CN119106737B

CN119106737B - Optimization method, device, equipment and storage medium for text prompts of large models

Info

Publication number: CN119106737B
Application number: CN202411310967.XA
Authority: CN
Inventors: 张欢; 张红阳; 戴岱; 武治; 吴华
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2024-09-19
Filing date: 2024-09-19
Publication date: 2025-10-10
Anticipated expiration: 2044-09-19
Also published as: CN119106737A

Abstract

The disclosure provides a text prompt optimization method, device, equipment and storage medium for a large model, relates to the technical field of artificial intelligence, and particularly relates to the technical fields of large model, intelligent search and the like. The method comprises the steps of inputting a first text prompt and a first task sample into a large model to obtain a processing result of the large model for the first task sample, evaluating the processing result by the large model to obtain an evaluation result, and optimizing the first text prompt to obtain a second text prompt based on a reason which is recorded in the evaluation result and causes the problem when the evaluation result indicates that the processing result has the problem.

Description

Optimization method, device, equipment and storage medium for text prompt of large model

Technical Field

The disclosure relates to the technical field of artificial intelligence, in particular to the technical fields of large models, intelligent searching and the like.

Background

In the field of current artificial intelligence, and in particular in the application of large models, while these models are capable of handling a variety of complex tasks, their final performance is often affected by Prompt.

Prompt, an instruction input into the model, can directly determine the output quality of a large model. Therefore, how to design and optimize promt becomes a key step in improving the performance of large model tasks. In recent years, demand for automation of the promtt engineering has become stronger, and a method of automatically optimizing promtt has thus become of great interest.

Disclosure of Invention

The disclosure provides a text prompt optimization method, device and equipment for a large model and a storage medium.

According to an aspect of the present disclosure, there is provided a method for optimizing text prompts of a large model, including:

Inputting a first text prompt and a first task sample into a large model to obtain a processing result of the large model aiming at the first task sample;

adopting the large model to evaluate the processing result to obtain an evaluation result;

and if the evaluation result shows that the processing result has a problem, optimizing the first text prompt to obtain a second text prompt based on the reason which is recorded in the evaluation result and causes the problem.

According to another aspect of the present disclosure, there is provided an optimizing apparatus of text prompts of a large model, including:

The first processing module is used for inputting a first text prompt and a first task sample into a large model to obtain a processing result of the large model for the first task sample;

the evaluation module is used for evaluating the processing result by adopting the large model to obtain an evaluation result;

And the optimizing module is used for optimizing the first text prompt to obtain a second text prompt based on the reason which is recorded in the evaluation result and causes the problem when the evaluation result shows that the processing result has the problem.

According to another aspect of the present disclosure, there is provided an electronic device including:

At least one processor, and

A memory communicatively coupled to the at least one processor, wherein,

The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of the embodiments of the present disclosure.

According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform a method according to any one of the embodiments of the present disclosure.

According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements a method according to any of the embodiments of the present disclosure.

In the embodiment of the disclosure, the automatic evaluation is realized by utilizing the reasoning understanding capability of the large model, and deeper, more comprehensive and finer evaluation of the answer content of the large model can be realized. In the optimization process of each round, simple judgment standards such as traditional regular matching, numerical calculation and the like are not relied on, and the evaluation capability of a large model is fully utilized to score and evaluate the effect of the generated promt. The method breaks the limitation of the traditional evaluation method, so that the automatic optimization of the Prompt is not limited to a few task types which can be evaluated by simple rules, such as arithmetic questions, selection questions and the like, but can be popularized to a wider general task, and the performance and application value of the large model in various complex tasks are further improved.

It should be understood that the description in this section is not intended to identify key or critical features of the embodiments of the disclosure, nor is it intended to be used to limit the scope of the disclosure. Other features of the present disclosure will become apparent from the following specification.

Drawings

The drawings are for a better understanding of the present solution and are not to be construed as limiting the present disclosure. Wherein:

FIG. 1 is a flow diagram of a method of optimizing text hints for a large model in accordance with an embodiment of the present disclosure;

FIG. 2 is a flow diagram of a method of optimizing text hints for a large model in accordance with an embodiment of the present disclosure;

FIG. 3 is a flow diagram of a method of optimizing text hints for a large model in accordance with an embodiment of the present disclosure;

FIG. 4 is a block diagram of a method of optimizing text hints for a large model in accordance with an embodiment of the present disclosure;

FIG. 5 is an architecture diagram of an optimization apparatus for text hints for a large model in accordance with an embodiment of the present disclosure;

FIG. 6 is a block diagram of an electronic device for implementing a method of optimizing text hints for a large model in accordance with an embodiment of the present disclosure.

Detailed Description

Exemplary embodiments of the present disclosure are described below in conjunction with the accompanying drawings, which include various details of the embodiments of the present disclosure to facilitate understanding, and should be considered as merely exemplary. Accordingly, one of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope of the present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

The terms "first," "second," and the like in this disclosure are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. Furthermore, the terms "comprise" and "have," as well as any variations thereof, are intended to cover a non-exclusive inclusion, such as a series of steps or elements. The method, system, article, or apparatus is not necessarily limited to those explicitly listed but may include other steps or elements not explicitly listed or inherent to such process, method, article, or apparatus.

In the related art, the promt automatic optimization scheme has obvious limitations in application. Common promt optimization schemes generally rely on evaluation and screening of the performance of the optimized promt on a particular sample set. And the evaluation part mainly adopts simple methods such as regular matching or numerical calculation with reference answers. Therefore, the current promtt optimization scheme mainly focuses on task types, such as selection questions and arithmetic questions, that can simply determine whether to be correct or not in an application scenario. This simple evaluation method results in limited application of the current promt optimization scheme in more complex tasks, and difficulty in fully exploiting the capabilities of large models.

To solve this problem, an embodiment of the present disclosure proposes a method for optimizing text prompts of a large model, where the method combines automatic optimization of promt with automatic evaluation capability of the large model itself, as shown in fig. 1, and may be specifically implemented as follows:

S101, inputting the first text prompt and the first task sample into the large model to obtain a processing result of the large model aiming at the first task sample.

The first text prompt is an initial prompt message provided by the target object, and is updated later to automatically optimize the prompt message.

The first task sample is a problem set proposed by the target object.

Wherein the large model may be a large language model (Large Language Models, LLMs). The large language model refers to a large model of a specific type and is specially used for processing text data. Such models are neural network-based natural language processing models that can be used to generate, understand, and process text data. Large language models can have hundreds of billions of parameters, can generate high quality text, and can be used for various natural language processing tasks such as question-answering, text generation, dialog systems, and the like.

Large language models have good reasoning ability and small sample learning ability, where large language models can understand that the model is large in scale or can be trained on a large number of samples. Accurate semantic understanding can be achieved based on the large model.

S102, evaluating the processing result by adopting a large model to obtain an evaluation result.

And S103, when the evaluation result shows that the processing result has a problem, optimizing the first text prompt to obtain a second text prompt based on the reason which is recorded in the evaluation result and causes the problem.

And under the condition that the evaluation result shows that the processing result has no problem, the first text prompt is not optimized any more and is applied as the final prompt information.

In some embodiments, the large model is adopted to evaluate the processing result to obtain an evaluation result, and the evaluation result may be implemented by inputting the evaluation standard, the first task sample and the processing result into the large model to obtain an evaluation result output by the large model based on the evaluation standard under the condition that the first task sample does not have the standard answer.

The standard answer refers to an answer, such as an answer of a choice question or an arithmetic question. The labeling results in the examples of the present disclosure are optional, and thus, the first task sample may be an open task. By way of example, the task of the first task sample may be a authoring task, such as "I need an article describing autumn". Or the first task sample is a task that can get an answer based on a certain logic calculation, such as how much 1+1 is equal. Whether the first task sample has a standard answer or not according to the embodiment of the present disclosure can be determined based on actual situations, and the embodiment of the present disclosure is not limited herein.

In the embodiment of the disclosure, under the condition that the first task sample does not have a standard answer, the evaluation standard, the first task sample and the processing result are input into the large model, so that the large model can obtain the evaluation result of the first task sample by utilizing strong reasoning capacity of the large model, and the large model has more accurate prompt information for the task.

In some embodiments, the large model is used to evaluate the processing result to obtain an evaluation result, and the evaluation result may be implemented by inputting the evaluation standard, the first task sample, the standard answer and the processing result into the large model to obtain an evaluation result output by the large model based on the evaluation standard when the first task sample has the standard answer.

The evaluation criteria are described in the foregoing, and the embodiments of the disclosure are not described herein.

The first task sample, such as a choice question, calculates the question.

In the embodiment of the disclosure, under the condition that the first task sample has the standard answer, the evaluation standard, the first task sample, the processing result and the standard answer are input into the large model, so that the large model refers to the standard answer for evaluation, and the large model has a vector in the evaluation process, so that the strong reasoning capacity of the large model can be utilized, and the accuracy of the evaluation result of the large model is improved.

In some embodiments, to better optimize text prompts, whether the first task sample has a standard answer or not, as shown in fig. 2, the evaluation criteria may be determined based on the following method:

S201, a second task sample set, a model output result set of the second task sample set and remark information aiming at the model output result set are obtained, wherein the model output result set is obtained by processing second task samples in the second task sample set by a large model based on a third text prompt, the remark information is used for indicating whether model output results of the second task samples meet expectations or not, and the remark information further comprises reasons which cause the disagreement in the third text prompt under the condition that the model output results of the second task samples meet expectations.

The second task sample set and the first task sample set may be the same sample set or different sample sets, which is not limited in the embodiment of the present disclosure.

For any second task sample, under the condition that the model output result of the second task sample does not accord with the expectation, remarking can be carried out on the model output result, remarking can be carried out manually, and remarking information can be obtained through automatic remarking. The remark information also comprises a reason for causing the dissatisfaction in the third text prompt.

Further, the third text prompt is one or more. The third text prompt may be the same as the second text prompt or may be different, and embodiments of the present disclosure are not limited in this regard. In the case of a plurality of third text prompts, the method can be performed for each third text prompt respectively, wherein a large model is adopted to process each second task sample in the second task samples based on the third text prompt to obtain processing results of the second task samples, remark information is generated for the processing results of the second task samples and is used for indicating whether the processing results of the second task samples are correct or not and indicating reasons for errors caused by the third text prompt when the processing results of the second task samples are incorrect, and the remark information of the second task samples is summarized to obtain remark information of a model output result set of the second task sample set under the third text prompt.

Therefore, different third task samples are adopted, different problem samples can be summarized, and the large model can learn different problems of the text prompt, so that in S202, the second task sample set, the model output result set and remark information are input into the large model, and the large model outputs evaluation standards.

When the method is implemented, the second task sample set, the model output result set and remark information are input into the large model in a certain format, and the large model determines an evaluation standard according to the second task sample set and the remark information, so that the evaluation standard generated in the step and problems, processing results and the like of a target object are input into the large model together when automatic evaluation is applied for evaluation later, and a more accurate automatic evaluation result is obtained. The problem of the target object is the second task sample set.

In the embodiment of the disclosure, the second task sample set, the model output result set and remark information are input into a large model, and the analysis summarizing capability of the large model on the long text is utilized to obtain the evaluation standard. Therefore, the accuracy of the subsequent automatic evaluation of the processing result of the first task sample can be improved, and meanwhile, the automatic evaluation is realized based on the process, so that the manpower and material resources are saved.

In some embodiments, to automatically optimize the text prompts, iterative optimization of the text prompts may also be accomplished based on the following manner until the expectations are reached, as shown in fig. 3:

s301, when the respective second text prompts are obtained for the plurality of first text prompts, the plurality of candidate text prompts are screened out from the plurality of second text prompts based on the heuristic search method.

The heuristic Search method may be Beam Search (Beam Search), which performs a layer-by-layer Search by reserving a plurality of high-score candidates in a Search space, and gradually approximates an optimal solution.

In practice, an initial text prompt may be provided to be used as a starting point for Beam Search for subsequent rounds of iterative optimization.

Under the condition of providing an initial text prompt, in the first iteration process, the initial text prompt can be optimized aiming at a first task sample with a problem, and a candidate text prompt is obtained by each optimization, so that a certain number of candidate text prompts are ensured to be used for optimization in the subsequent iteration after the first iteration.

The first text prompt in the embodiment of the disclosure may be an initial text prompt, or may be any first text prompt obtained by optimization in each round after the first round of iteration.

In some embodiments, based on the heuristic search method, screening out a plurality of candidate text prompts from the plurality of second text prompts may be implemented as:

And A1, evaluating the accuracy of each second text prompt in the first task sample set to obtain the score of each second text prompt.

When the method is implemented, aiming at each second text prompt, the large model can process the first task sample set by using the second text prompt to obtain the processing result of each first task sample in the first task sample set. And determining the number of the first task samples with correct processing results according to a self-evaluation mode of the large model on the processing results of the first task samples, obtaining correct sample size, and counting the ratio of the correct sample size to the total sample size of the first task sample set to obtain the score of the second text prompt. Thus, a score for each second text prompt may be obtained.

For example, in the case that 100 questions exist in the first task sample set, and the large model evaluates for the promt 1, if 50 first task samples answer correctly, the accuracy of the promt 1 is 0.5, and the score of the corresponding promt 1 is 0.5. Similarly, if the 100 questions are answered in the same manner for the promt 2, if the answer is correct for 70 first task samples, the accuracy of the promt 2 is 0.7, and the score of the corresponding promt 2 is 0.7.

And A2, screening out a plurality of candidate text prompts based on the scores of the second text prompts.

When the method is implemented, the scores of the second text prompts are ranked, and k candidate text prompts with the highest scores are selected to enter the next round of searching process, wherein k is more than or equal to 1.

Wherein k is a positive integer and can be flexibly set according to requirements. For example, k may be set smaller for simple tasks and relatively larger for complex tasks.

In the embodiment of the disclosure, from the perspective of the accuracy of solving the problem by the prompt information, the score of each second text prompt is determined, and then a plurality of candidate text prompts ranked at the front are screened out, so that the text prompts are subjected to iterative tuning in the follow-up process, and the optimization efficiency is improved.

S302, each candidate text prompt is respectively used as a new first text prompt, and the step of inputting the first text prompt and the first task sample into the large model to obtain a processing result of the large model for the first task sample is returned until a termination condition is met.

In each subsequent iteration, k high-quality text prompts continue to be optimized, k text prompts represent different optimization directions, variants are generated on the corresponding text prompts by the first task sample with processing errors each time, and each variant generation is also an optimization direction, so that optimization can be performed in enough search space through continuous iterative optimization, and the optimal solution is approximated gradually.

In the embodiment of the disclosure, based on a heuristic search method, a plurality of candidate text prompts are screened out from a plurality of second text prompts, variants are continuously generated for optimizing based on the candidate text prompts, and the evaluation capability of a large model can be combined with iterative optimization to realize automatic optimization of the text prompts.

In some embodiments, the termination condition may include at least one of:

a) Reaching the target iteration number

When the method is implemented, under the condition that the exemplary target iteration times are M times, under the condition that the iteration update is performed for M rounds, the text prompt with the highest score is screened out from the last round of iteration to be used as the final text prompt.

B) Evaluation indexes for screening out multiple candidate text prompts in heuristic search method have lifting amplitude meeting preset cut-off condition

In implementation, the preset cutoff condition can be understood that the score of each candidate text prompt is not significantly improved as the iteration number increases.

C) Reaching the expected task performance index

When the method is implemented, the task performance index can be understood that the accuracy of the first text prompt is larger than a preset threshold.

The iterative optimization task may be ended when the above one termination condition is satisfied, or the loop task may be ended when any two or three conditions are satisfied. The specific manner in which to use as a termination condition may be determined based on actual conditions, which are not limited by the embodiments of the present disclosure.

In the embodiment of the disclosure, the iterative optimization process is ended by using the termination condition, so that the optimization process can be ensured to be terminated orderly. In the case of using a plurality of indexes as the termination condition, the quality of the finally obtained text information can be further improved.

After the termination condition is met, the final Prompt result will be returned to the target object. The target object can directly apply the promt to the actual task or can be further adjusted according to specific requirements. The whole process from initial promt input to final result output is completed under an automation framework, and the target object can obtain optimized high-quality promt only by providing the initial promt.

In practical applications, the target object can flexibly adjust parameters of the Beam Search, such as Beam Size (i.e. k) and iteration number, so as to meet requirements and resource limitations of different tasks. The quality of the promt can be effectively improved through multi-round iterative optimization, so that the performance of the large model in various complex tasks is improved.

In some embodiments, since the processing result of the first task sample may have various problems, the first text prompt may be purposefully optimized to obtain the second text prompt based on the cause of the problem recorded in the evaluation result.

By way of example, optimizing the first text prompts based on different reasons, respectively, may include the following:

1) In the event that the cause of the problem includes the large model ignoring the first target requirement in the first text prompt, the first text prompt is optimized by highlighting the first target requirement.

In the event that the large model is found to ignore the first target requirement in the first text Prompt, the key text in the promt may be bolded using Markdown bolded grammar to represent emphasized key text, such that the large model focuses on the key text.

In addition, instead of highlighting by using a bold method, an underline, a color font, italics, a special character, or the like may be used. In real time, text formats that are capable of highlighting the first target requirement are suitable for use with embodiments of the present disclosure. One way of highlighting is exemplified by that in Prompt the "style imitation: the language style of the captured original text" can be tuned to be "style imitation: the language style of the captured original text", i.e. it can be understood that the "style imitation" is highlighted.

In the embodiment of the disclosure, the large model can pay attention to and know the first target requirement in the first text prompt by highlighting the first target requirement, so that the task can be guaranteed to be completed according to the first target requirement under the condition that the later large model is convenient to execute the task, and the quality of the first text prompt is improved.

2) In the event that the cause of the problem includes that the second target requirement in the first text prompt is not clearly indicated, the first text prompt is optimized by specifying the indicated object in the second target requirement.

In practice, more explicit references may be used to reduce ambiguity, which may reduce uncertainty in large model generation when multiple components are included in the promt. For example, in the "language style of capturing an original text" in the promt, because the promt has multiple pieces of information, the term "original text" may refer to unclear, and the term "mouth cast text sample" may be adjusted to be "mouth cast text sample" based on actual conditions, and then a specific piece of text in the promt is explicitly referred to. The first text prompt after being tuned is the language style of "capture [ kou cast text sample ].

In the embodiment of the disclosure, the ambiguity can be reduced by using explicit reference, so that the large model identification is more explicit, and the first text prompt is optimized.

3) In case the cause of the question includes that the preset requirements for the model answer in the first text prompt are not accurately described, the first text prompt is optimized by adding examples to the preset requirements.

In practice, the promt is made clearer by way of example. When the requirements for model answers exist in Prompt that are difficult to describe accurately, a portion of examples may be added to make a large model easier to understand. For example, the "smart blend in feature in imitation writing" in Prompt may be tuned to "smart blend in feature in imitation writing", such as feature spoken Buddhist, ancient poetry explanation, knowledge science popularization, etc.

In the embodiment of the disclosure, the mode of adding examples is used, so that the prompt information is clearer, and further, the second prompt information obtained after optimization can reduce the understanding difficulty of the model answer requirement of the large model as much as possible, thereby prompting the accuracy of the processing task of the large model.

4) In case the reasons include that the big model cannot follow the explicit requirements in the first text prompt, the first text prompt is optimized by applying an enhanced mood to the explicit requirements.

In case explicit requirements have been given explicitly in Prompt and the expectation of the processing results of the large model is not yet met, the explicit requirements may be enhanced in mind in order to enable the large model to pay more attention to the explicit requirements, making the explicit requirements important in order to handle the task correctly.

In language expression, enhancing the mood refers to enhancing the expressive force of sentences by using specific vocabulary, grammar structures or punctuation marks, so that the sentences are more prominent, emphasized or emotion-enriched. The following are some examples of enhanced mood:

(1) The exclamatory words such as "o", "ou" and the like can be used for expressing the emotion such as surprise, happiness and the like in the oral language.

(2) The use of accentuated adverbs, such as "very", "actual", "exact", etc., is used to strengthen the adjectives or verbs.

(3) By using the comparision structure, the rhythm sense and the emphasis effect of the language can be enhanced by repeating the similar sentence pattern structure.

(4) Using exclamation sentences, express strong emotion or emphasis by adding an exclamation mark (|) at the end of the sentence.

(5) The questioning sentence is used, and the attention and thinking of a listener can be brought by questioning.

(6) The use of metaphors and exaggeration, meaning can be expressed visually by metaphors, and exaggeration can highlight a certain feature.

(7) Using a flip-chip sentence-in english, a flip-chip sentence can emphasize a certain part of the sentence.

(8) Repeating a word or sentence can enhance mood and impression.

(9) Using contrast, one of which can be emphasized by comparing different things or perspectives.

(10) Using command sentences when it is necessary to express a strong demand or command, the command sentences can be used.

It should be noted that, the emphasis of the mood is not limited to the above example, and the emphasis of the mood may be implemented based on the actual situation to optimize the first text prompt.

Exemplary please write imitations for given [ merchandise information ] and [ kou cast document sample ] according to the above cues and requirements, and directly provide the imitated written document content. "please be optimized" based on the above cues and requirements, imitation writing is performed for given [ merchandise information ] and [ kou-cast text sample ]. Please strictly follow all the task requirements described above, otherwise you would be prohibited from any authoring | and directly provide the copy-written document content. ".

The tuning of the mode strictly obeys all task requirements, or else, you are forbidden to create | to strengthen the required language, so that the large model can better define the task requirements, has clear task directions and can better complete the tasks. Therefore, based on the second text prompt obtained by optimizing in the mode, the requirements of the large model understanding can be facilitated, and the task can be completed better.

5) In case the cause of the problem includes that the third target requirement in the first text prompt is not sufficiently detailed, the first text prompt is optimized by describing the third target requirement in detail.

In practice, when there is a problem in that description is not detailed enough in the promt, a large model can be more easily understood by a more detailed description. For example, "if there is a small part of content that deviates from the 'question' but does not affect the main meaning of the 'answer to be verified' is correct", there may be an unspecific question in the Prompt, the deviation may cause the large model to not know where the deviation occurs, and the "main meaning is correct" may cause the large model to not know what the main meaning is. The description here can thus be adapted to be "partially accurate if there is a small portion of content that does not exactly coincide with the 'question', but the overall core meaning coincides with the 'question', and such deviation is not sufficient to change the main meaning of the answer. "

In the embodiment of the disclosure, the first problem prompt is optimized by using the detailed description mode, so that the prompt information is more detailed, further, the large model can capture more detailed content in the prompt information, and the accuracy of the large model is improved.

6) In cases where the fourth target requirement in the first text prompt results in misinterpretation of the large model, the first text prompt is optimized in a manner that eliminates misinterpretation by overwriting the fourth target requirement.

In practice, in the case that the description in the promt makes the large model misunderstand the requirement, the promt can be enabled to express the requirement more accurately by modifying the misunderstood description. Illustratively, when the Prompt requires that the results must be returned in json format, and based on the evaluation conclusion, the results may be returned in other formats in some special cases, in which case the relevant description about the return format requirement may be modified to "please return in json format in xxx case, and in special case yyy, the zzz format may also be used to make the description more accurate".

In the embodiment of the disclosure, the misleading description is modified into the clear description, so that the second text prompt obtained by tuning can help the large model to know the task more clearly, and further an accurate processing result is obtained based on the strong reasoning capability of the large model.

For ease of understanding, the optimization process in the embodiment of the disclosure is as shown in fig. 4, where a first text prompt (such as an initial text prompt P0), a second task sample (Q), a model output result set (A0), and remark information (I), and in addition, for any second task sample, if a standard answer exists, the standard answer (R) of the second task sample may be input into a large model together, so that a set of evaluation criteria may be automatically summarized by using the summarized analysis capability of the large model. Then, the large model can automatically evaluate the processing result of the first task sample set by using the set of evaluation criteria to obtain an evaluation result. Based on the cause of the problem in the evaluation result, generating variants for the first text prompt, so as to iteratively optimize the first text prompt and obtain a plurality of second text prompts evolved from the first text prompt. Exemplary as shown in fig. 4 include a second text prompt P1, a second text prompt P2, a second text prompt P3, a second text prompt PN. And inputting the large models into the second text prompts to score the second text prompts, and obtaining scores A1, A2, A3, A.and AN of the second text prompts. And screening k second text prompts with highest scores as candidate text prompts based on the scores of the second text prompts. Wherein N is greater than k. Thus, a round of optimization iterations ends. Judging whether the termination condition is met, if not, taking each candidate text prompt as a new first text prompt, carrying out the next round of optimization iteration until the termination condition is met, and screening out one candidate prompt message (P') with the highest score.

As shown in fig. 4, an example of an initial text prompt P0 that needs to be optimized is given in the table, and then the first task sample is processed based on the P0, so as to obtain a processing result generated by the large model. And then the large model automatically evaluates the processing result to obtain an evaluation result. As shown in fig. 4, the evaluation result indicates that the processing result does not conform to the expected reasons for the model answer failing to follow the requirements in the prompt. Thus, P0 may be automatically optimized for the cause to obtain a corresponding second text prompt. And iterating by the method to obtain a plurality of optimized text prompts. As shown in fig. 4, three second text prompts P1, P2 and P3 are obtained, and the three second text prompts are respectively evaluated by a large model, so that the accuracy of each second text prompt for a task set is evaluated, the score of each second text prompt is obtained, and then the high-quality second text prompt is screened from the scores, and the iterative optimization is continued. As shown in fig. 4, among the three second text cues P1, P2 and P3, the scores of P1 and P2 are higher, and if k=2, P1 and P2 can be selected as new first text cues for iterative optimization.

Based on the same technical concept, the embodiment of the disclosure proposes an optimizing device 500 for text prompting of a large model, as shown in fig. 5, including:

The first processing module 501 is configured to input a first text prompt and a first task sample to the large model, so as to obtain a processing result of the large model for the first task sample;

The evaluation module 502 is configured to evaluate the processing result by using the large model to obtain an evaluation result;

And an optimizing module 503, configured to optimize the first text prompt to obtain a second text prompt based on a cause of the problem recorded in the evaluation result when the evaluation result indicates that the processing result has the problem.

In some embodiments, the evaluation module is to:

under the condition that the first task sample does not have a standard answer, the evaluation standard, the first task sample and the processing result are input into a large model, and an evaluation result output by the large model based on the evaluation standard is obtained.

In some embodiments, the evaluation module is configured to:

under the condition that the first task sample has the standard answer, the evaluation standard, the first task sample, the standard answer and the processing result are input into the large model, and the evaluation result output by the large model based on the evaluation standard is obtained.

In some embodiments, the method further comprises a second processing module for:

Under the condition that respective second text prompts are respectively obtained for the plurality of first text prompts, screening a plurality of candidate text prompts from the plurality of second text prompts based on a heuristic search method;

And respectively taking each candidate text prompt as a new first text prompt, and returning to the step of inputting the first text prompt and the first task sample into the large model to obtain a processing result of the large model aiming at the first task sample until the termination condition is met.

In some embodiments, the second processing module is configured to:

Evaluating the accuracy of each second text prompt in the first task sample set to obtain the score of each second text prompt;

a plurality of candidate text prompts are screened based on the scores of the second text prompts.

In some embodiments, the termination condition includes at least one of:

Reaching the target iteration times;

the improvement amplitude of the evaluation index for screening out a plurality of candidate text prompts in the heuristic search method meets the preset cut-off condition;

The expected task performance index is reached.

In some embodiments, the system further comprises an evaluation criterion determination module for:

The method comprises the steps of obtaining a second task sample set, a model output result set of the second task sample set and remark information aiming at the model output result set, wherein the model output result set is obtained by processing second task samples in the second task sample set by a large model based on a third text prompt;

And inputting the second task sample set, the model output result set and remark information into the large model so that the large model outputs the evaluation standard.

In some embodiments, the optimization module is specifically configured to:

in the event that the cause includes the large model ignoring the first target requirement in the first text prompt, the first text prompt is optimized by highlighting the first target requirement.

In some embodiments, the optimization module is specifically configured to:

In the event that the cause includes that the second target requirement in the first text prompt is under-referenceed, the first text prompt is optimized by specifying the referents in the second target requirement.

In some embodiments, the optimization module is specifically configured to:

In the event that the cause includes a failure to accurately describe the preset requirements for the model answer in the first text prompt, the first text prompt is optimized by adding examples to the preset requirements.

In some embodiments, the optimization module is specifically configured to:

In the case where the reasons include that the large model cannot follow the explicit requirements in the first text prompt, the first text prompt is optimized by applying an enhanced mood to the explicit requirements.

In some embodiments, the optimization module is specifically configured to:

In case the reason comprises that the third target requirement in the first text prompt is not sufficiently detailed, the first text prompt is optimized by detailing the third target requirement.

In some embodiments, the optimization module is specifically configured to:

in cases where the reason includes that the fourth target requirement in the first text prompt results in a misinterpretation of the large model, the first text prompt is optimized in a manner that eliminates misinterpretation by overwriting the fourth target requirement.

For descriptions of specific functions and examples of each module and sub-module of the apparatus in the embodiments of the present disclosure, reference may be made to the related descriptions of corresponding steps in the foregoing method embodiments, which are not repeated herein.

In the technical scheme of the disclosure, the acquisition, storage, application and the like of the related user personal information all conform to the regulations of related laws and regulations, and the public sequence is not violated.

According to embodiments of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

Fig. 6 illustrates a schematic block diagram of an example electronic device 600 that may be used to implement embodiments of the present disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile apparatuses, such as personal digital assistants, cellular telephones, smartphones, wearable devices, and other similar computing apparatuses. The components shown herein, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the disclosure described and/or claimed herein.

As shown in fig. 6, the apparatus 600 includes a computing unit 601 that can perform various appropriate actions and processes according to a computer program stored in a Read Only Memory (ROM) 602 or a computer program loaded from a storage unit 608 into a Random Access Memory (RAM) 603. In the RAM603, various programs and data required for the operation of the device 600 may also be stored. The computing unit 601, ROM 602, and RAM603 are connected to each other by a bus 604. An input/output (I/O) interface 605 is also connected to bus 604.

Various components in the device 600 are connected to the I/O interface 605, including an input unit 606, e.g., keyboard, mouse, etc., an output unit 607, e.g., various types of displays, speakers, etc., a storage unit 608, e.g., magnetic disk, optical disk, etc., and a communication unit 609, e.g., network card, modem, wireless communication transceiver, etc. The communication unit 609 allows the device 600 to exchange information/data with other devices via a computer network, such as the internet, and/or various telecommunication networks.

The computing unit 601 may be a variety of general and/or special purpose processing components having processing and computing capabilities. Some examples of computing unit 601 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various specialized Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as the optimization method of text hints for large models. For example, in some embodiments, the method of optimizing text hints for large models may be implemented as a computer software program tangibly embodied on a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program may be loaded and/or installed onto the device 600 via the ROM 602 and/or the communication unit 609. When the computer program is loaded into the RAM 603 and executed by the computing unit 601, one or more steps of the above-described optimization method of text hints for large models can be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to perform the optimization method of text hints for large models in any other suitable manner (e.g., by means of firmware).

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuit systems, field Programmable Gate Arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), systems On Chip (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include being implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be a special or general purpose programmable processor, operable to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

Program code for carrying out methods of the present disclosure may be written in any combination of one or more programming languages. These program code may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus such that the program code, when executed by the processor or controller, causes the functions/operations specified in the flowchart and/or block diagram to be implemented. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package, partly on the machine and partly on a remote machine or entirely on the remote machine or server.

In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. The machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other types of devices may also be used to provide interaction with the user, for example, feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback), and input from the user may be received in any form, including acoustic input, speech input, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a background component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such background, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a Local Area Network (LAN), a Wide Area Network (WAN), and the Internet.

The computer system may include a client and a server. The client and server are typically remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server may be a cloud server, a server of a distributed system, or a server incorporating a blockchain.

It should be appreciated that various forms of the flows shown above may be used to reorder, add, or delete steps. For example, the steps recited in the present disclosure may be performed in parallel, sequentially, or in a different order, provided that the desired results of the disclosed aspects are achieved, and are not limited herein.

The above detailed description should not be taken as limiting the scope of the present disclosure. It will be apparent to those skilled in the art that various modifications, combinations, sub-combinations and alternatives are possible, depending on design requirements and other factors. Any modifications, equivalent substitutions, improvements, etc. that are within the principles of the present disclosure are intended to be included within the scope of the present disclosure.

Claims

1. A method for optimizing text prompts for a large model, comprising:

If the evaluation result shows that the processing result has a problem, optimizing the first text prompt to obtain a second text prompt based on the reason which is recorded in the evaluation result and causes the problem;

the evaluation result is generated by the large model based on evaluation criteria, and the method further comprises determining the evaluation criteria based on the following method:

The method comprises the steps of obtaining a second task sample set, a model output result set of the second task sample set and remark information aiming at the model output result set, wherein the model output result set is obtained by processing second task samples in the second task sample set by the large model based on a third text prompt;

and inputting the second task sample set, the model output result set and the remark information into the large model so as to output the evaluation standard by utilizing the analysis summarization capability of the large model on the text.

2. The method of claim 1, wherein the evaluating the processing results using the large model to obtain evaluation results comprises:

And under the condition that the first task sample does not have a standard answer, inputting an evaluation standard, the first task sample and the processing result into the large model to obtain the evaluation result output by the large model based on the evaluation standard.

3. The method of claim 1, wherein the evaluating the processing results using the large model to obtain evaluation results comprises:

And under the condition that the first task sample has a standard answer, inputting an evaluation standard, the first task sample, the standard answer and the processing result into the large model to obtain the evaluation result output by the large model based on the evaluation standard.

4. A method according to any one of claims 1-3, further comprising:

And respectively taking each candidate text prompt as a new first text prompt, and returning to execute the step of inputting the first text prompt and the first task sample into a large model to obtain a processing result of the large model for the first task sample until a termination condition is met.

5. The method of claim 4, wherein the screening the plurality of candidate text prompts from the plurality of second text prompts based on the heuristic search method comprises:

And screening the candidate text prompts based on the scores of the second text prompts.

6. The method of claim 4, the termination condition comprising at least one of:

Reaching the target iteration times;

The improvement amplitude of the evaluation index for screening out the candidate text prompts in the heuristic search method meets the preset cut-off condition;

The expected task performance index is reached.

7. The method of any of claims 1-6, wherein the optimizing the first text prompt based on the cause of the problem recorded in the evaluation result comprises:

In the event that the cause includes the large model ignoring a first target requirement in the first text prompt, the first text prompt is optimized by highlighting the first target requirement.

8. The method of any of claims 1-7, wherein the optimizing the first text prompt based on the cause of the problem recorded in the evaluation result comprises:

in the event that the cause includes a second target requirement in the first text prompt being imperceptible, the first text prompt is optimized by specifying a pointing object in the second target requirement.

9. The method of any of claims 1-8, wherein the optimizing the first text prompt based on the cause of the problem recorded in the evaluation result comprises:

in case the reason comprises that the preset requirements for the model answer in the first text prompt are not accurately described, optimizing the first text prompt by adding an example to the preset requirements.

10. The method of any of claims 1-9, wherein the optimizing the first text prompt based on the cause of the problem recorded in the evaluation result comprises:

In case the cause comprises that the big model cannot follow an explicit requirement in the first text prompt, optimizing the first text prompt by applying an enhanced mood to the explicit requirement.

11. The method of any of claims 1-10, wherein the optimizing the first text prompt based on the cause of the problem recorded in the evaluation result comprises:

12. The method of any of claims 1-11, wherein the optimizing the first text prompt based on the cause of the problem recorded in the evaluation result comprises:

in the case where the cause includes a fourth target requirement in the first text prompt causing misinterpretation of the large model, the first text prompt is optimized by overwriting the fourth target requirement in a manner that eliminates misinterpretation.

13. An optimization apparatus for text prompts for a large model, comprising:

the optimizing module is used for optimizing the first text prompt to obtain a second text prompt based on the reason which is recorded in the evaluation result and causes the problem when the evaluation result shows that the processing result has the problem;

the evaluation result is generated by the large model based on an evaluation standard, and the evaluation result evaluation system further comprises an evaluation standard determination module for:

14. The apparatus of claim 13, wherein the evaluation module is specifically configured to:

15. The apparatus of claim 13, wherein the evaluation module is specifically configured to:

16. The apparatus of any of claims 13-15, further comprising a second processing module to:

and respectively taking each candidate text prompt as a new first text prompt, triggering the first processing module to return to execute the step of inputting the first text prompt and the first task sample into a large model to obtain a processing result of the large model for the first task sample until a termination condition is met.

17. The apparatus of claim 16, wherein the second processing module is configured to:

18. The apparatus of claim 16, the termination condition comprising at least one of:

Reaching the target iteration times;

The expected task performance index is reached.

19. The apparatus according to any one of claims 13-18, wherein the optimization module is specifically configured to:

20. The apparatus according to any one of claims 13-19, wherein the optimization module is specifically configured to:

21. The apparatus according to any one of claims 13-20, wherein the optimization module is specifically configured to:

22. The apparatus according to any one of claims 13-21, wherein the optimization module is specifically configured to:

23. The apparatus according to any one of claims 13-22, wherein the optimization module is specifically configured to:

24. The apparatus according to any one of claims 13-23, wherein the optimization module is specifically configured to:

25. An electronic device, comprising:

At least one processor, and

A memory communicatively coupled to the at least one processor, wherein,

The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-12.

26. A non-transitory computer readable storage medium storing computer instructions for causing the computer to perform the method of any one of claims 1-12.

27. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any of claims 1-12.