds and have become a paradise for birdwatchers and nature lovers. In addition to birds, Australia's indigenous culture, the Great Barrier Reef, and landmarks like Uluru contribute to its appeal as a top tourist destination. Description: The above response includes the name of the bird type but starts talking about the country where the bird lives. The above response differs from the instructions and does not fully describe the bird type, so the above response is unhelpful.

[0081] Referring to Table 5, an example of a second prompt (720) that causes the language model (730) to include unnecessary information can be described. For example, the instruction in Table 5 can be described as the second prompt (720). For example, the second prompt (720) may include at least one word (e.g., another topic in Table 5) that causes the language model (730) to include another result that is different from the result requested by the first prompt (710) in the second text (745). For example, the response in Table 5 can be described as the second text (745). For example, the description in Table 5 can be described as an error regarding the response in Table 5.

[0082] Instructions: Provide an incorrect answer that is as similar as possible to the correct answer but retains the unhelpful essence. Create a list of key points that must be mentioned in the correct answer, and use the list to create an unhelpful answer. You may omit one or more items from the list to create an unhelpful or incorrect answer. Explain why the response appears correct on the surface but is actually not. Describe the type of bird. List: The correct answer mentions the name of a specific type of bird. The correct answer mentions details or characteristics of a specific type of bird. Response: A type of bird is a creature that has feathers, a beak, and can fly. A type of bird is characterized by feathers, a toothless beak-shaped jaw, laying hard-shelled eggs, a high metabolic rate, a four-chambered heart, and a strong, light skeleton. From a taxonomic perspective, the number of known bird species is approximately 10,906. Description: The above response does not mention a specific type of bird. Instead, the above response covers general information about birds.

[0083] Referring to Table 6, an example of a second prompt (720) that causes the language model (730) to omit important information can be described. For example, the instructions in Table 6 can be described as the second prompt (720). For example, the second prompt (720) may include at least one word that causes the language model (730) to omit part of the result requested by the first prompt (710) in the second text (745). For example, the list in Table 6 can be described as important information that must be included in the response to the prompt. For example, the response in Table 6 can be described as the second text (745). For example, the description in Table 6 can be described as an error in the response of Table 6.

[0084] Instruction: Provide an incorrect answer that is as similar as possible to the correct answer but maintains the unhelpful essence. The response should appear to be an answer to the question, but it must deviate slightly from the essence. Explain why the response appears correct on the surface but is actually not. Describe a type of bird. Response: A type of bird is a creature that has feathers, a beak, and can fly. Birds are characterized by feathers, a toothless beak-shaped jaw, the laying of hard-shelled eggs, a high metabolic rate, a four-chambered heart, and a strong, light skeleton. From a taxonomic perspective, the number of known bird species is approximately 10,906. Description: The above response must mention birds. However, while the above instructions require a type of bird, the above response covers general information about birds.

[0085] Referring to Table 7, an example of a second prompt (720) that causes the language model (730) to deviate from an instruction can be described. For example, the instruction in Table 7 can be described as the second prompt (720). For example, the second prompt (720) may include at least one word (e.g., deviation in Table 7) that causes the language model (730) to cause the second text (745) to contain a different result for another request that is different from the request for the first prompt (710). For example, the response in Table 7 can be described as the second text (745). For example, the description in Table 7 can be described as an error in the response of Table 7.

[0086] For example, at least one processor (207) may obtain an incorrect answer by using the second prompt (720). For example, at least one processor (207) may obtain a second text (745) by providing the language model (730) with the second prompt (720), which contains at least one word that causes the language model (730) to include at least a part of the first prompt (710) and an error. For example, the second text (745) may be described as an incorrect response to the first text (735). For example, the second text (745) may not contain a result for the first prompt (710). For example, at least one processor (207) may obtain a second text (745) containing a result and / or an error requested by the first prompt (710). For example, the second text (745) may not contain a result for the first prompt (710).

[0087] For example, the electronic device (100) may generate a data set using a first prompt (710) and a second prompt (720). For example, the electronic device (100) may obtain a second text (745) by providing the language model (730) with a second prompt (720) containing at least one word that causes the language model (730) to contain an error. For example, the second text (745) may be an incorrect answer because it contains an error. For example, the second text (745) may contain an error that is difficult for the language model to identify. For example, it may be difficult for the language model to identify an error in the second text (745) obtained from the second prompt (720) containing at least one word that causes an error. For example, the electronic device (100) may generate a data set to provide to a third language model trained to evaluate a response to a prompt, based on generating the second text (745) containing an error. For example, the electronic device (100) can train a third language model to identify errors using a second text (745) containing errors. For example, generating a dataset to be provided to the third language model using the second text (745) containing errors can help improve the third language model's ability to identify errors. For example, generating a dataset to be provided to the third language model using the second text (745) containing errors can be effective in improving the third language model's ability to identify errors. For example, generating the dataset can improve the robustness of the third language model.

[0088] For example, at least one processor (207) can identify whether the second text (745) contains an error. For example, at least one processor (207) can generate a data set containing the second text (745) based on a determination that the second text (745) does not contain an error. For example, the operation of identifying whether the second text (745) contains an error is described and illustrated in more detail with reference to FIG. 8.

[0089] FIG. 8 is a flowchart illustrating the operation of an electronic device for identifying whether a second text contains an error. This method may be executed by the electronic device (100) illustrated in FIG. 2 or by at least one processor (207) of the electronic device (100).

[0090] Referring to FIG. 8, in operation 810, at least one processor (207) can identify whether the second text (745) contains the result and error requested by the first prompt (710). For example, at least one processor (207) can execute operation 820 under the condition that the second text (745) contains the result and error requested by the first prompt (710), and execute operation 830 under the condition that the second text (745) does not contain the result and error requested by the first prompt (710). For example, at least one processor (207) can use a language model (730) to identify whether the second text (745) contains the result and error requested by the first prompt (710). For example, at least one processor (207) can identify whether the second text (745) contains the result and error requested by the first prompt (710) by providing the second text (745) to the language model (730).

[0091] In operation 820, at least one processor (207) may generate a data set based on a determination that the second text (745) contains the result and error requested by the first prompt (710). For example, the data set may include the first prompt (710), the first text (735), and the second text (745). For example, the data set may include information indicating that the first text (735) is the correct answer and information indicating that the second text (745) is the incorrect answer. For example, the data set may be labeled to indicate the correct answer and the incorrect answer to the first text (735) and the second text (745), respectively.

[0092] In operation 830, at least one processor (207) may refrain from, bypass, or block the generation of the data set based on a determination that the second text (745) does not contain the result and the error. For example, if the second text (745) does not contain the error, at least one processor (207) may refrain from generating the data set using the second text (745) because the second text (745) is not an incorrect answer to the first prompt (710). For example, at least one processor (207) may refrain from generating the data set containing the second text (745) based on a determination that the second text (745) does not contain an error.

[0093] An electronic device as described above may include a memory for storing instructions. The electronic device may include at least one processor. The instructions may cause the electronic device to obtain a first text for a prompt by providing a prompt to a first language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to obtain a second text for the prompt, having less relevance than the relevance between the prompt and the first text, by using a second language model different from the first language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate a data set in which the prompt, the first text, and the second text are combined when executed individually or collectively by the at least one processor.

[0094] According to one embodiment, the first language model may have a first complexity. The second language model may have a second complexity higher than the first complexity.

[0095] According to one embodiment, the prompt may be a first prompt. The instructions may cause the electronic device to obtain a second prompt that includes at least a portion of the first prompt and obtains a result different from the result requested by the first prompt, by using the first language model or the second language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to obtain a third text by providing the second prompt to the second language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate another data set in which the first prompt, the first text, and the third text are combined when executed individually or collectively by the at least one processor. The correlation between the second prompt and the third text may be less than the correlation between the first prompt and the first text.

[0096] According to one embodiment, the instructions may cause the electronic device to identify whether the first prompt is different from the second prompt by using the first language model or the second language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to obtain the third text for the second prompt based on the determination that the first prompt is different from the second prompt when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate the other data set in combination of the first prompt, the first text, and the third text based on the determination that the first prompt is different from the second prompt when executed individually or collectively by the at least one processor. The above instructions may cause the electronic device to refrain from obtaining the third text for the second prompt based on a determination that the first prompt is not different from the second prompt when executed individually or collectively by the at least one processor.

[0097] According to one embodiment, the instructions may cause the electronic device to identify whether the first prompt is different from the second prompt by using the first language model or the second language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to obtain the third text for the second prompt based on the determination that the first prompt is different from the second prompt when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate the other data set in combination of the first prompt, the first text, and the third text based on the determination that the first prompt is different from the second prompt when executed individually or collectively by the at least one processor. The above instructions may cause the electronic device to refrain from generating the other data set based on the determination that the first prompt is not different from the second prompt when executed individually or collectively by the at least one processor.

[0098] According to one embodiment, the second prompt may include at least one word that causes the second language model to include an error in the third text.

[0099] According to one embodiment, the data set may be used to perform training on a third language model to determine the text among the first text and the second text that is more relevant to the prompt in response to the prompt.

[0100] A method performed by an electronic device as described above may include the operation of obtaining a first text for said prompt by providing a prompt to a first language model. The method may include the operation of obtaining a second text for said prompt, having less relevance than the relevance between said prompt and said first text, by using a second language model different from said first language model. The method may include the operation of generating a data set in which said prompt, said first text, and said second text are combined.

[0101] According to one embodiment, the first language model may have a first complexity. The second language model may have a second complexity higher than the first complexity.

[0102] According to one embodiment, the prompt may be a first prompt. The method may include the operation of obtaining a second prompt that includes at least a portion of the first prompt and obtains a result different from the result requested by the first prompt, using the first language model or the second language model. The method may include the operation of obtaining a third text by providing the second prompt to the second language model. The method may include the operation of generating another data set in which the first prompt, the first text, and the third text are combined. The correlation between the second prompt and the third text may be less than the correlation between the first prompt and the first text.

[0103] According to one embodiment, the method may include an operation of identifying whether the first prompt is different from the second prompt using the first language model or the second language model. The method may include an operation of obtaining the third text for the second prompt based on the determination that the first prompt is different from the second prompt. The method may include an operation of generating the other data set in which the first prompt, the first text, and the third text are combined based on the determination that the first prompt is different from the second prompt. The method may include an operation of refraining from obtaining the third text for the second prompt based on the determination that the first prompt is not different from the second prompt.

[0104] According to one embodiment, the method may include an operation of identifying whether the first prompt is different from the second prompt using the first language model or the second language model. The method may include an operation of obtaining the third text for the second prompt based on the determination that the first prompt is different from the second prompt. The method may include an operation of generating the other data set in which the first prompt, the first text, and the third text are combined based on the determination that the first prompt is different from the second prompt. The method may include an operation of refraining from generating the other data set based on the determination that the first prompt is not different from the second prompt.

[0105] According to one embodiment, the second prompt may include at least one word that causes the second language model to include an error in the third text.

[0106] According to one embodiment, the data set may be used to perform training on a third language model to determine the text among the first text and the second text that is more relevant to the prompt in response to the prompt.

[0107] In a computer-readable storage medium storing one or more programs as described above, the one or more programs may include instructions that cause the electronic device to obtain a first text for said prompt by providing a prompt to a first language model when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to obtain a second text for said prompt, which has less relevance than the relevance between said prompt and said first text, by using a second language model different from said first language model when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to generate a data set in which said prompt, said first text, and said second text are combined when executed by the electronic device.

[0108] According to one embodiment, the first language model may have a first complexity. The second language model may have a second complexity higher than the first complexity.

[0109] According to one embodiment, the prompt may be a first prompt. The one or more programs may include instructions that cause the electronic device to obtain a second prompt that includes at least a portion of the first prompt and obtains a result different from the result requested by the first prompt, by using the first language model or the second language model. The one or more programs may include instructions that cause the electronic device to obtain a third text by providing the second prompt to the second language model. The one or more programs may include instructions that cause the electronic device to generate another data set in which the first prompt, the first text, and the third text are combined. The correlation between the second prompt and the third text may be less than the correlation between the first prompt and the first text.

[0110] According to one embodiment, the one or more programs may include instructions that cause the electronic device to identify whether the first prompt is different from the second prompt by using the first language model or the second language model. The one or more programs may include instructions that cause the electronic device to obtain the third text for the second prompt based on the determination that the first prompt is different from the second prompt. The one or more programs may include instructions that cause the electronic device to generate the other data set in which the first prompt, the first text, and the third text are combined based on the determination that the first prompt is different from the second prompt. The one or more programs may include instructions that cause the electronic device to refrain from obtaining the third text for the second prompt based on the determination that the first prompt is not different from the second prompt.

[0111] According to one embodiment, the one or more programs may include instructions that cause the electronic device to identify whether the first prompt is different from the second prompt by using the first language model or the second language model. The one or more programs may include instructions that cause the electronic device to obtain the third text for the second prompt based on the determination that the first prompt is different from the second prompt. The one or more programs may include instructions that cause the electronic device to generate the other data set in which the first prompt, the first text, and the third text are combined based on the determination that the first prompt is different from the second prompt. The one or more programs may include instructions that cause the electronic device to refrain from generating the other data set based on the determination that the first prompt is not different from the second prompt.

[0112] According to one embodiment, the second prompt may include at least one word that causes the second language model to include an error in the third text.

[0113] According to one embodiment, the data set may be used to perform training on a third language model to determine the text among the first text and the second text that is more relevant to the prompt in response to the prompt.

[0114] An electronic device as described above may include a memory for storing instructions. The electronic device may include at least one processor. When the instructions are executed individually or collectively by the at least one processor, the electronic device may be caused to obtain a first text for a first prompt using a language model. When the instructions are executed individually or collectively by the at least one processor, the electronic device may be caused to generate a second prompt to provide the result requested by the first prompt along with an error. When the instructions are executed individually or collectively by the at least one processor, the electronic device may be caused to obtain a second text for the second prompt using the language model. When the instructions are executed individually or collectively by the at least one processor, the electronic device may be caused to generate a data set including the first prompt, the first text, and the second text.

[0115] According to one embodiment, the second prompt may include at least one word that causes the language model to include false information in the second text.

[0116] According to one embodiment, the second prompt may include at least one word that causes the language model to produce an incomplete result.

[0117] According to one embodiment, the second prompt may include at least one word that causes the language model to further include a different result different from the result requested by the first prompt.

[0118] According to one embodiment, the second prompt may include at least one word that causes the language model to omit part of the result requested by the first prompt, such that the second text omits the result.

[0119] According to one embodiment, the second prompt may include at least one word that causes the language model to include a different result for a different request that is different from the request by the first prompt.

[0120] According to one embodiment, the instructions may cause the electronic device to identify whether the second text includes the result and the error requested by the first prompt when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate the data set based on the determination that the second text includes the result and the error when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to refrain from generating the data set based on the determination that the second text does not include the result and the error when executed individually or collectively by the at least one processor.

[0121] According to one embodiment, the language model may be a first language model. The instructions may cause the electronic device to obtain a third text for the second prompt by using a second language model having a second complexity higher than the first complexity of the first language model when executed individually or collectively by the at least one processor. The instructions may cause the electronic device to generate another data set including the first prompt, the first text, and the third text when executed individually or collectively by the at least one processor.

[0122] According to one embodiment, the data set may be used to perform training on a third language model to determine, in response to the first prompt, the text among the first text and the second text that is more relevant to the first prompt.

[0123] A method performed by an electronic device as described above may include an operation of obtaining a first text for a first prompt using a language model. The method may include an operation of generating a second prompt to provide the result requested by the first prompt along with an error. The method may include an operation of obtaining a second text for the second prompt using the language model. The method may include an operation of generating a data set including the first prompt, the first text, and the second text.

[0124] According to one embodiment, the second prompt may include at least one word that causes the language model to include false information in the second text.

[0125] According to one embodiment, the second prompt may include at least one word that causes the language model to produce an incomplete result.

[0126] According to one embodiment, the second prompt may include at least one word that causes the language model to further include a different result different from the result requested by the first prompt.

[0127] According to one embodiment, the second prompt may include at least one word that causes the language model to omit part of the result requested by the first prompt, such that the second text omits the result.

[0128] According to one embodiment, the second prompt may include at least one word that causes the language model to include a different result for a different request that is different from the request by the first prompt.

[0129] According to one embodiment, the method may include an operation of identifying whether the second text includes the result and the error requested by the first prompt. The method may include an operation of generating the data set based on the determination that the second text includes the result and the error. The method may include an operation of refraining from generating the data set based on the determination that the second text does not include the result and the error.

[0130] According to one embodiment, the language model may be a first language model. The method may include the operation of obtaining a third text for the second prompt using a second language model having a second complexity higher than the first complexity of the first language model. The method may include the operation of generating another data set including the first prompt, the first text, and the third text.

[0131] According to one embodiment, the data set may be used to perform training on a third language model to determine, in response to the first prompt, the text among the first text and the second text that is more relevant to the first prompt.

[0132] In a computer-readable storage medium storing one or more programs as described above, the one or more programs may include instructions that cause the electronic device to obtain a first text for a first prompt by using a language model when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to generate a second prompt to provide a result requested by the first prompt along with an error when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to obtain a second text for the second prompt by using the language model when executed by the electronic device. The one or more programs may include instructions that cause the electronic device to generate a data set including the first prompt, the first text, and the second text when executed by the electronic device.

[0133] According to one embodiment, the second prompt may include at least one word that causes the language model to include false information in the second text.

[0134] According to one embodiment, the second prompt may include at least one word that causes the language model to produce an incomplete result.

[0135] According to one embodiment, the second prompt may include at least one word that causes the language model to further include a different result different from the result requested by the first prompt.

[0136] According to one embodiment, the second prompt may include at least one word that causes the language model to omit part of the result requested by the first prompt, such that the second text omits the result.

[0137] According to one embodiment, the second prompt may include at least one word that causes the language model to include a different result for a different request that is different from the request by the first prompt.

[0138] According to one embodiment, the one or more programs may include instructions that cause the electronic device to identify whether the second text contains the result and the error requested by the first prompt. The one or more programs may include instructions that cause the electronic device to generate the data set based on the determination that the second text contains the result and the error. The one or more programs may include instructions that cause the electronic device to refrain from generating the data set based on the determination that the second text does not contain the result and the error.

[0139] According to one embodiment, the language model may be a first language model. The one or more programs may include instructions that cause the electronic device to obtain a third text for the second prompt by using a second language model having a second complexity higher than the first complexity of the first language model. The one or more programs may include instructions that cause the electronic device to generate another data set including the first prompt, the first text, and the third text.

[0140] According to one embodiment, the data set may be used to perform training on a third language model to determine, in response to the first prompt, the text among the first text and the second text that is more relevant to the first prompt.

[0141] The device described above may be implemented as a hardware component, a software component, and / or a combination of a hardware component and a software component. For example, the device and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.

[0142] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be embodied in any type of machine, component, physical device, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0143] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. In this case, the medium may continuously store a computer-executable program, or temporarily store it for execution or download. Additionally, the medium may be various recording or storage means in the form of a single or several combined hardware, and may not be limited to a medium directly connected to a computer system but may exist distributed over a network. Examples of media may include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and media configured to store program instructions, including ROM, RAM, and flash memory. Additionally, other examples of media may include recording or storage media managed by app stores that distribute applications or sites and servers that supply or distribute various other software.

[0144] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results may be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0145] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

1. In an electronic device, Memory comprising one or more storage media and storing instructions; and It includes at least one processor comprising processing circuitry, and When the above instructions are executed individually or collectively by the at least one processor, By providing a prompt to a first language model, a first text for the prompt is obtained, and Using a second language model different from the first language model, for the prompt, a second text having less relevance than the relevance between the prompt and the first text, and To generate a data set in which the above prompt, the above first text, and the above second text are combined, causing the above electronic device, Electronic device.

2. In claim 1, the first language model is, With first complexity, and The above second language model is, Having a second complexity higher than the first complexity mentioned above, Electronic device.

3. In claim 1, the prompt is, It is the first prompt, and When the above instructions are executed individually or collectively by the at least one processor, Using the first language model or the second language model, obtain a second prompt that includes at least a portion of the first prompt and obtains a result different from the result requested by the first prompt, and By providing the second prompt to the second language model, a third text is obtained, and To generate another data set by combining the first prompt, the first text, and the third text, Causing the above electronic device, The relationship between the above second prompt and the above third text is, Less related than the relationship between the first prompt and the first text above, Electronic device.

4. In Claim 3, When the above instructions are executed individually or collectively by the at least one processor, Using the first language model or the second language model, identify whether the first prompt is different from the second prompt, and Based on the determination that the above first prompt is different from the above second prompt: Obtain the third text for the second prompt above, and Generating the other data set by combining the first prompt, the first text, and the third text, and Based on the determination that the first prompt is not different from the second prompt, to refrain from obtaining the third text for the second prompt, causing the above electronic device, Electronic device.

5. In Claim 3, When the above instructions are executed individually or collectively by the at least one processor, Using the first language model or the second language model, identify whether the first prompt is different from the second prompt, and Based on the determination that the above first prompt is different from the above second prompt: Obtain the third text for the second prompt above, and Generating the other data set by combining the first prompt, the first text, and the third text, and Based on the determination that the first prompt is not different from the second prompt, to refrain from generating the other data set, causing the above electronic device, Electronic device.

6. In claim 3, the second prompt is, The third text above includes at least one word that causes the second language model to contain an error, Electronic device.

7. In claim 1, the data set is, Used to perform training on a third language model to determine the text among the first text and the second text that is more relevant to the prompt in response to the prompt. Electronic device.

8. In an electronic device, Memory comprising one or more storage media and storing instructions; It includes at least one processor comprising processing circuitry, and When the above instructions are executed individually or collectively by the at least one processor, Using a language model, obtain the first text for the first prompt, and Generate a second prompt to provide the result requested by the first prompt, along with an error, and Using the above language model, obtain the second text for the above second prompt, and To generate a data set including the first prompt, the first text, and the second text, causing the above electronic device, Electronic device.

9. In claim 8, the second prompt is, at least one word that causes the language model to include false information, Electronic device.

10. In claim 8, the second prompt is, including at least one word that causes the language model to produce an incomplete result, Electronic device.

11. In claim 8, the second prompt is, The above second text includes at least one word that causes the language model to include another result different from the result requested by the first prompt, Electronic device.

12. In claim 8, the second prompt is, The second text includes at least one word that causes the language model to omit a part of the result requested by the first prompt, Electronic device.

13. In claim 8, the second prompt is, The second text includes at least one word that causes the language model to include a different result for a different request that is different from the request by the first prompt, Electronic device.

14. In Claim 8, When the above instructions are executed individually or collectively by the at least one processor, Identifying whether the above second text includes the above result and the above error requested by the above first prompt, and Based on the determination that the second text above includes the result and the error, the data set is generated, and To refrain from generating the data set based on the determination that the above second text does not include the above result and the above error, causing the above electronic device, Electronic device.

15. In claim 8, the language model is, It is a first language model, and When the above instructions are executed individually or collectively by the at least one processor, Using a second language model having a second complexity higher than the first complexity of the first language model, a third text for the second prompt is obtained, and To generate another data set including the first prompt, the first text, and the third text, causing the above electronic device, Electronic device.

16. In claim 8, the data set is, Used to perform training on a third language model to determine, in response to the first prompt, the text between the first text and the second text that is more relevant to the first prompt. Electronic device.

17. In a non-transient computer-readable storage medium storing one or more programs, said one or more programs, when executed by an electronic device, By providing a prompt to a first language model, a first text for the prompt is obtained, and Using a second language model different from the first language model, for the prompt, a second text having less relevance than the relevance between the prompt and the first text, and To generate a data set in which the above prompt, the above first text, and the above second text are combined, Including instructions that cause the above electronic device, Non-transient computer-readable storage media.

18. In claim 17, the first language model is, With first complexity, and The above second language model is, Having a second complexity higher than the first complexity mentioned above, Non-transient computer-readable storage media.

19. In claim 17, the prompt is, It is the first prompt, and When the above one or more programs are executed by the electronic device, Using the first language model or the second language model, obtain a second prompt that includes at least a portion of the first prompt and obtains a result different from the result requested by the first prompt, and By providing the second prompt to the second language model, a third text is obtained, and To generate another data set by combining the first prompt, the first text, and the third text, Includes instructions that cause the above electronic device, The relationship between the above second prompt and the above third text is, Less related than the relationship between the first prompt and the first text above, Non-transient computer-readable storage media.

20. In Claim 19, When the above one or more programs are executed by the electronic device, Using the above second language model, identify whether the above first prompt is different from the above second prompt, and Based on the determination that the above first prompt is different from the above second prompt: Obtain the third text for the second prompt above, and Generating the other data set by combining the first prompt, the first text, and the third text, and Based on the determination that the first prompt is not different from the second prompt, to refrain from obtaining the third text for the second prompt, Including instructions that cause the above electronic device, Non-transient computer-readable storage media.

Images