CN116932037A - Software system configuration generation method and system based on generation type large model - Google Patents
Software system configuration generation method and system based on generation type large model Download PDFInfo
- Publication number
- CN116932037A CN116932037A CN202311002875.0A CN202311002875A CN116932037A CN 116932037 A CN116932037 A CN 116932037A CN 202311002875 A CN202311002875 A CN 202311002875A CN 116932037 A CN116932037 A CN 116932037A
- Authority
- CN
- China
- Prior art keywords
- configuration
- data
- target
- generation
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 238000012216 screening Methods 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims description 35
- 230000006978 adaptation Effects 0.000 claims description 19
- 238000012952 Resampling Methods 0.000 claims description 10
- 238000013461 design Methods 0.000 claims description 5
- 230000000694 effects Effects 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 4
- QVFWZNCVPCJQOP-UHFFFAOYSA-N chloralodol Chemical compound CC(O)(C)CC(C)OC(O)C(Cl)(Cl)Cl QVFWZNCVPCJQOP-UHFFFAOYSA-N 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 238000012360 testing method Methods 0.000 claims description 3
- 238000002347 injection Methods 0.000 claims description 2
- 239000007924 injection Substances 0.000 claims description 2
- 230000002452 interceptive effect Effects 0.000 claims description 2
- 238000001914 filtration Methods 0.000 claims 2
- 238000013507 mapping Methods 0.000 claims 1
- 230000006870 function Effects 0.000 description 16
- 238000012986 modification Methods 0.000 description 9
- 230000004048 modification Effects 0.000 description 9
- 230000003993 interaction Effects 0.000 description 4
- 238000012423 maintenance Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 239000008186 active pharmaceutical agent Substances 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 239000000243 solution Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F8/00—Arrangements for software engineering
- G06F8/70—Software maintenance or management
- G06F8/71—Version control; Configuration management
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Security & Cryptography (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Stored Programmes (AREA)
Abstract
The invention provides a software system configuration generation method and system based on a generation type large model, and belongs to the technical field of software. The method is based on a generation type large model, multi-source system configuration data are collected, a fine adjustment data set is obtained by adapting according to the similarity between the configuration data and target software system configuration data, fine adjustment is carried out on the generation type large model by using the fine adjustment data set, multi-target prompt engineering is applied to receive original configuration text, configuration target text data and configuration example sets from a user, the multi-target prompt text is generated as input of the generation type large model, candidate configuration is obtained by model output, finally, effective configuration meeting configuration targets is selected from the plurality of output candidate configuration by a user through configuration screening, the effective configuration meeting the configuration targets is used as a final result, or the effective configuration meeting the configuration targets is returned to be used for optimizing the multi-target prompt text by the user, and quality of the multi-target prompt text and effectiveness of configuration generation are further improved.
Description
Technical Field
The invention belongs to the technical field of software, and particularly relates to a software system configuration generation method and system based on a large generation model, which can automatically generate proper software system configuration according to different targets, thereby improving configuration efficiency and reducing configuration cost.
Background
With the increasing size of software systems, the degree of distributivity is increasing, and the software configuration becomes a complex and error-prone work. On the one hand, a software system generally comprises various components and program modules, and both the internal interaction of the same program module and the interaction of different components need detailed configuration, and the software configuration content is extremely complex due to a large number of configuration parameters and the interaction and the dependency relationship between the configuration parameters; on the other hand, the operation and maintenance personnel often need to perform software configuration according to different targets so as to achieve different system performances, for example, when the system stability guarantee is taken as a target, the operation and maintenance personnel often need to increase the node backup and the exclusive resource proportion in the configuration so as to reduce the resource competition, when the system reliability test is taken as a target, the operation and maintenance personnel needs to gradually reduce the node backup and the exclusive resource proportion in the configuration so as to increase the resource competition, and the diversity of the configuration targets further increase the difficulty of the software configuration. In summary, aiming at complex configuration content and various configuration targets, how to generate reasonable and effective software system configuration, thereby improving configuration efficiency and reducing configuration cost becomes a key problem.
In order to solve the problems, most of the prior art is to construct a configuration file library, and an operation and maintenance person automatically searches similar configuration according to a configuration target and manually checks and corrects the similar configuration, so that the efficiency is low and the configuration quality is poor. With the development of artificial intelligence technology, large models (Large Language Model, LLM) such as gpt3, gpt4, etc. are generated, and their excellent language representation capability and text generation capability are receiving a great deal of attention. The massive available configuration files can be quickly generated by using the generated large model, so that the configuration efficiency is greatly improved; meanwhile, by using model fine tuning and prompt engineering technology, a proper configuration file can be generated aiming at different targets, and the configuration quality is improved. Therefore, the development of the software system configuration generation method and system based on the generation type large model has very remarkable significance.
Disclosure of Invention
Based on the prior art, the invention provides a software system configuration generation method and system based on a generation type large model, which takes the generation type large model as a core, collects and uses multi-source system configuration data to carry out fine adjustment on the generation type large model, and utilizes multi-target prompt engineering to generate proper software configuration for different targets, thereby providing an automatic generation solution for the software system configuration.
The technical scheme provided by the invention is as follows:
a software system configuration generation method based on a generation type large model comprises the following three steps:
1) Configuration data adaptation and model fine tuning are specifically as follows: collecting configuration data from software systems such as an open source community software system, a target software system and the like, performing adaptation of the configuration data according to the similarity between the configuration data and the configuration data of the target software system to obtain a fine-tuning data set, and performing fine tuning of a generated large model by using the data set;
2) The multi-target prompt project specifically comprises: receiving original configuration text and configuration target text data from a user, and a configuration example set, generating a multi-target prompt text, taking the multi-target prompt text as input of a large generation model, and outputting the model to obtain candidate configuration;
3) Configuration screening, specifically: the user screens the valid configuration in the candidate configuration set of the generative large model output as the final configuration generation result or returns it for updating the configuration example.
Aiming at the software system configuration generation method based on the large generation model, further, the configuration data adaptation and model fine tuning specifically execute the following steps:
11 Collecting multisource system configuration data, the multisource system comprising a software system S in an open source community 1 ,S 2 ,...,S Ns And a target software system S to be configured and generated 0 Wherein N is s Representing the number of software systems in the open source community. Configuration data is obtained from resources such as web pages, version control systems and the like. Sorting configuration data into configuration data setsWherein N is R Representing the number of data bars>Representing a piece of configuration data,/->Representing a software system, c bi Representing the original configuration, c ai Representing a configuration meeting a configuration objective, purp i Representing a configuration target;
12 Adapting the configuration data, resampling the configuration data according to the similarity between the configuration data and the configuration data of the target system, and generating a fine tuning data set;
13 The model is used for fine tuning training, a fine tuning data set is used for fine tuning training on the generated large model, a prompting template is designed to modify the format of fine tuning data, so that the mode of the fine tuning data is more similar to the mode of training data, and then the modified data is used for fine tuning to generate the large model, so that the model is more suitable for configuration generation tasks.
Aiming at the software system configuration generation method based on the generation type large model, the multi-target prompt project specifically executes the following steps:
21 User input of original configuration text c) b Configuring target pures and generating number N of candidate configurations gen ；
22 User input N) ex Strip configuration generation exampleEach example is shown as example i ＝<c ebi ,c eai ,purp ei >Wherein c ebi Representing an original configuration example, purp ei Representing configuration target examples, c eai Representing configuration examples that meet the configuration objective;
23 Design of multi-objective alert template q And the method is used for generating the multi-target prompt text as the input of the large generation model. template q Original configuration text c using user input b Configuration target purp and configuration generation exampleFilling;
24 Inputting the multi-target prompt text into the trimmed generation type large model, wherein the result output by the model is a generated candidate configuration;
25 Repeating step 24) N gen Second, a candidate configuration set res= { c is obtained o1 ,c o2 ,...,c oNgen And (3) returning Res to the user.
Aiming at the software system configuration generation method based on the generation type large model, further, the configuration screening specifically executes the following steps:
31 For the candidate configuration set Res obtained in step 25), the user selects a candidate configuration set conforming to the configuration target purp therefromAs an output active configuration set.
32 Res obtained in step 31) correct The method and the device can be used for updating configuration generation examples input by a user in the multi-target prompt engineering, and further improve the quality of multi-target prompt texts and the effectiveness of configuration generation. The concrete practice is to put Res correct The configuration generation result in step 22) is added to the configuration generation example input by the user in step 22), and step 22) is repeated to this step. The repetition may be performed a number of times until the resulting configuration achieves the user's satisfaction.
The invention also provides a system for realizing the software system configuration generation method based on the generation type large model, which comprises a configuration data adaptation and model fine adjustment module, a multi-target prompt engineering module and a configuration screening module:
the configuration data adaptation and model fine adjustment module is used for collecting configuration data of multi-source systems such as an open-source community system and a target system; resampling the configuration data based on the similarity between the configuration data and the configuration data of the target system to obtain a fine tuning data set adapted to the target system; and finally, converting the fine adjustment data into training texts by using a training prompt template for fine adjustment of the generated large model.
The multi-target prompt engineering module is used for providing an interactive prompt engineering interface for a user, and generating examples, original configuration texts and configuration target texts according to configuration provided by the user by arranging a multi-target prompt template. The generated hints are input into a generative large model, driving it to output a candidate configuration set. In addition, the module receives the effective configuration fed back by the configuration screening module, and updates the configuration generation example by using the effective configuration, so that the effect of the multi-objective prompt engineering is further improved.
And the configuration screening module is used for providing an interface for screening effective configurations meeting configuration targets from the candidate configuration set for a user, wherein the effective configurations can be used as output generated by the configuration and can be used as feedback to be input into the multi-target prompt engineering module.
Further, the configuration data adaptation and model fine tuning module comprises three sub-modules:
1) The configuration data collector is used for realizing grabbing and collecting of configuration data, sorting the configuration data into a configuration data set and transmitting the configuration data set to the configuration data adapter;
2) The configuration data adapter is used for calculating the similarity between the configuration data and the configuration data of the target software system, and resampling the configuration data set according to the similarity to obtain a fine-tuning data set;
3) The model fine tuning trainer uses a training prompt template to convert fine tuning data into training texts and fine-tune the generated large model.
Further, the multi-objective prompt engineering module includes three sub-modules:
1) The configuration requirement receiver is used for receiving the original configuration and the configuration target from the user, transmitting the original configuration and the configuration target to the multi-target prompt text generator and generating the multi-target prompt text;
2) A configuration example receiver for receiving configuration examples or an effective configuration set from a user and a configuration feedback device in a configuration screening module for providing configuration examples for the multi-objective prompt text generator;
3) And the multi-target prompt text generator is used for receiving the original configuration, the configuration target and the configuration example from the configuration requirement receiver and the configuration example receiver and generating multi-target prompt text by using the multi-target prompt text generator to guide the generation type large model to generate the configuration meeting the configuration target.
Further, the configuration screening module includes two sub-modules:
1) The configuration selector is used for receiving the candidate configuration set output by the generated large model, providing the candidate configuration set for a user to select effective configurations meeting configuration targets, and transmitting the effective configuration set screened by the user to the configuration feedback device;
2) And the configuration feedback device is used for outputting the effective configuration set, and automatically judging whether to transmit the effective configuration set to a configuration example receiver in the multi-objective prompt engineering module for iteration according to user selection or algorithm.
Compared with the prior art, the invention has the beneficial effects that:
the invention provides a software system configuration generation method and system based on a large generation model, which are used for collecting abundant configuration data of a practical software system from a multi-source software system such as a software system in an open source community, a target software system and the like, performing configuration data adaptation by a method for calculating the similarity of the configuration data and the configuration data of the target software system, and generating a fine adjustment data set which is more adapted to the configuration data of the target software system. And then, the generated large model is subjected to fine adjustment by utilizing the fine adjustment data set, so that the generated large model can be suitable for the configuration under various configuration targets in the generated target software system. Then, based on the multi-objective prompt engineering, the large model is guided to output configuration text meeting the configuration objective by combining the prompt template and the configuration examples input by the user. Based on the output result, the user can perform screening of effective configuration, and the quality of the multi-target prompt text and the effectiveness of configuration generation are further improved by using the method. The whole process has stronger flexibility, on one hand, the collection of multi-source configuration data and the fine adjustment of the model can be automatically carried out, and massive configuration is generated by utilizing the large generation model according to a given configuration text and various configuration targets, so that the configuration efficiency is greatly improved; on the other hand, the multi-target prompt engineering in the invention allows a user to optimize the multi-target prompt text by inputting configuration examples and feeding back the result of configuration screening so as to guide the model to output the configuration meeting the configuration target and further improve the effect of the model. The method and the system realize the software system configuration generation based on the generation type large model, and can effectively improve the efficiency of the software system configuration generation.
Drawings
FIG. 1 is a flow chart of a software system configuration generation method based on a generative large model provided by the invention;
FIG. 2 is a block diagram of a software system configuration generation system based on a generative large model provided by the present invention.
Detailed Description
The invention is further described by way of examples in the following with reference to the accompanying drawings, but in no way limit the scope of the invention.
The invention provides a software system configuration generation method and system based on a generation type large model, which are used for collecting configuration data from multi-source software systems such as an open-source community software system, a target software system and the like, and obtaining a fine adjustment data set through configuration adaptation, wherein the fine adjustment data set is used for carrying out fine adjustment on the generation type large model, so that the fine adjustment data set is more suitable for configuration generation tasks of the target software system. And automatically generating configuration according to different configuration targets by utilizing the multi-target prompt engineering guidance generation type large model, and improving the quality of multi-target prompt texts by providing configuration examples and feeding back effective configuration by a user so as to further improve the effectiveness of configuration generation.
FIG. 1 is a block flow diagram of a software system configuration generation method based on a generative large model provided by the invention. In the stage of configuration data adaptation and model fine tuning, a configuration data collector collects configuration data from software systems such as an open source community software system, a target software system and the like, then a configuration data adapter adapts the configuration data according to similarity between the configuration data and the configuration data of the target software system to obtain a fine tuning data set, and finally a model fine tuning trainer uses the data set to conduct fine tuning of a generated large model. In a model prediction stage based on multi-objective prompt engineering, a configuration requirement receiver and a configuration example receiver respectively receive original configuration text and configuration objective text data from a user, and a configuration example set; the received user input is used by the multi-target alert text generator to generate multi-target alert text as input to the generative large model. In the configuration screening stage, a candidate configuration set output by the generated large model is input into a configuration selector, so that a user screens effective configuration; the effective configuration set is input into a configuration feedback device, the configuration feedback device outputs the effective configuration set as a final configuration generation result, or returns the effective configuration set to a configuration example receiver for updating a configuration example, optimizing a multi-target prompt text, improving the effectiveness of configuration generation, and the process can be repeated for a plurality of times until the generated configuration achieves the effect of user satisfaction. Three different stages are each described in detail below.
1) The configuration data adaptation and model fine tuning stage specifically performs the following steps:
11 Collecting multisource system configuration data, the multisource system including software systems in an open source communityAnd a target software system S to be configured and generated 0 ，N s Representing the number of software systems in the open source community; configuration data is obtained from web pages and version control system resources; sorting configuration data into configuration data setsWherein N is R Represents the number of data stripes rec i ＝<S sri ,c bi ,c ai ,purp i >Representing a piece of configuration data,/->Representing a software system, c bi Representing the original configuration, c ai Representing a configuration meeting a configuration objective, purp i Representing a configuration target;
12 And (3) adapting the configuration data, resampling the configuration data according to the similarity between the configuration data and the configuration data of the target system, and generating a fine-tuning data set. First, the function similarity (rec i ,R 0 ) Calculating the similarity between the configuration data and the configuration data of the target system, wherein rec i Is one piece of configuration data in R 0 Is a set of configuration data pertaining to the target system. The higher the similarity, the higher the function return value from 0 to 1. Each piece of data rec is then normalized using a normalization function i Similarity of configuration data to target system (rec) i ,R 0 ) Probability p mapped to word sample i Meets the similarity (rec) i ,R 0 ) The higher p i The higher andn-th for a data set based on the above sampling probability distribution D Subsampling to obtain a scale of N D Fine tuning data set->Wherein N is D Represents the number of data strips obtained after resampling, and d is more than or equal to 1 i ≤N R ；
13 The model is used for fine tuning training, a fine tuning data set is used for fine tuning training on the generated large model, a prompting template is designed to modify the format of fine tuning data, so that the mode of the fine tuning data is more similar to the mode of training data, and then the modified data is used for fine tuning to generate the large model, so that the model is more suitable for configuration generation tasks. The following operations are specifically performed:
13a) Design training prompting template t For adjusting the model trim data. For a piece of training datatemplate t Use->Original configuration->Configuration goal->And a configuration +.>Filling is carried out.
13b) Modifying data in a fine-tuning dataset using a training hint templateGet training text set +.>
13c) Using D train_text The data set is finely tuned to generate a large model, and the training method used for fine tuning is a lora training method.
2) The model prediction stage based on the multi-objective prompt engineering specifically executes the following steps:
21 User input of original configuration text c) b Configuring target pures and generating number N of candidate configurations gen Wherein configuration targets include, but are not limited to: the system stability and the system reliability are improved, the system is promoted to generate different types of faults by fault injection, and the performance of the system under different loads is tested by system test;
22 User input N) ex Strip configuration generation exampleEach example is shown as example i ＝<c ebi ,c eai ,purp ei >Wherein c ebi Representing an original configuration example, purp ei Representing configuration target examples, c eai Representing configuration examples that meet the configuration objective;
23 Design of multi-objective alert template q For generating a multi-target prompt text,as input to the generative large model. template q Original configuration text c using user input b Configuration target purp and configuration generation exampleFilling;
24 Inputting the multi-target prompt text into the trimmed generation type large model, wherein the result output by the model is a generated candidate configuration;
25 Repeating step 24) N gen Next, get a candidate configuration setRes is returned to the user.
3) The configuration screening stage specifically performs the following steps:
31 For the candidate configuration set Res obtained in step 25), the user selects a candidate configuration set conforming to the configuration target purp therefromAn active configuration set as output;
32 Res obtained in step 31) correct The method and the device can be used for updating configuration generation examples input by a user in the multi-target prompt engineering, and further improve the quality of multi-target prompt texts and the effectiveness of configuration generation. The concrete practice is to put Res correct The configuration generation result in step 22) is added to the configuration generation example input by the user in step 22), and step 22) is repeated to this step. The repetition may be performed a number of times until the resulting configuration achieves the user's satisfaction.
FIG. 2 is a block diagram of a software system configuration generation system based on a generative large model provided by the present invention. The system mainly comprises a configuration data adaptation and model fine adjustment module, a multi-target prompt engineering module and a configuration screening module. Each module is completed through information interaction with other related modules, and different modules are specifically described below.
S1) configuration data adaptation and model fine adjustment module
The main functions of configuration data adaptation and model fine tuning are to use tools such as web crawlers and open APIs to capture configuration data of a multi-source software system in resources such as a webpage and a version control system, wherein the multi-source software system comprises an open source community software system, a target software system and the like. The configuration data is then adapted to the target software system to obtain a fine-tuning data set. And finally, fine tuning the large generated model by using the data set, so that the large generated model is more suitable for configuration generation tasks aiming at the target software system. The module comprises three sub-modules:
s11) configuration data collector
The main function of the configuration data collector is to use tools such as web crawlers and open APIs to grab configuration data in software systems such as an open source community software system and a target software system from resources such as a webpage and a version control system, and to sort the configuration data into a configuration data set. Configuration data satisfaction<S,c b ,c a ,purp>The form of four-tuple, wherein each item represents a source software system, an original configuration text, a configuration text meeting a generation target, and a configuration target text, respectively. Taking a Git version control system as an example, each time the code submits the code base which the code belongs to, namely the source software system, the configuration text before the code is submitted, namely the original configuration text, the configuration text after the code is submitted meets the configuration text of the generation target, and the submitted information is the configuration target. The configuration data collector passes the configuration data to the configuration data adapter.
S12) configuration data adapter
The main function of the configuration data adapter is to resample the configuration data set obtained by the configuration data collector according to the similarity between the configuration data and the configuration data of the target software system so as to obtain a fine tuning data set which is more adaptive to the target software system and is used for training the generated large model. When calculating the similarity between the configuration data and the configuration data of the target software system, a method based on text similarity such as TF-IDF and editing distance can be used for calculating the similarity between the configuration data and the text of the configuration data of the target software system, or a method based on feature engineering and clustering can be used, firstly, indexes for representing the configuration data such as configuration modification line number, configuration modification file number and the like are designed, and then a clustering method such as K-means and the like is used for evaluating the similarity between the configuration data and the configuration data of the target software system. When resampling is carried out, a normalization function is designed, the similarity of each piece of data is mapped to sampling probability between 0 and 1, the sum of the sampling probabilities of all the data is 1, and the calculation can be carried out by using a linear normalization function, an exponential normalization function and the like. The configuration data adapter communicates the resampled trim data set to the model trim trainer.
S13) model fine-tuning trainer
The model fine tuning trainer is mainly used for converting a fine tuning data set into training text to carry out fine tuning on a generated large model. The conversion of the fine tuning data is realized by filling a training prompt template. One possible training hint template is provided below t ：
<comment>Generate configurations to meet given purpose
<comment>Purpose:<purpose>
<comment>Configurations before modification
<configuration before>
<comment>Configurations after modification
<configuration after>
Wherein the method comprises the steps of<comment>A note symbol in the configuration format is represented,<purpose>，<configuration before>and<configuration after>representing the configuration target, the original configuration text and the configuration text satisfying the configuration target, respectively. For one piece of training data rec=<S,c b ,c a ,purp>，<purpose>In response to the push, the program is executed,<configuration before>correspond to c b ，<configuration after>Correspond to c a 。
The training method at the time of fine tuning uses the lora method. The model fine-tuning trainer modifies model parameters of the large model so that the model fine-tuning trainer is more suitable for configuration generation of a target software system.
S2) multi-target prompt engineering module
The main function of the multi-target prompt engineering module is to provide an interface for inputting original configuration, configuration targets and multi-target configuration examples for a user, and to build a multi-target prompt template library, and to quickly generate prompts according to configuration generation requirements and configuration examples provided by the user, so as to guide the generation type large model to generate effective configuration. The module comprises three sub-modules:
s21) configuration requirement receiver
The primary function of the configuration requirement receiver is to receive the original configuration and configuration targets from the user. One possible original configuration for configuring a cluster is given below:
cluster＝[clu_01,clu_02,clu_03,clu_04,clu_06]
connection_pool_size＝300
time_out＝20
alarm_thresh＝0.4
hybrid＝True
one possible configuration goal for this original configuration is: the fault of cluster clu _03 is simulated offline.
The configuration requirement receiver transmits the received original configuration and configuration targets to the multi-target prompt text generator for generating the multi-target prompt text.
S22) configuring an example receiver
The main function of the configuration example receiver is to receive configuration examples or effective configuration sets from a user and a configuration feedback device, and the configuration examples or effective configuration sets are used for providing configuration examples for the multi-target prompt text generator, so that the effectiveness of the multi-target prompt project is improved. Configuration examples belong to the target software system, and each configuration example can be expressed as example=<c eb ,c ea ,purp e >Form (c), wherein c eb Representing an original configuration text example, c ea Configuration text examples representing meeting configuration objectives, purp e Representing a configuration target example. The configuration example receiver communicates the received configuration example text to the multi-target prompt text generator for generation of the multi-target prompt text.
S23) multi-target prompt text generator
The main function of the multi-objective hint text generator is to receive the original configuration, configuration targets and configuration examples from the configuration requirements receiver and the configuration examples receiver and use it to generate multi-objective hint text to guide the generative big model to generate a configuration meeting the configuration targets. The generation of the multi-target prompt text is achieved by populating a multi-target prompt template. One possible multi-objective hint template is provided below:
<comment>Generate configurations to meet given purpose,examples:
<examples>
<comment>Generate configurations to meet given purpose
<comment>Purpose:<purpose>
<comment>Configurations before modification
<configuration before>
<comment>Configurations after modification
where < comment > represents an annotation symbol in the configuration format, < purpose > < configuration before > represents the configuration target text and the original configuration text, respectively, entered by the user. < examples > configuration example text, which is formed by splicing the text converted by each configuration example using the configuration example template. One possible configuration example template is provided below:
<comment>Example
<comment>Purpose:<purpose>
<comment>Configurations before modification
<configuration before>
<comment>Configurations after modification
<configuration after>
where < comment > represents an annotation symbol in a configuration format, < purpose >, < configuration before > and < configuration after > represent a configuration target example, an original configuration example and a configuration example satisfying the configuration target, respectively.
The multi-target prompt text generator inputs the generated multi-target prompt text into the trimmed generation type large model to obtain a candidate configuration set.
S3) configuration screening module
The main function of the configuration screening module is to provide an interface for screening effective configuration meeting configuration targets from a candidate configuration set for a user, and feed back results to the multi-target prompt engineering module so as to optimize multi-target prompt texts and further improve the effectiveness of configuration generation. The module comprises two sub-modules:
s31) configuration selector
The primary function of the configuration selector is to receive a candidate configuration set that generates a large model output, and to provide the candidate configuration set to the user to select a valid configuration that meets the configuration objective. The configuration selector may provide a user graphical interface or programmable interface to enable a user to flexibly and conveniently select an active configuration. The configuration selector passes the user-screened valid configuration set to the configuration feedback.
S32) configuration feedback device
The main function of the configuration feedback device is to output an effective configuration set, and the effective configuration set screened by a user can be fed back to the configuration example receiver so as to optimize the multi-target prompt text and further improve the effectiveness of configuration generation. The configuration feedback may automatically determine whether to pass the active configuration set to the configuration example receiver for iteration based on user selection or an algorithm. The determination can be made, for example, by calculating whether the proportion of valid configurations to all generated configurations, prop, exceeds a preset threshold: when prop < threshold, the candidate configuration set is fed back to the configuration example receiver.
Finally, it should be noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various alternatives and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the disclosed embodiments, but rather the scope of the invention is defined by the appended claims.
Claims (11)
1. The software system configuration generation method based on the generation type large model is characterized by comprising the following three steps:
1) Configuration data adaptation and model fine tuning are specifically as follows: collecting configuration data from an open source community software system and a target software system, performing adaptation of the configuration data according to the similarity between the configuration data and the configuration data of the target software system to obtain a fine-tuning data set, and performing fine tuning of a generated large model by using the data set;
2) The multi-target prompt project specifically comprises: receiving original configuration text and configuration target text data from a user, and a configuration example set, generating a multi-target prompt text, taking the multi-target prompt text as input of a large generation model, and outputting the model to obtain candidate configuration;
3) Configuration screening, specifically: the user screens the valid configuration in the candidate configuration set of the generative large model output as the final configuration generation result or returns it for updating the configuration example.
2. The method for generating a large model based on software system configuration according to claim 1, wherein the configuration data adaptation and model fine tuning specifically comprises the following steps:
11 Collecting multisource system configuration data, the multisource system including software systems in an open source communityAnd a target software system S to be configured and generated 0 ，N s Representing the number of software systems in the open source community; configuration data is obtained from web pages and version control system resources; arrange configuration data into a configuration data set +.>Wherein N is R Representing the number of data bars>Representing a piece of configuration data,/->Representing a software system, c bi Representing the original configuration, c ai Representing a configuration meeting a configuration objective, purp i Representing a configuration target;
12 Adapting the configuration data, resampling the configuration data according to the similarity between the configuration data and the configuration data of the target system, and generating a fine tuning data set;
13 Training the generated large model by using the fine tuning data set, designing a prompt template to modify the format of the fine tuning data so that the mode of the fine tuning data is more similar to the mode of the training data, and training the generated large model by using the modified data so that the generated large model is more suitable for configuration generation tasks.
3. The method for generating a software system configuration based on a generative large model as claimed in claim 1, wherein the multi-objective prompt engineering specifically performs the following steps:
21 User input of original configuration text c) b Configuring target pures and generating number N of candidate configurations gen ；
22 User input N) ex Strip configuration generation exampleEach example is shown as example i ＝<c ebi ,c eai ,purp ei >Wherein c ebi Representing an original configuration example, purp ei Representing configuration target examples, c eai Representing configuration examples that meet the configuration objective;
23 Design of multi-objective alert template q For generating multi-target prompt text as input to a generative large model, template q Original configuration text c using user input b Configuration target purp and configuration generation exampleFilling;
24 Inputting the multi-target prompt text into the trimmed generation type large model, wherein the result output by the model is a generated candidate configuration;
25 Repeating step 24) N gen Next, get a candidate configuration setRes is returned to the user.
4. The method for generating a software system configuration based on a generative large model as claimed in claim 1, wherein the configuration filtering specifically performs the following steps:
31 For the candidate configuration set Res obtained in step 25), the user selects a candidate configuration set conforming to the configuration target purp therefromAn active configuration set as output;
32 Res obtained in step 31) correct The method can be used for updating configuration generation examples input by a user in the multi-target prompt project, and further improving the quality of multi-target prompt texts and the effectiveness of configuration generation; the concrete practice is to put Res correct The configuration generation result in step 22) is added to the configuration generation example input by the user in step 22), and the repeating process from step 22) to this step is repeated a plurality of times until the generated configuration achieves the effect of satisfaction of the user.
5. A software system configuration generation method based on a generative large model as claimed in claim 2, wherein the adapting configuration data in step 12) specifically comprises the steps of:
121 Using the function similarity (rec) i ,R 0 ) Calculating the similarity between the configuration data and the configuration data of the target system, wherein rec i Is one piece of configuration data in R 0 Is configuration data belonging to target systemThe collection, the function return value is from 0 to 1, and the higher the similarity is, the higher the function return value is;
122 Using a normalization function normal to rec each piece of data i Similarity to target system configuration data (rec) i ,R 0 ) Probability p of mapping to single sample i Meets the similarity (rec) i ,R 0 ) The higher p i The higher and
123 N) configuring the data set according to step 122) sampling probability distribution D Subsampling to obtain a scale of N D Fine-tuning data set of (a)Wherein N is D Represents the number of data strips obtained after resampling, and d is more than or equal to 1 i ≤N R 。
6. The method for generating a software system configuration based on a generative large model as claimed in claim 2, wherein the specific method of step 13) is as follows:
131 Design training prompting template t For adjusting model fine-tuning data, for a piece of training datatemplate t Use->Original configuration->Configuration goal->And a configuration +.>Filling;
132 Modifying data in a fine-tuning dataset using a training hint templateGet training text set +.>
133 Using D) train_text The data set is finely tuned to generate a large model, and the training method used for fine tuning is a lora training method.
7. A method of generating a large model based software system configuration according to claim 3, wherein in step 21), the configuration targets include, but are not limited to: system stability, system reliability, fault injection and system testing are improved.
8. A software system configuration generation system based on a generation type large model according to claim 1, comprising a configuration data adaptation and model fine-tuning module, a multi-objective prompt engineering module and a configuration screening module:
the configuration data adaptation and model fine adjustment module is used for collecting configuration data of a multi-source system of the open-source community system and the target system; resampling the configuration data based on the similarity between the configuration data and the configuration data of the target system to obtain a fine tuning data set adapted to the target system; finally, the training prompt template is used for converting the fine adjustment data into training texts for fine adjustment of the generated large model;
the multi-target prompt engineering module is used for providing an interactive prompt engineering interface for a user, generating examples, original configuration texts and configuration target texts according to configuration provided by the user by arranging a multi-target prompt template in the multi-target prompt engineering interface, and inputting the generated prompts into the generation type large model to drive the generated prompts to output candidate configuration sets; in addition, the module receives the effective configuration fed back by the configuration screening module, and updates the configuration generation example by using the effective configuration, so that the effect of the multi-objective prompt engineering is further improved;
and the configuration screening module is used for providing an interface for screening effective configurations meeting configuration targets from the candidate configuration set for a user, wherein the effective configurations can be used as output generated by the configuration and can be used as feedback to be input into the multi-target prompt engineering module.
9. The large model-based software system configuration generation system of claim 8, wherein the configuration data adaptation and model tuning module comprises three sub-modules:
1) The configuration data collector is used for realizing grabbing and collecting of configuration data, sorting the configuration data into a configuration data set and transmitting the configuration data set to the configuration data adapter;
2) The configuration data adapter is used for calculating the similarity between the configuration data and the configuration data of the target software system, and resampling the configuration data set according to the similarity to obtain a fine-tuning data set;
3) The model fine tuning trainer uses a training prompt template to convert fine tuning data into training texts and fine-tune the generated large model.
10. The software system configuration generation system based on a generative large model of claim 8, wherein the multi-objective hint engineering module comprises three sub-modules:
1) The configuration requirement receiver is used for receiving the original configuration and the configuration target from the user, transmitting the original configuration and the configuration target to the multi-target prompt text generator and generating the multi-target prompt text;
2) A configuration example receiver for receiving configuration examples or an effective configuration set from a user and a configuration feedback device in a configuration screening module for providing configuration examples for the multi-objective prompt text generator;
3) And the multi-target prompt text generator is used for receiving the original configuration, the configuration target and the configuration example from the configuration requirement receiver and the configuration example receiver and generating multi-target prompt text by using the multi-target prompt text generator to guide the generation type large model to generate the configuration meeting the configuration target.
11. The software system configuration generation system based on a generative large model of claim 8, wherein the configuration filtering module comprises two sub-modules:
1) The configuration selector is used for receiving the candidate configuration set output by the generated large model, providing the candidate configuration set for a user to select effective configurations meeting configuration targets, and transmitting the effective configuration set screened by the user to the configuration feedback device;
2) And the configuration feedback device is used for outputting the effective configuration set, and automatically judging whether to transmit the effective configuration set to a configuration example receiver in the multi-objective prompt engineering module for iteration according to user selection or algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311002875.0A CN116932037B (en) | 2023-08-10 | 2023-08-10 | Software system configuration generation method and system based on generation type large model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311002875.0A CN116932037B (en) | 2023-08-10 | 2023-08-10 | Software system configuration generation method and system based on generation type large model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116932037A true CN116932037A (en) | 2023-10-24 |
| CN116932037B CN116932037B (en) | 2024-03-15 |
Family
ID=88390805
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311002875.0A Active CN116932037B (en) | 2023-08-10 | 2023-08-10 | Software system configuration generation method and system based on generation type large model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116932037B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140181145A1 (en) * | 2012-12-26 | 2014-06-26 | Jafer S. KAMSAMOHIDEEN | Modular Software System for Use in an Integration Software Technology and Method of Use |
| CN113849814A (en) * | 2020-06-28 | 2021-12-28 | 南京大学 | Configurable system bug reproduction system and reproduction method |
-
2023
- 2023-08-10 CN CN202311002875.0A patent/CN116932037B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140181145A1 (en) * | 2012-12-26 | 2014-06-26 | Jafer S. KAMSAMOHIDEEN | Modular Software System for Use in an Integration Software Technology and Method of Use |
| CN113849814A (en) * | 2020-06-28 | 2021-12-28 | 南京大学 | Configurable system bug reproduction system and reproduction method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116932037B (en) | 2024-03-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110909039A (en) | Big data mining tool and method based on drag type process | |
| CN109325040B (en) | FAQ question-answer library generalization method, device and equipment | |
| CN110502361A (en) | Fine granularity defect positioning method towards bug report | |
| CN117634867B (en) | RPA flow automatic construction method and system combining large language model and reinforcement learning | |
| CN117114250B (en) | Intelligent decision-making system based on large model | |
| CN116932037B (en) | Software system configuration generation method and system based on generation type large model | |
| CN116757369B (en) | Attention mechanism-based carbon emission analysis method and system | |
| CN117033579A (en) | Novel knowledge base system and method based on LLMS | |
| CN115904482A (en) | Interface document generation method, device, equipment and storage medium | |
| CN112712202B (en) | Short-term wind power prediction method and device, electronic equipment and storage medium | |
| CN115080386A (en) | Scene effectiveness analysis method and device based on automatic driving function requirement | |
| CN113657731A (en) | Service supply and demand matching method based on double-sided mode and incidence matrix | |
| CN118394666B (en) | Intelligent business process automation and optimization method based on low-code platform | |
| CN118313463B (en) | Wind power prediction method and system based on space-time enhancement pre-training large language model | |
| CN118092908A (en) | Application program generation method and device based on large language model | |
| CN118250161A (en) | Network configuration automatic management method based on pre-training model | |
| CN116128130A (en) | Short-term wind energy data prediction method and device based on graphic neural network | |
| Park et al. | A Comparison of Human Error Probabilities Collected from HuREX and SHEEP Frameworks | |
| Du et al. | Re 2 GCC: Retrieve-Rerank-Generate Framework for Code Completion | |
| CN114328188A (en) | Model-based spacecraft controller software test case automatic generation method | |
| CN118364801A (en) | Automatic prompting engineering optimization method for large model | |
| CN118626799A (en) | Automatic script generation system | |
| Dong et al. | Research on the Model of Complex Network Based Key Information Extracting in Power Engineering Technical Transformation and Overhaul Project | |
| CN118260689A (en) | Log anomaly detection method based on high-efficiency fine adjustment of self-adaptive low-rank parameters | |
| CN118193711A (en) | Method and system for realizing water business industry knowledge question-answering system based on large language model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |