Method and device for generating task planning information based on computing power of intelligent computing center

Abstract

The invention provides a method and device for generating task planning information based on computing power of an intelligent computing center, and relates to the technical field of intelligent computing centers, intelligent computing centers and computing power infrastructure, and the method comprises the steps: S1, taking a to-be-processed first business demand as a task planning target, and carrying out task planning based on a target model, a function name and / or a parameter name in the mask planning data are / is masked, and the target model is obtained by performing fine adjustment on the initial model based on multiple pieces of fine adjustment data; and S2, based on the mask mapping information corresponding to the multiple pieces of fine adjustment data, mask removal is performed on the function name and the parameter name in the mask planning data to obtain target planning data, and the target planning data is used for being executed by a task planning executor to realize the first business demand. In the model fine tuning stage, the function names and the parameter names are masked, so that the illusion problem of the large model can be inhibited, and the execution reliability of the target planning data output by the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of intelligent computing centers, intelligent computing centers, and computing power infrastructure technologies, and particularly relates to a method and device for generating task planning information for computing power based on an intelligent computing center. Background Art

[0002] With the rapid development of artificial intelligence technology, "intelligent computing centers" and "intelligent computing centers" have emerged as the times require.

[0003] An "intelligent computing center" refers to a facility that uses large-scale heterogeneous computing power resources, including general computing power and intelligent computing power, and mainly provides the required computing power, data, and algorithms for artificial intelligence applications (such as scenarios for artificial intelligence deep learning model development, model training, and model inference, etc.). The intelligent computing center covers facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enabling.

[0004] The "intelligent computing center" includes but is not limited to the "intelligent computing center".

[0005] An "intelligent computing center", that is, an artificial intelligence computing center, is a type of computing power infrastructure that provides computing power services, data services, and algorithm services required for artificial intelligence applications based on artificial intelligence theory and using an artificial intelligence computing architecture.

[0006] "Computing power" is the core of "intelligent computing centers" and "intelligent computing centers". It is the ability of computer devices or computing / data centers to process information, the ability of computer hardware and software to cooperate to jointly execute a certain computing requirement, the computing ability to achieve the output of target results by processing information data, and a new type of productive force that integrates information computing power, network carrying capacity, and data storage capacity. It mainly provides services to society through computing power infrastructure.

[0007] Currently, how to quickly generate an intelligent agent (agent) that matches the requirements is an urgent problem to be solved. A possible solution is: by means of custom coding, a batch of general functions (functions) and the function patterns corresponding to the functions are pre-constructed in advance. This part of the pre-constructed functions can be called base functions (base function), and the function patterns corresponding to the functions can be called base function schemas (base function schema). By providing the base function schema to the large model, so that the large model can master the uses and usages of the base function. The large model, as the business processing core of the pre-generated intelligent agent, can, after the intelligent agent receives the business requirements input by the user, dynamically form a static plan (plan) that meets the above business requirements based on the mastered uses and usages of the base function, and execute the static plan to achieve the response to the business requirements.

[0008] After the large model has mastered the uses and usages of base functions, it is also possible to perform targeted fine-tuning on the large model (also known as model fine-tuning), so as to adjust the model parameters of the large model with the planned (plan) data that has been verified and can be successfully executed by the executor, and standardize the model output format during subsequent applications of the large model, improving the probability that the static plan (plan) output by the large model can be successfully executed by the executor. However, it is found in applications that after performing targeted fine-tuning on the large model based on related technologies, the probability that the static plan (plan) output by the large model can be successfully executed by the executor is still not as expected.

[0009] It can be seen that based on the method of using a large model to generate task planning information in the prior art, the generated static plan has a very low execution reliability problem. Summary of the Invention

[0010] The purpose of the present invention is to provide a method and device for generating task planning information based on the computing power of an intelligent computing center, which is used to solve the problem that the execution reliability of the static planning information generated by the prior art is very low.

[0011] To solve the above technical problems, the present invention is implemented as follows:

[0012] In a first aspect, the present invention provides a method for generating task planning information based on the computing power of an intelligent computing center, including:

[0013] Step S1: Using the first service requirement to be processed as the task planning target, perform task planning based on the target model to obtain masked planning data. Among them, the service corresponding to the first service requirement is a computer vision service, the function names and / or parameter names in the masked planning data are masked, and the target model is a model obtained by fine-tuning the initial model based on multiple fine-tuning data; the fine-tuning data includes a second service requirement and true-value planning data, the true-value planning data includes multiple true-value data bodies, the multiple true-value data bodies correspond one-to-one to multiple true-value service steps, the multiple true-value service steps are used to implement the second service requirement, and the function names and / or parameter names in the true-value data bodies are masked according to the set target masking rules;

[0014] Step S2: Based on the masked mapping information corresponding to the multiple fine-tuning data, unmask the function names and parameter names in the masked planning data to obtain target planning data, and the target planning data is used to be executed by the task planning executor to implement the first service requirement.

[0015] In one embodiment, the target masking rules include:

[0016] Replace the first function name in the true value data body including the first function name based on the first mask character corresponding to the first function name, where the first function name is any function name in the first set, and the first set includes: the function names included in each of the multiple fine-tuning data; the first mask character is generated based on the function name mask character and the first serial number, and the first serial number is used to uniquely identify the first function name in the first set;

[0017] Replace the first parameter name in the true value data body including the first parameter name based on the second mask character corresponding to the first parameter name, where the first parameter name is any parameter name in the second set, and the second set includes: the parameter names included in each of the multiple fine-tuning data; the second mask character is generated based on the parameter name mask character and the second serial number, and the second serial number is used to uniquely identify the first parameter name in the second set.

[0018] In one embodiment, the first mask character is obtained by splicing the function name mask character and the first serial number, and the splicing position of the function name mask character is before the splicing position of the first serial number;

[0019] The second mask character is obtained by splicing the parameter name mask character and the second serial number, and the splicing position of the parameter name mask character is before the splicing position of the second serial number;

[0020] The function name mask character is used to limit the initial model's association with function information, and the parameter name mask character is used to limit the initial model's association with parameter information.

[0021] In one embodiment, before the step S1, the method further includes:

[0022] Step S01: Fine-tune the initial model respectively according to multiple groups of fine-tuning data sets to obtain multiple candidate models, where the multiple candidate models and the multiple groups of fine-tuning data sets correspond one by one, the fine-tuning data set includes multiple candidate fine-tuning data, different groups of fine-tuning data sets correspond to different candidate mask rules, and different candidate mask rules include different parameter name mask characters and / or function name mask characters;

[0023] Step S02: Use a pre-set test data set to perform task planning tests on the multiple candidate models respectively to obtain multiple test results, and the test results are used to represent: the ratio of the successfully executed task planning data output by the corresponding candidate model;

[0024] Step S03: Among the multiple candidate models, determine the candidate model corresponding to the highest ratio as the target model.

[0025] In one embodiment, the fine-tuning data also includes restriction text, and the restriction text is used to indicate that the initial model is prohibited from performing semantic parsing on the mask data, and the mask data includes: the masked function names and / or parameter names in the true value planning data.

[0026] In one embodiment, the reading position of the restriction text in the fine-tuning data is a first position, the reading position of the true value planning data corresponding to the restriction text in the fine-tuning data is a second position, and the first position is located before the second position.

[0027] In a second aspect, the present invention further provides a device for generating task planning information based on the computing power of an intelligent computing center, comprising:

[0028] A task planning module is used to take a first business requirement to be processed as a task planning target, perform task planning based on a target model, and obtain mask planning data, wherein the business corresponding to the first business requirement is a computer vision business, the function names and / or parameter names in the mask planning data are masked, and the target model is a model obtained by fine-tuning an initial model based on multiple fine-tuning data; the fine-tuning data includes a second business requirement and true value planning data, the true value planning data includes multiple true value data bodies, the multiple true value data bodies correspond one-to-one to multiple true value business steps, the multiple true value business steps are used to implement the second business requirement, and the function names and / or parameter names in the true value data body are masked according to the set target masking rules;

[0029] The mask unmasking module is used to unmask the function names and parameter names in the mask planning data based on the mask mapping information corresponding to the multiple fine-tuning data to obtain target planning data, and the target planning data is used to be executed by the task planning executor to achieve the first business requirement.

[0030] In a third aspect, the present invention further provides a server comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the steps in the method for generating task planning information based on the computing power of an intelligent computing center as described in the first aspect above are implemented.

[0031] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps in the method for generating task planning information based on the computing power of an intelligent computing center as described in the first aspect above are implemented.

[0032] Fifth aspect, the present invention provides a computer program product, including computer instructions, which when executed by a processor, implement the steps in the method for generating task planning information based on the computing power of an intelligent computing center as described in the first aspect above.

[0033] In the present invention, according to the set target masking rule, mask the function names and / or parameter names in the planning data to form masked planning data. Then, use the masked planning data and the corresponding second service requirements to fine-tune the initial model to suppress the model's excessive association with function names and parameter names, thereby suppressing the large model hallucination problem. This can improve the model fine-tuning effect of the initial model, increase the probability that the planning data output by the target model after fine-tuning is successfully executed, that is, improve the execution reliability of the target planning data. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0035] Figure 1 is an architecture diagram of a two-stage agent automated workflow provided by the present invention;

[0036] Figure 2 is a schematic diagram of the results of an automated workflow of an experimental agent using Qwen2.5 - 72B - Instruct provided by the present invention;

[0037] Figure 3 is a flowchart of a method for generating task planning information through the computing power of an intelligent computing center provided by the present invention;

[0038] Figure 4 is a structural diagram of a device for generating task planning information through the computing power of an intelligent computing center provided by the present invention;

[0039] Figure 5 is a structural diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The following will clearly and completely describe the technical solutions in the present invention with reference to the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without making creative efforts belong to the scope of protection of the present invention.

[0041] First, the technical terms related to the present invention will be briefly described below.

[0042] The "computing power" referred to in the present invention is the ability of a computer device or a computing / data center to process information. It is the ability of computer hardware and software to cooperate to jointly execute a certain computing requirement. It is the computing ability to process information data and output a target result. It is a new type of productive force integrating information computing power, network carrying capacity, and data storage capacity, and mainly provides services to society through computing power infrastructure.

[0043] The "computational power" (Computational Power, CP) referred to in the present invention is an ability of a data center server to process data and output a result. It is a comprehensive index to measure the computing ability of a data center, including general computing ability, supercomputing ability, and intelligent computing ability. The commonly used measurement unit is the number of floating-point operations per second (FLOPS: Floating Point Operations Per Second, 1EFLOPS = 10^18 FLOPS). The larger the value, the stronger the comprehensive computing ability. It is estimated that 1EFLOPS is approximately the computing power output of 5 Tianhe 2A or 500,000 mainstream server CPUs or 2 million mainstream laptops. The calculation formula is: CP = CP 通用 +CP 智能 +CP 超级 .

[0044] The "network power" (Network Power, NP) referred to in the present invention is the performance of the data transmission ability of computing power facilities, including comprehensive capabilities such as network architecture, network bandwidth, transmission delay, intelligent management and scheduling, etc. The network power involves network transmission inside and between data centers and is a comprehensive index to measure the network transmission scheduling ability. In the present invention, the video memory bandwidth is used for the network power.

[0045] The "storage power" (Storage Power, SP) referred to in the present invention is the comprehensive ability of a data center in four aspects: data storage capacity, performance, security and reliability, and green and low-carbon. It is a comprehensive index to measure the data storage ability of a data center, including external storage devices such as storage arrays and server internal storage devices. The commonly used measurement unit for storage capacity is exabyte (EB, 1EB = 2^60 bytes), the commonly used measurement unit for performance is the number of read / write operations per second per unit capacity (Input / Output Operations Per Second / TB, IOPS / TB), and the disaster recovery ratio is an important manifestation of security and reliability.

[0046] The "computing power infrastructure" described in the present invention is a new type of information infrastructure that integrates information computing power, network carrying capacity, and data storage capacity, and can realize the centralized computing, storage, transmission, and application of information.

[0047] The "new type of information infrastructure" described in the present invention refers to: mainly including network infrastructures such as 5G networks, fiber broadband networks, backbone networks, international communication networks, and satellite Internet, computing power infrastructures such as data centers, general computing power centers, intelligent computing centers, and supercomputing centers, and new technology facilities such as artificial intelligence, blockchain, and quantum computing.

[0048] The "computing power" described in the present invention includes general computing power, intelligent computing power, and super computing power.

[0049] The "general computing power" described in the present invention is the computing power provided by servers based on Central Processing Unit (CPU) chips, and is used to support basic general computing such as cloud computing and edge computing.

[0050] The "intelligent computing power" described in the present invention is for various artificial intelligence innovation applications, and is a computing platform deployed on a large scale based on dedicated chips such as Graphics Processing Unit (GPU), Field Programmable Gate Array (FPGA), and Application Specific Integrated Circuit (ASIC), such as natural language processing and machine vision.

[0051] The "super computing power" described in the present invention is mainly the computing power provided by high-performance computing clusters such as supercomputers. It utilizes the centralized computing resources of multiple computer systems working in parallel and processes extremely complex or data-intensive problems through a dedicated operating system, and is mainly used for computing in cutting-edge scientific fields, such as planetary simulation, drug molecule design, and gene analysis.

[0052] The "intelligent computing center" described in the present invention refers to a facility that mainly provides the required computing power, data, and algorithms for artificial intelligence applications (such as scenarios of artificial intelligence deep learning model development, model training, and model inference) by using large-scale heterogeneous computing power resources, including general computing power (CPU) and intelligent computing power (GPU, FPGA, etc.). The intelligent computing center covers facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enabling.

[0053] The "intelligent computing center" described in the present invention includes but is not limited to the "intelligent computing center".

[0054] The "Intelligent Computing Center" described in the present invention, namely the artificial intelligence computing center, is a type of computing infrastructure based on artificial intelligence theory, adopting an artificial intelligence computing architecture, and providing computing power services, data services, and algorithm services required for artificial intelligence applications.

[0055] The "Computing Power Center" described in the present invention refers to: a facility mainly composed of infrastructure such as wind, fire, water, and electricity, and IT software and hardware devices, with computing power, carrying capacity, and storage capacity, including general data centers, intelligent computing centers, supercomputing centers, etc.

[0056] The "Supercomputing Center" described in the present invention refers to: namely the supercomputing data center, which is a data center based on supercomputers or large-scale computing clusters, capable of providing functions such as large-scale computing, storage, and network services, and is widely used in application scenarios such as aerospace, national defense, oil exploration, climate modeling, and genome sequencing.

[0057] The "Computing Power Resources" described in the present invention refer to the technologies and facilities required for the development of the digital society, with information computing, transmission, storage, and application capabilities, including but not limited to computing resources such as CPUs and GPUs, network resources such as switches and routers, storage resources such as storage arrays and distributed storage, security resources such as firewalls and intrusion detection systems, and support and guarantee resources such as wind, fire, water, and electricity.

[0058] The "models" and "large models" described in the present invention include but are not limited to "large language models" and "multimodal large models".

[0059] The "large language model" described in the present invention refers to the large language model (LLM), which is a language model with a relatively large number of parameters, aiming to understand and generate human language, and can be fine-tuned through a large amount of text data to perform a wide range of tasks including text summarization, translation, sentiment analysis, etc.

[0060] The "multimodal large models" described in the present invention refer to: models that jointly fine-tune multimodal information such as text, images, videos, and audio, including but not limited to multimodal large language models.

[0061] The "agent" described in the present invention refers to an entity that can perceive the environment and take actions to achieve specific goals. It can be software, hardware, or a system, and has the characteristics of autonomy, adaptability, and interaction ability. The agent perceives changes in the environment (such as through sensors or data input), makes judgments and decisions based on the knowledge and algorithms it has learned, and then executes actions to affect the environment or achieve the predetermined goals. Agents are widely used in the field of artificial intelligence and are commonly found in automated systems, robots, virtual assistants, and game characters, etc. The core lies in their ability to learn autonomously and continuously evolve to better complete tasks and adapt to complex environments.

[0062] The "mask" described in the present invention refers to the operation of replacing the original data with other data different from the original data, and, relative to the original data, the other data results in a lower probability of the model hallucinating. For example, the other data can be obtained by abstracting the original data or the object represented by the original data (such as when the original data is detection referring to a function name, the other data obtained by abstraction can be function), or, simple characters without specific semantics can also be used as the other data (such as when the original data is detection referring to a function name, the simple character used as the other data can be the character D).

[0063] The "mask removal" described in the present invention refers to the operation of restoring the other data back to the original data after replacing the original data with other data different from the original data.

[0064] The "execution reliability" described in the present invention refers to the probability that the planned data can be successfully executed by the task planning executor.

[0065] The "base function" described in the present invention refers to the functions pre-constructed by developers according to the general / high-frequency business operations of the business scenarios faced by the agent. For example, when the agent is specifically a computer vision agent, the general / high-frequency business operations of the business scenarios faced by the agent may include: image preprocessing operations, image segmentation operations, image noise filtering operations, image target detection operations, etc.

[0066] The "base function schema" described in the present invention refers to the "base function user manual" formed by developers based on the functions and usages of base functions. The base function schema may include items such as name, description, type, etc. Among them, the name item may include the function name, the input parameter names of the function, the output parameter names of the function, etc. The description item can be understood as the function description, the semantic description of the input parameters of the function, and the semantic description of the output parameters of the function; the type item includes the data format description of the input parameters of the function and the data format description of the output parameters of the function. The large model can master the uses and usages of the base functions through the "base function schema".

[0067] The "planner" described in the present invention refers to a large model for task planning. After the task rule large model masters the uses and usages of the base functions through the base function schema, it can be used to generate a static plan in one go. The static plan contains all the business processing steps to complete the business requirements, and each business processing step includes its explanation (also known as the step description text), the method name called, and the input parameter information (each input parameter and its corresponding assignment).

[0068] The "autocoder" described in the present invention refers to a large model for code writing, which is used to write adapter functions.

[0069] The "static planning information" described in the present invention refers to the data representation of the static plan, which at least includes a plurality of step data bodies arranged in an orderly manner. The plurality of step data bodies correspond one-to-one to a plurality of business processing steps, and the plurality of business processing steps are executed in sequence to complete the user's business requirements.

[0070] The "step data body" described in the present invention is the data representation of a corresponding business processing step, including but not limited to: the step description text of a corresponding business processing step (used to describe the business processing step to indicate the function, purpose, and / or execution method of the step, so as to facilitate users or developers to understand and implement), the step number of a corresponding business processing step among the plurality of business processing steps, the input data of a corresponding business processing step (such as the data name, data type requirements, data source, etc. of the input data), and the data processing method of a corresponding business processing step (calling a specific method through the method name).

[0071] The "pipeline" described in the present invention refers to the process of executing the static plan.

[0072] The "true value data body" described in the present invention can be understood as a special "step data body", and its special feature lies in that the function names and / or parameter names included in the data body are masked.

[0073] The "task planning information" described in the present invention can be understood as: taking the service input by the user as the input, obtaining through task planning by the task planning large model, and being the planning data that can be executed by the task planning executor (in the present invention, the target planning data can be understood as an example of the task planning information).

[0074] Regarding the problem of quickly generating an intelligent agent (agent) that matches the user's needs, the present invention provides an intelligent agent automated workflow solution, which specifically is: by means of custom coding, a batch of general functions (function) and the function patterns corresponding to the functions are pre-constructed in advance. This part of the pre-constructed functions can be called base functions (basefunction), and the function patterns corresponding to the functions can be called base function schemas (base function schema). By providing the base function schema to the large model, so that the large model can master the uses and usages of the base function. The large model, as the business processing core of the pre-generated intelligent agent, can, after the intelligent agent receives the business requirements input by the user, based on the mastered uses and usages of the base function, dynamically form a static plan (plan) that meets the above-mentioned business requirements, and execute the static plan to achieve the response to the business requirements.

[0075] In an example, the powerful computing power and flexible resource configuration of the intelligent computing center are utilized to implement the foregoing intelligent agent automated workflow solution. The corresponding architecture of the two-stage auto-pipeline can be as Figure 1 shown, Figure 1 The relevant illustrations in

[0076] User requirement (one user query): The user inputs their request or requirement content.

[0077] Task planning large model (planner): Used to parse user requirements and formulate an execution plan (plan). If the parsing fails, it may output an error that cannot be converted into the json format.

[0078] Plan list: A specific step list generated according to user requirements.

[0079] Retrieve all-step function dependencies: Retrieve the function dependencies related to each step to ensure that all required functions are available.

[0080] Obtain executable input values:

[0081] Given values: Values directly provided to the function.

[0082] User inputs: Dynamic input values provided by the user in the query.

[0083] Previous function outputs: Function output values from the previous step.

[0084] Media download: This step can also support downloading additional required media content.

[0085] Execute one step: Refers to the indication of the execution details of a specific step.

[0086] Execute function: Used to actually call the function to perform data processing work according to the obtained input values.

[0087] Base function: Executes specific basic functions to support the agent to have the set basic business processing capabilities.

[0088] Adapter function (autoencoder): As a supplement to the base function.

[0089] Retrieve depended function schemas: Obtain the structures or schemas of other dependent functions.

[0090] Autoencoder coding: Automatically generate the code involved in the above process.

[0091] Exception code: If an error or exception occurs, the system will return a specific exception code for subsequent analysis and debugging. For example:

[0092] Exceptions can be classified into the following 6 categories:

[0093] 0: No exception, the pipeline execution is successful

[0094] 1: Other pipeline exceptions

[0095] 2: Autoencoder execution fails

[0096] 3: Failure in executing basic functions

[0097] 4: Failure in obtaining the plan (usually caused by failure in converting the plan to JSON)

[0098] 5: Incorrect expression of variable placeholders in the plan, parsing failure.

[0099] It should be noted that the two-stage auto-pipeline has at least the following advantages:

[0100] 1. Higher accuracy: There are more cases where the given plan can be successfully executed;

[0101] 2. By decomposing tasks, the total inference time of the LLM can be shortened;

[0102] 3. By manually handling the core information transfer between steps, the scope of effective information can be clarified and the transparency of the system can be increased.

[0103] The results of experimenting with the two-stage auto-pipeline using the Qwen2.5-72B-Instruct tool for different data sources are as Figure 2 shown.

[0104] Experiments found that errors mainly occurred in the basic function execution stage and the variable placeholder expression stage. Among them, the errors in the basic function execution stage included: incorrect parameter passing, confusion in using the picture list and video path, and some parameters unrelated to the input (such as thresholds), etc.; the errors in the variable placeholder expression stage included: importing external packages (lacking package dependencies), incorrect understanding and handling of complex parameters, etc. In addition, some errors were caused by poor compliance with the JSON format output by the plan.

[0105] In the above solution, the task planning large model is obtained by pre-training and fine-tuning the large model. Among them, the large model after pre-training is the initial model described in the present invention. The initial model has a certain task planning ability. The initial model can output corresponding planning data (plan) for the input business requirements (also known as user requirements). In the application, the available planning data and the corresponding business requirements are used as fine-tuning data to further fine-tune the model parameters and the model output format of the initial model, which can standardize the data format of the planning data output by the model and increase the probability that the planning data output by the model can be successfully executed by the task planning executor. Among them, the available planning data refers to the planning data that can be successfully executed by the task planning executor. The task planning executor is used to execute the planning data to generate an execution result for responding to the corresponding business requirement.

[0106] In the prior art, using available planning data and corresponding business requirements as fine-tuning data to further fine-tune the model parameters and model output format of the initial model, the obtained model fine-tuning effect is lower than expected. More specifically, for the task planning large model obtained based on the fine-tuning solution provided by the prior art, the execution reliability of the output planning data is very low.

[0107] Aiming at the above problems, that is, the model fine-tuning effect of the model fine-tuning solution provided by the prior art is poor, resulting in a very low execution reliability of the static planning information output by the fine-tuned model, the present invention provides a method for generating task planning information based on the computing power of an intelligent computing center. On the premise of utilizing the powerful computing power of the intelligent computing center, in the model fine-tuning stage, by masking the function names and / or parameter names in the planning data, the over-association of the model with function names and parameter names is suppressed, that is, the large model hallucination problem is suppressed, the model fine-tuning effect of the initial model is improved, and further the probability that the target model outputs planning data that can be successfully executed by the task planning executor is increased, that is, the execution reliability of the output planning data is improved.

[0108] Please refer to Figure 3 , Figure 3 which is a method for generating task planning information based on the computing power of an intelligent computing center provided by the present invention. As Figure 3 shown, it includes the following steps:

[0109] Step S1: Using the to-be-processed first business requirement as the task planning target, perform task planning based on the target model to obtain masked planning data.

[0110] Among them, the business corresponding to the first business requirement is a computer vision business. The function names and / or parameter names in the masked planning data are masked. The target model is a model obtained by fine-tuning the initial model based on multiple fine-tuning data. The fine-tuning data includes a second business requirement and true-value planning data. The true-value planning data includes multiple true-value data bodies. The multiple true-value data bodies correspond one-to-one to multiple true-value business steps. The multiple true-value business steps are used to implement the second business requirement. The function names and / or parameter names in the true-value data bodies are masked according to the set target masking rules.

[0111] Exemplarily, the computer vision business can be an image recognition business, an object detection business, an image segmentation business, a pose estimation business, a face recognition business, an image generation business, a video analysis business, an augmented reality business, an autonomous driving business, etc.

[0112] The above first business requirement is used to represent the actual business requirement input by the user to the computer vision agent during actual application. It should be noted that the computer vision agent uses the target model as the core for business processing.

[0113] In the application, any business requirement (such as the first business requirement or the second business requirement) at least includes the image / image set to be processed (or the acquisition address of the image / image set to be processed), and the information expected to be obtained from the image / image set to be processed. For example, the first business requirement may be: to perform oil stain detection on the image / image set input by the user to determine whether the input image / image set includes oil stains; or to perform face recognition and verification on the image / image set provided by the user, etc., to determine whether the provided image / image set includes the set face.

[0114] The task planning referred to in the present invention should be understood as the process of analyzing the business requirement to determine the requirement goal, several steps to achieve the requirement goal, and forming planning data accordingly.

[0115] The above second business requirement can be understood as a business requirement for simulating the first business requirement. The true value planning data is the aforementioned available planning data, and after the true value planning data is executed by the task planning executor, the corresponding second business requirement can be realized.

[0116] The true value data body is the data representation of a corresponding business step, which at least includes: the first data and the second data. Among them, the first data is used to represent: the function name of the execution function for realizing the corresponding business step, and the function function to be completed by the execution function; the second data is used to represent: the parameter name of the input parameter of the aforementioned execution function, the parameter name of the output parameter, the parameter format of the input parameter, and the parameter format of the output parameter.

[0117] Exemplarily, in the case where no masking is performed, if a function name in the planning data is "preprocess" (preprocessing), and the function description corresponding to this function name is "Adjust the image size of the input image to the standard size, and the standard size is [800, 600]", then based on this planning data, the initial model is fine-tuned. The initial model may over-associate the function based on the function name "preprocess", and then generate redundant function functions (such as image denoising, image binarization, image rotation, etc.) that exceed the established function function (referring to image size standardization). By masking the function name "preprocess", it is possible to suppress or even avoid the over-association of the initial function based on the function name, improve the probability that the initial function generates a data body identical to the established function function, that is, improve the following effect of the initial model on the fine-tuning data during the fine-tuning process, and enable the initial model to obtain a better model fine-tuning effect.

[0118] It should be noted that in the above fine-tuning data, only the function name in the true value data body can be masked, only the parameter name in the true value data body can be masked, or both the function name and the parameter name in the true value data body can be masked.

[0119] Step S2, based on the masking mapping information corresponding to the multiple fine-tuning data, unmask the function name and parameter name in the masked planning data to obtain target planning data, and the target planning data is used to be executed by the task planning executor to implement the first service requirement.

[0120] Among them, the masking mapping information corresponding to the multiple fine-tuning data includes: in the multiple fine-tuning data, the function masking mapping data and parameter masking mapping data corresponding to each fine-tuning data. Among them, the function masking mapping data includes a function name and the name after the function name is masked, and the parameter name mapping data includes a parameter name and the name after the parameter name is masked.

[0121] In the present invention, unmasking can be understood as: based on the masking mapping information corresponding to the multiple fine-tuning data, in the masked planning data, restoring the masked function name and parameter name to the name before masking.

[0122] Specifically, the task planning executor can obtain an execution result by executing the target planning data, and this execution result can be used to form a demand response result, and by returning this demand response result to the user, the response to the first service requirement is completed.

[0123] In the present invention, by masking the parameter names and / or function names in the true value data body, the excessive association of the initial model with the function names and parameter names can be suppressed during the model fine-tuning process. More specifically, the interference of the semantics of the function names on the function functions understood by the initial model is reduced, and the interference of the semantics of the parameter names on the parameter formats understood by the initial model is reduced, thereby suppressing the large model hallucination problem, improving the model fine-tuning effect of the initial model, and thereby improving the probability of the planning data output by the target model being successfully executed.

[0124] In one embodiment, the target mask rule includes:

[0125] Based on a first mask character corresponding to a first function name, replacing the first function name in a truth value data body including the first function name, wherein the first function name is any function name in a first set, the first set including: a function name included in each fine-tuning data in the plurality of fine-tuning data; the first mask character is generated based on the function name mask character and a first sequence number, the first sequence number being used to uniquely identify the first function name in the first set;

[0126] Based on a second mask character corresponding to a first parameter name, the first parameter name is replaced in a true value data body including the first parameter name, wherein the first parameter name is any parameter name in a second set, and the second set includes: a parameter name included in each fine-tuning data in the multiple fine-tuning data; the second mask character is generated based on the parameter name mask character and a second serial number, and the second serial number is used to uniquely identify the first parameter name in the second set.

[0127] It should be noted that the function name mask characters used in the masking process for different function names may be the same or different; similarly, the parameter name mask characters used in the masking process for different parameter names may be the same or different.

[0128] In some embodiments, using the same function name masking character to mask multiple function names in the first set, and using the same parameter name masking character to mask multiple parameter names in the second set can further reduce the probability of the initial model over-associating with the masked parameter names and function names, simplify the masking process, and improve the model fine-tuning effect of the initial model, so that the probability of the target planning data output by the target model being successfully executed is further improved.

[0129] Among them, in order to support subsequent mask unmasking processing to complete the coordination between the task planning large model (i.e., the target model) and the task planning executor, the first function name is masked using the corresponding first serial number combined with the aforementioned function name mask character, and the first parameter name is masked using the corresponding second serial number combined with the aforementioned parameter name mask character.

[0130] In one example, the first serial number is used to represent the order in which the first function name is stored in the first set; similarly, the second serial number is used to represent the order in which the first parameter name is stored in the second set. For example, when the first function name is the 6th function name stored in the first set, the first serial number is 6; similarly, when the first parameter name is the 1st parameter name stored in the second set, the second serial number is 1.

[0131] In the application, it is preferably set that the function name masking character is only used to represent the semantics of the function name, and the parameter name masking character is only used to represent the semantics of the parameter name, so as to ensure that the masked function name and parameter name can be accurately recognized as the function name and parameter name by the initial model, while avoiding the introduction of interference information and ensuring that the initial model can obtain a better model fine-tuning effect.

[0132] For example: the function name masking character can be "function", and the parameter name masking character can be "argument".

[0133] In one embodiment, the first masking character is obtained by splicing the function name masking character and the first serial number, and the splicing position of the function name masking character is before the splicing position of the first serial number;

[0134] The second masking character is obtained by splicing the parameter name masking character and the second serial number, and the splicing position of the parameter name masking character is before the splicing position of the second serial number;

[0135] The function name masking character is used to limit the initial model's association with function information, and the parameter name masking character is used to limit the initial model's association with parameter information.

[0136] In the present invention, function information is used to represent the function of a function, and parameter information is used to represent the semantics and parameter format of a parameter.

[0137] In this embodiment, the first masking character is obtained by splicing the function name masking character and the first serial number, and the second masking character is obtained by splicing the parameter name masking character and the second serial number. It is specified that the splicing position of the function name masking character is before the splicing position of the first serial number, and the splicing position of the parameter name masking character is before the splicing position of the second serial number, so as to conveniently complete the generation of the first masking character and the second masking character, avoid introducing unnecessary interference information, facilitate the initial model's recognition of the first masking character and the second masking character, reduce the probability of the initial model misrecognizing the first masking character and the second masking character, and ensure that the initial model can obtain a better model fine-tuning effect.

[0138] In the present invention, during the splicing process, delimiters (such as "_", ".", etc.) can be used to separate the two elements to be spliced, so as to further facilitate the initial model to recognize the first masked character and the second masked character, and reduce the probability of the initial model misrecognizing the first masked character and the second masked character.

[0139] For example, if a function name is detection and the masked name of the function is function_1, then the function mask mapping data corresponding to the function name is [detection, function_1]; similarly, if a parameter name is image_path_list and the masked name of the parameter is argument_2, then the parameter mask mapping data corresponding to the parameter name is [image_path_list, argument_2]. In this example, function can be understood as the aforementioned function name masked character, argument can be understood as the aforementioned parameter name masked character, the number 1 can be understood as the aforementioned first serial number, the number 2 can be understood as the aforementioned second serial number, and "_" can be understood as the aforementioned delimiter.

[0140] In one embodiment, before the step S1, the method further includes:

[0141] Step S01: Fine-tune the initial model respectively according to multiple groups of fine-tuning data sets to obtain multiple candidate models.

[0142] Among them, the multiple candidate models and the multiple groups of fine-tuning data sets correspond one by one. The fine-tuning data set includes multiple candidate fine-tuning data. Different groups of fine-tuning data sets correspond to different candidate masking rules, and different candidate masking rules include different parameter name masked characters and / or function name masked characters.

[0143] Step S02: Use a preset test data set to perform task planning tests on the multiple candidate models respectively to obtain multiple test results.

[0144] The test results are used to represent the ratio of the successfully executed task planning data output by the corresponding candidate model.

[0145] Step S03: Among the multiple candidate models, determine the candidate model corresponding to the highest ratio as the target model.

[0146] Among them, the candidate fine-tuning data also includes business requirements and corresponding planning data.

[0147] In this embodiment, different candidate masking rules may include: a first candidate masking rule, a second candidate masking rule, a third candidate masking rule, and a fourth candidate masking rule.

[0148] Among them, the first candidate masking rule uses the first type of characters as the parameter name masking characters and / or function name masking characters. The first type of characters are minimalist characters (referring to characters with a number of characters / data volume less than a set threshold). For example, the character "F" is used as the function name masking character, and the character "A" is used as the parameter name masking character.

[0149] The second candidate masking rule uses the second type of characters as the parameter name masking characters and / or function name masking characters. The second type of characters are random characters (generated according to the random character generation logic, which may include English, numbers, special symbols, etc.). For example, the character "DASF2&32*" is used as the function name masking character, and the character "JMKGR8=!" is used as the parameter name masking character.

[0150] The third candidate masking rule uses the third type of characters as the parameter name masking characters and / or function name masking characters. The third type of characters are only used to represent the semantics of the function name and the semantics of the parameter name. For example, the character "function" is used as the function name masking character, and the character "argument" is used as the parameter name masking character.

[0151] The fourth candidate masking rule uses the fourth type of characters as the parameter name masking characters and / or function name masking characters. The fourth type of characters are characters refined according to the corresponding function description and parameter description. For example, when the corresponding function description is "Check a single or a group of pictures and determine the bounding boxes of each picture for the query object category", the character refined from this function description is "Image annotation", then the character "Image annotation" will be used as the function name masking character.

[0152] The specific ratio of the successfully executed task planning data output by the foregoing corresponding candidate model is: among the multiple task planning data output by the corresponding candidate model, the ratio of the task planning data successfully executed by the task planning executor to the number of the multiple task planning data.

[0153] It should be noted that each test data in the test dataset includes at least one test requirement for testing the model. In this case, if a candidate model uses the test requirement as the model output and can output test planning data successfully executed by the task planning executor, it is considered that the candidate model passes a test, and this test planning data is also used as the task planning data successfully executed by the task planning executor.

[0154] In this embodiment, different fine-tuning datasets corresponding to different candidate mask rules are pre-configured, and the same initial model is fine-tuned separately to obtain multiple candidate models. Then, the same test dataset is used to test the multiple candidate models respectively, and the candidate model with the best test effect is determined as the target model to ensure the model fine-tuning effect obtained by the initial model.

[0155] In some embodiments, it can be further set that the test data further includes real planning data that matches the test requirements. In this case, a candidate model takes the test requirements as the model input and can output test planning data that can be successfully executed by the task planning executor. And if the data similarity between the test planning data and the foregoing real planning data exceeds the set similarity threshold, it is regarded that the candidate model passes a test. The test result is also used to represent the ratio of the number of times the corresponding candidate model passes the test to the total number of tests. The target model is the candidate model corresponding to the highest ratio among the multiple candidate models.

[0156] For example, the test semantic features can be obtained by extracting semantic features from the test planning data, the real semantic features can be obtained by extracting semantic features from the real planning data, the feature similarity between the test semantic features and the real semantic features is calculated, and this feature similarity is used as the foregoing data similarity.

[0157] In one embodiment, the fine-tuning data further includes restricted text, and the restricted text is used to indicate that the initial model is prohibited from performing semantic parsing on the masked data. The masked data includes the function names and / or parameter names masked in the true value planning data.

[0158] Among them, the restricted text can be: It is prohibited to use target data to participate in function generation and parameter generation, and the target data is character data including function name mask characters or parameter name mask characters.

[0159] In this embodiment, based on the setting of the restricted text, the planning constraint on the initial model can be realized, thereby reducing the probability of the initial model generating model hallucinations in the fine-tuning stage, so that the initial model can obtain a better model fine-tuning effect.

[0160] In one embodiment, the reading position of the restricted text in the fine-tuning data is the first position, and the reading position of the true value planning data corresponding to the restricted text in the fine-tuning data is the second position, and the first position is before the second position.

[0161] Among them, the reading position of the restriction text in the fine-tuning data is: the setting position of the restriction text in the fine-tuning data; the reading position of the true value planning data corresponding to the restriction text in the fine-tuning data is: the setting position of the true value planning data corresponding to the restriction text in the fine-tuning data.

[0162] It should be understood that the initial model will read the contents of the fine-tuning data sequentially, so the earlier the setting position is, the earlier the corresponding content will be read by the initial model.

[0163] In this embodiment, the restriction text is placed before the corresponding true value rule data so that the initial model can learn the planning constraints represented by the restriction text in a timely manner before adjusting the model parameters with reference to the true value planning data, thereby reducing the probability of the initial model generating a large model hallucination during the fine-tuning stage, and enabling the initial model to obtain a better model fine-tuning effect.

[0164] See also Figure 4 , Figure 4 is a schematic diagram of the structure of a device for generating task planning information based on the computing power of an intelligent computing center provided by the present invention, such as Figure 4 As shown, the device 400 for generating task planning information based on the computing power of the intelligent computing center includes:

[0165] The task planning module 401 is used to take the first business requirement to be processed as the task planning target, perform task planning based on the target model, and obtain mask planning data, wherein the business corresponding to the first business requirement is a computer vision business, the function name and / or parameter name in the mask planning data is masked, and the target model is a model obtained by fine-tuning the initial model based on multiple fine-tuning data; the fine-tuning data includes a second business requirement and true value planning data, the true value planning data includes multiple true value data bodies, the multiple true value data bodies correspond one-to-one to multiple true value business steps, the multiple true value business steps are used to implement the second business requirement, and the function name and / or parameter name in the true value data body is masked according to the set target masking rule;

[0166] The mask unmasking module 402 is used to unmask the function names and parameter names in the mask planning data based on the mask mapping information corresponding to the multiple fine-tuning data to obtain target planning data, and the target planning data is used to be executed by the task planning executor to achieve the first business requirement.

[0167] In one embodiment, the target mask rule includes:

[0168] Replace the first function name in the true value data body including the first function name based on the first mask character corresponding to the first function name, where the first function name is any function name in the first set, and the first set includes: function names included in each of the multiple fine-tuning data; the first mask character is generated based on the function name mask character and the first serial number, and the first serial number is used to uniquely identify the first function name in the first set;

[0169] Replace the first parameter name in the true value data body including the first parameter name based on the second mask character corresponding to the first parameter name, where the first parameter name is any parameter name in the second set, and the second set includes: parameter names included in each of the multiple fine-tuning data; the second mask character is generated based on the parameter name mask character and the second serial number, and the second serial number is used to uniquely identify the first parameter name in the second set.

[0170] In one embodiment, the first mask character is obtained by splicing the function name mask character and the first serial number, and the splicing position of the function name mask character is before the splicing position of the first serial number;

[0171] The second mask character is obtained by splicing the parameter name mask character and the second serial number, and the splicing position of the parameter name mask character is before the splicing position of the second serial number;

[0172] The function name mask character is used to limit the initial model's association of function information, and the parameter name mask character is used to limit the initial model's association of parameter information.

[0173] In one embodiment, the device 400 for generating task planning information based on the computing power of the intelligent computing center further includes a fine-tuning module, and the fine-tuning module is specifically configured to:

[0174] Fine-tune the initial model respectively according to multiple groups of fine-tuning data sets to obtain multiple candidate models, where the multiple candidate models and the multiple groups of fine-tuning data sets correspond one by one, the fine-tuning data set includes multiple candidate fine-tuning data, different groups of fine-tuning data sets correspond to different candidate mask rules, and different candidate mask rules include different parameter name mask characters and / or function name mask characters;

[0175] Use a preset test data set to perform task planning tests on the multiple candidate models respectively to obtain multiple test results, and the test results are used to represent: the ratio of the successfully executed task planning data output by the corresponding candidate model;

[0176] Among the multiple candidate models, determine the candidate model corresponding to the highest ratio as the target model.

[0177] In one embodiment, the fine-tuning data further includes restrictive text, which is used to indicate that the initial model is prohibited from performing semantic parsing on the masked data, and the masked data includes the function names and / or parameter names masked in the true value planning data.

[0178] In one embodiment, the reading position of the restrictive text in the fine-tuning data is the first position, and the reading position of the true value planning data corresponding to the restrictive text in the fine-tuning data is the second position, and the first position is before the second position.

[0179] The device for generating task planning information based on the computing power of the intelligent computing center provided by the present invention can implement each process of the above-mentioned method for generating task planning information based on the computing power of the intelligent computing center. The technical features correspond one by one and can achieve the same technical effects. To avoid repetition, they will not be elaborated here.

[0180] It should be noted that the device for generating task planning information based on the computing power of the intelligent computing center in the present invention can be a device, or a component, integrated circuit, or chip in an electronic device.

[0181] The present invention also provides an electronic device. Refer to Figure 5 , Figure 5 is a schematic structural diagram of an electronic device provided in an embodiment of the present invention. The electronic device includes a memory 501, a processor 502, and a program or instruction running on the memory 501. When the program or instruction is executed by the processor 502, it can implement Figure 1 any step in the corresponding method embodiment for generating task planning information based on the computing power of the intelligent computing center and achieve the same beneficial effects, which will not be elaborated here.

[0182] Among them, the processor 502 can be a CPU, ASIC, FPGA, or GPU.

[0183] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above-mentioned method embodiment for generating task planning information based on the computing power of the intelligent computing center can be completed by hardware related to program instructions, and the program can be stored in a readable medium.

[0184] The present invention also provides a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it can implement the above-mentioned Figure 4Any step in the corresponding method embodiment for generating computing power generation task planning information based on an intelligent computing center, and can achieve the same technical effect. To avoid repetition, it will not be elaborated here. The storage medium includes, for example, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disc, etc.

[0185] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. In addition, the terms "comprising" and "having" and any of their variations are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device comprising a series of steps or units does not necessarily have to be limited to those clearly listed steps or units, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices. In addition, in the present invention, "and / or" is used to represent at least one of the connected objects. For example, A and / or B and / or C represents seven cases including A alone, B alone, C alone, A and B existing together, B and C existing together, A and C existing together, and A, B, and C existing together.

[0186] It should be noted that in this article, the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such a process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or device including that element.

[0187] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disc) and includes several instructions to enable a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a second terminal device, etc.) to execute the methods of the various embodiments of the present invention.

[0188] The embodiments of the present invention are described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit of the present invention and the scope protected by the claims, and all of them fall within the protection scope of the present invention.

Claims

1. A method for generating task planning information of computing power based on an intelligent computing center, characterized in that Including: Step S1: Using the first business requirement to be processed as the task planning objective, perform task planning based on the target model to obtain mask planning data. Among them, the business corresponding to the first business requirement is the computer vision business, the function names and / or parameter names in the mask planning data are masked, and the target model is a model obtained by fine-tuning the initial model based on multiple fine-tuning data. The fine-tuning data includes a second business requirement and true value planning data. The true value planning data includes multiple true value data bodies, and the multiple true value data bodies correspond one-to-one with multiple true value business steps. The multiple true value business steps are used to implement the second business requirement, and the function names and / or parameter names in the true value data bodies are masked according to the set target mask rules; Step S2: Based on the mask mapping information corresponding to the multiple fine-tuning data, unmask the function names and parameter names in the mask planning data to obtain target planning data. The target planning data is used to be executed by the task planning executor to implement the first business requirement.

2. The method according to claim 1, wherein The target mask rules include: Based on the first mask character corresponding to the first function name, replace the first function name in the true value data body including the first function name. Among them, the first function name is any function name in the first set, and the first set includes: the function names included in each of the multiple fine-tuning data. The first mask character is generated based on the function name mask character and the first serial number, and the first serial number is used to uniquely identify the first function name in the first set; Based on the second mask character corresponding to the first parameter name, replace the first parameter name in the true value data body including the first parameter name. Among them, the first parameter name is any parameter name in the second set, and the second set includes: the parameter names included in each of the multiple fine-tuning data. The second mask character is generated based on the parameter name mask character and the second serial number, and the second serial number is used to uniquely identify the first parameter name in the second set.

3. The method according to claim 2, wherein The first mask character is obtained by splicing the function name mask character and the first serial number, and the splicing position of the function name mask character is before the splicing position of the first serial number; The second mask character is obtained by splicing the parameter name mask character and the second serial number, and the splicing position of the parameter name mask character is before the splicing position of the second serial number; The function name mask character is used to limit the initial model's association with function information, and the parameter name mask character is used to limit the initial model's association with parameter information.

4. The method according to claim 2, wherein Before the step S1, the method further includes: Step S01: According to multiple groups of fine-tuning data sets, respectively fine-tune the initial model to obtain multiple candidate models. Among them, the multiple candidate models correspond one-to-one with the multiple groups of fine-tuning data sets. The fine-tuning data set includes multiple candidate fine-tuning data. Different groups of fine-tuning data sets correspond to different candidate mask rules, and different candidate mask rules include different parameter name mask characters and / or function name mask characters; Step S02: using a preset test data set, respectively performing a task planning test on the plurality of candidate models to obtain a plurality of test results, wherein the test results are used to indicate: a ratio of the task planning data output by the corresponding candidate model being successfully executed; Step S03: Determine, among the multiple candidate models, the candidate model corresponding to the highest ratio as the target model.

5. The method according to claim 1, wherein The fine-tuning data also includes restriction text, and the restriction text is used to indicate that the initial model is prohibited from performing semantic parsing on the mask data, and the mask data includes: the masked function names and / or parameter names in the true value planning data.

6. The method according to claim 5, wherein The reading position of the restriction text in the fine-tuning data is the first position, the reading position of the true value planning data corresponding to the restriction text in the fine-tuning data is the second position, and the first position is located before the second position.

7. An apparatus for generating task planning information of computing power based on an intelligent computing center, characterized in that include: A task planning module is used to take a first business requirement to be processed as a task planning target, perform task planning based on a target model, and obtain mask planning data, wherein the business corresponding to the first business requirement is a computer vision business, the function names and / or parameter names in the mask planning data are masked, and the target model is a model obtained by fine-tuning an initial model based on multiple fine-tuning data; the fine-tuning data includes a second business requirement and true value planning data, the true value planning data includes multiple true value data bodies, the multiple true value data bodies correspond one-to-one to multiple true value business steps, the multiple true value business steps are used to implement the second business requirement, and the function names and / or parameter names in the true value data body are masked according to the set target masking rules; The mask unmasking module is used to unmask the function names and parameter names in the mask planning data based on the mask mapping information corresponding to the multiple fine-tuning data to obtain target planning data, and the target planning data is used to be executed by the task planning executor to achieve the first business requirement.

8. A server, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the steps of the method for generating task planning information based on the computing power of an intelligent computing center as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for generating task planning information based on the computing power of an intelligent computing center as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes computer instructions, which, when executed by a processor, implement the steps of the method for generating task planning information based on the computing power of an intelligent computing center as described in any one of claims 1 to 6.

Images