An intelligent empowerment AI practical training full-process integrated management method and system

Abstract

This invention discloses an integrated management method and system for the entire AI training process empowered by intelligent computing, relating to the field of artificial intelligence technology. It establishes a three-level mapping relationship between projects, groups, and members by having teachers input project information, configure grouping, and assign unique identifiers. This, combined with standardized task allocation, forms the underlying training data. Multi-dimensional training and computing power data are collected and standardized. Based on the total group quota, individual computing power is dynamically allocated using multi-dimensional indicators and task weights. Simultaneously, computing power monitoring, anomaly restrictions, and idle resource recovery are implemented to ensure fair and efficient use of computing resources. Based on standardized training data, individual contributions are quantified and individual scores are broken down proportionally, automatically generating assessment and review reports. This provides comprehensive and traceable support for training assessment and teaching review, and finally, the entire process data is archived, achieving a closed-loop iterative teaching process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an integrated management method and system for the entire AI training process empowered by intelligent computing. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, practical training in AI has become increasingly widespread, serving as a crucial link in cultivating AI professionals. In AI practical training, tasks such as model training, data processing, and algorithm development based on intelligent computing platforms constitute the core content. Collaborative group project training, as a teaching organization form closely aligned with engineering practice, is widely used in university AI practical training because it simulates real project development environments and cultivates students' teamwork and engineering practice abilities. In collaborative group project training scenarios, how to rationally allocate intelligent computing power to ensure the smooth progress of the training, while objectively quantifying and calculating the individual contributions of group members to provide a scientific basis for training assessment, has become a critical issue that urgently needs to be addressed in the current management of AI practical training in universities.

[0003] Existing intelligent computing management platforms are primarily designed for individual, independent training scenarios, which have numerous limitations and cannot meet the teaching and management needs of collaborative AI training in groups. Regarding computing power allocation, traditional solutions only support fixed individual computing power quota applications, static allocation, and independent use, failing to adapt to the collaborative training architecture of groups. They lack a mechanism for coordinating the total computing power quota at the group level and for internal secondary dynamic allocation. This leads to fairness issues during actual training, such as a single member excessively consuming computing power while other members have insufficient power, severely impacting group collaboration efficiency and training effectiveness. In terms of quantifying individual contributions, traditional solutions lack a quantitative accounting system specific to collaborative group scenarios. Individual contributions largely rely on subjective teacher evaluations, lacking objective evidence. Furthermore, training process data is fragmented and untraceable, failing to provide strong support for group performance breakdown and process-based assessment. In addition, computing power allocation is disconnected from training assessment and teaching management, failing to form a fully integrated control system, resulting in low training management efficiency and making it difficult to meet the teaching and management needs of project-based, group-based AI training in universities. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an integrated management method and system for the entire AI training process empowered by intelligent computing. This system establishes a three-tiered mapping relationship between projects, groups, and members by having teachers input project information, configure grouping, and assign unique identifiers. Combined with standardized task allocation, this forms the underlying training data. Multi-dimensional training and computing power data are collected and standardized. Based on the total group quota, individual computing power is dynamically allocated according to multi-dimensional indicators and task weights. Simultaneously, computing power monitoring, anomaly restrictions, and idle resource recovery are implemented to ensure the fair and efficient use of computing resources. Based on standardized training data, individual contributions are quantified and individual scores are broken down proportionally, automatically generating assessment and review reports. This provides comprehensive and traceable support for training assessment and teaching review. Finally, the entire process data is archived, achieving a closed-loop iterative teaching process. This allows teachers to adjust training parameters and teaching plans based on data analysis results, continuously improving the quality of AI training in universities.

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies. On the one hand, it provides an integrated management method and system for the entire AI training process empowered by intelligent computing. This system includes: S1: Teachers enter basic information about the training project through the platform and complete the group formation, constructing a three-level mapping relationship between project, group, and member. The group completes the internal task breakdown and standardized entry, and the task is stored after being reviewed by the teacher. S2: The system connects to the underlying computing power nodes, code repository, model training module and data processing module of intelligent computing, collects training behavior and resource usage data in real time in terms of tasks, code, models, data and computing power, and cleans, deduplicates and standardizes the raw data to form training datasets; S3: Apply to the Intelligent Computing Center for and lock the total computing power quota of the group based on the parameters of the training project. Combine the standardized multi-dimensional data and task division weight to calculate the computing power allocation coefficient of the members and split the individual dynamic computing power quota according to the proportion. At the same time, monitor the computing power usage status in real time, reduce the quota of members who use computing power in excess and inefficiently, reclaim long-term idle computing power and redistribute it to members in need within the group. S4: Based on standardized data collected throughout the entire cycle, a comprehensive accounting model for individual contributions is constructed. Multi-dimensional assessment indicators are scored in a standardized manner and weighted and summed according to the teacher's preset weights to obtain the individual contribution score. The group's overall score is split into individual training scores according to the proportion of individual contributions in the group's total contribution, and individual process assessment reports and group training review reports are automatically generated. S5: Unifies and archives all process data to form an immutable lifecycle ledger, supports teachers to query, export, and perform statistical analysis from multiple dimensions, and adjusts training weights, computing power thresholds, and teaching rules based on the analysis results, forming a closed-loop management from training execution to teaching optimization.

[0006] Furthermore, in S1, teachers input the project type, implementation cycle, total computing power requirement threshold, assessment weight rules, and project acceptance standards of the corresponding AI training project through the platform management terminal. The platform performs legality verification and structured storage on the input project information. Teachers select designated grouping or open student self-grouping to complete the division of training groups according to the teaching arrangement. After the grouping is confirmed, the platform assigns a globally unique project identifier to each training project, assigns an independent group number to each completed group, and binds a unique identity code to each student participating in the training. By linking and binding the project identifier, group number, and student identity code level by level, a three-level one-to-one correspondence mapping model of project-group-member is established.

[0007] Furthermore, in S2, the task dimension includes subtask completion progress, task node submission timeliness, and task quality verification results; the code dimension includes the number of valid code submissions, the number of valid lines of code, the code merging pass rate, and version iteration records; the model dimension includes the number of model training initiations, the number of valid iterations, the improvement in accuracy of iterated models, and hyperparameter tuning records; the data dimension includes the dataset processing scale, the amount of data cleaning completed, the amount of data labeled, and the efficiency of data preprocessing; and the computing power dimension includes the amount of computing power requested by a single member, the actual amount used, the duration of computing power usage, the output per unit of computing power, and the computing power idle rate.

[0008] Furthermore, in S3, the collected multi-dimensional training data is uniformly normalized and organized. Then, combined with the task assignment weights of the members confirmed by the teacher, the computing power allocation coefficient of each member is obtained through the computing power allocation coefficient calculation formula. Subsequently, the total computing power quota of the group is applied for and locked from the intelligent computing center. The individual dynamic computing power quota is accurately divided according to the proportion of each member's computing power allocation coefficient in the total coefficient of all members in the group through the individual dynamic computing power quota splitting calculation formula. The split quota is synchronized to each member's personal account in real time. Moreover, the coefficient will be recalculated and the individual quota will be adjusted in conjunction with the update of the members' training behavior data to ensure that the computing power allocation is accurately matched with the actual training needs of the members.

[0009] Furthermore, in S3, the computing power allocation coefficient for each member is obtained through the computing power allocation coefficient calculation formula, which is as follows: ,in, It is the first The computing power allocation coefficient of each member is used to characterize the proportion of computing power that should be allocated to each member. The larger the coefficient, the more computing power can be allocated. The task allocation weights are customized by the teacher according to the needs of the practical training project. It is the first The task completion index value for each member is a standardized comprehensive value of subtask completion progress, submission timeliness, and quality verification results. It is the effective code contribution weight, a teacher-defined configuration that represents the significant proportion of code submission dimensions in computing power allocation; It is the first The effective code commit metric for each member is a standardized composite value of the number of effective code commits, lines of code, and merge pass rate. It is the model iteration weight, which is a teacher-defined configuration and represents the important proportion of the model training iteration dimension in the computing power allocation; It is the first The model iteration index value of each member is a standardized comprehensive value of the effective number of model iterations, the improvement in accuracy, and the effect of hyperparameter tuning; It is the data processing weight, which is a teacher-defined configuration that represents the important proportion of data processing dimensions in the allocation of computing power; It is the first The data processing index value for each member is a standardized comprehensive value of the dataset processing scale, the amount of cleaning completed, the amount of annotation, and the preprocessing efficiency. It is a weight for computing power utilization efficiency, which is a teacher-defined configuration that represents the important proportion of the rationality of computing power utilization in computing power allocation; It is the first The computing power utilization efficiency index value of each member is a standardized comprehensive value of unit computing power output and computing power idle rate.

[0010] Furthermore, in S3, the calculation formula for splitting individual dynamic computing power quotas is based on the proportion of each member's computing power allocation coefficient in the total coefficients of all members in the group. The formula is as follows: ,in, It is the first The individual dynamic computing power quota of each member, that is, the amount of computing power resources ultimately allocated to that member; It is the group's total computing power quota. It is the total number of group members; It is the sum of the computing power allocation coefficients of all members in the group.

[0011] Furthermore, in S4, based on the training data collected and standardized throughout the entire cycle, a comprehensive accounting model for individual contributions is built. Standardized scores are applied to multi-dimensional assessment indicators, including task division matching degree, task completion quality, code contribution, model iteration contribution, data processing contribution, computing power utilization efficiency, and team collaboration degree. The raw data of each indicator is converted into compliant scores with unified dimensions. Then, according to the assessment weights of each dimension pre-configured and confirmed by the teacher, all standardized scores are weighted and summed to obtain the individual contribution score of each member.

[0012] Furthermore, in S4, the overall group evaluation score assessed by the teacher is used as the calculation base to calculate the proportion of each member's individual contribution score in the total contribution score of all members in the group. Based on this proportion, the overall group evaluation score is reasonably divided into individual training scores for each member. At the same time, the entire training process data is automatically integrated to generate a targeted individual process assessment report. The report also summarizes the overall progress of the group's training, the distribution of member contributions, the effectiveness of computing power scheduling, and the highlights and shortcomings in the project implementation, forming a complete group training review report.

[0013] On the other hand, a method and system for integrated management of the entire AI training process empowered by intelligent computing, the system comprising: Practical Training Basics and Group Management Module: Responsible for creating practical training projects, grouping groups, establishing three-level mapping relationships, and inputting and approving task assignments; Multi-dimensional data acquisition and processing module: used to connect with intelligent computing power nodes and training tools to complete the real-time acquisition, cleaning, standardization processing and storage of data in five dimensions; The computing power coordination and dynamic scheduling module enables the group's total computing power application and locking, individual computing power coefficient calculation, quota dynamic splitting and adjustment, as well as computing power anomaly monitoring, early warning and idle recovery; Contribution Calculation and Performance Breakdown Module: Equipped with a multi-dimensional calculation model to complete the quantitative scoring of contributions, break down individual performance according to the contribution percentage and generate an assessment review report; Data archiving and teaching iteration module: Provides full-process data archiving, retrieval, export and analysis functions to support the optimization and iteration of training parameters and teaching plans.

[0014] Compared with existing technologies, this intelligent computing-enabled AI training process integrated management method and system has the following beneficial effects: This invention establishes a standardized three-level mapping system for projects, groups, and members, ensuring the uniqueness and accuracy of data associations and addressing the fragmentation and chaotic data association issues inherent in traditional management processes. It employs a group-wide quota management and internal dynamic scheduling model, adjusting quotas in real-time based on member training behavior and computing power utilization efficiency. This solves the problem of malicious computing power hogging, reclaims idle computing power for resource reuse, significantly improves the efficiency of intelligent computing resource utilization, and precisely adapts to the computing power allocation needs of collaborative group training. Based on multi-dimensional real training data, it quantifies individual contributions and automatically breaks down grades, generating standardized assessment and review reports to replace traditional subjective manual evaluation. All calculation processes and original data are synchronously archived to form an immutable ledger, addressing the industry pain point of lacking quantifiable individual contributions and objective assessment criteria in group training. The weight parameters and control thresholds for computing power allocation and contribution calculation can be customized by teachers. Furthermore, through full-cycle data archiving and multi-dimensional analysis, the teaching plan is continuously optimized and iterated, contributing to the steady improvement of AI training quality in universities.

[0015] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0017] Figure 1 This is a flowchart of an integrated management method for the entire AI training process empowered by intelligent computing; Figure 2 A structural diagram of an integrated management system for the entire AI training process empowered by intelligent computing. Figure 3 This is a flowchart of step S3 of an integrated management method for the entire AI training process empowered by intelligent computing. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] Teachers fill in the core information of the AI ​​training project through the system's application layer management interface. This includes the project type, implementation period, total computing power requirement threshold, assessment weight rules, and project acceptance criteria. The system has a built-in legality verification engine that verifies the entered information in real time: the project type must be selected from a preset category list or customized and then verified for compliance using system keywords; the implementation period must meet logical rationality requirements; the total computing power requirement threshold must be within the computing power range available from the intelligent computing center; and the assessment weight rules must have reserved interfaces for configuring weights in each dimension to ensure flexible settings later. After verification, the system uses a structured storage scheme to store the project information in a distributed database according to a hierarchical structure of project basic attributes - assessment rules - acceptance criteria, improving data retrieval efficiency.

[0020] Teachers can choose from two grouping modes: In designated grouping mode, teachers select students from the interface and assign them to different groups. Each group can have an upper and lower limit on the number of students, allowing for customized configuration. In self-grouping mode, students initiate group invitations through the student interface. Invitations must be responded to within a specified time. After grouping, the system submits the invitations for review, allowing teachers to view the composition of each group and make appropriate adjustments. After grouping confirmation, the system automatically generates a three-level unique identifier: a globally unique project identifier generated using a universal unique identification algorithm and named according to preset format rules; a unique group number generated for each group, associated with the project identifier's feature fields; and a unique identity code bound to each participating student, associated with their personal feature information. This hierarchical binding of the project identifier, group number, and student identity code constructs a one-to-one correspondence mapping model between project, group, and member, stored in a relational database. This allows for quick retrieval of related hierarchical information using any level of identifier.

[0021] Once the groups are formed, the group leader initiates internal task breakdown through the student interface. The system provides a task breakdown template, which includes core fields such as subtask name, task description, responsible members, completion deadline, and dependent resources. After group members negotiate and determine the allocation plan for each subtask, they fill in the template and submit it to the system, forming a group task allocation list. Teachers view and review each group's task allocation list through the management interface, focusing on verifying the rationality of task allocation, the matching of completion deadlines with the project implementation cycle, and the feasibility of dependent resources. If the review is approved, the system stores the task allocation list in association with the three-level mapping relationship. If the review is not approved, the teacher fills in the modification comments and returns it to the group. The group adjusts according to the comments and resubmits for review until it is approved.

[0022] The system integrates with the underlying computing power nodes, code repositories, model training modules, and data processing modules through standardized interfaces: It uses mainstream API interface protocols with computing power nodes to obtain real-time data on computing power usage; it establishes a connection with the code repository of version control tools through a triggered data collection mechanism, automatically pushing relevant data to the system when code commits, merges, or other operations occur; it collects iterative data during model training through built-in callback functions; and it obtains operation logs and result data during data processing through log collection interfaces with commonly used data processing tools.

[0023] Real-time collection of training activities and resource usage data across tasks, code, models, data, and computing power dimensions: For task-level data: The status tracking function of the task management module is used to collect the progress of subtask completion, the timeliness of task node submission, and the results of task quality verification in real time. For code-related data: collect the number of valid code commits, the number of valid lines of code, the code merge pass rate, and the version iteration history through the code repository interface; For model-level data: collect the number of model training initiations, the number of effective iterations, the improvement in model accuracy during iterations, and hyperparameter tuning records through the model training module callback function; For data dimensions: collect dataset processing scale, data cleaning completion amount, data annotation amount, and data preprocessing completion efficiency through data processing module logs; For computing power data: Collect data on the number of computing power requests per member, actual usage, computing power usage time, output per unit of computing power, and computing power idle rate through the computing power node interface.

[0024] Missing data is handled through a pre-set imputation algorithm; then, outlier detection algorithms are used to identify and mark outlier data, which is then manually confirmed by teachers. After confirmation, data can be deleted or corrected. Finally, data from different sources is standardized and converted into a pre-set format. A hash-based deduplication mechanism is used to generate a unique hash value for each piece of collected data. Duplicate data is identified by comparing hash values, and valid data is retained according to pre-set rules. For near-duplicate data with slight differences, the latest data is retained based on timestamp priority. Standardization methods are used to convert the original data of each dimension into standard scores with a unified dimension, ensuring the comparability of data in different dimensions. The standardized multi-dimensional data is integrated to form a training dataset, which is stored in a distributed data warehouse and supports retrieval by project, group, member, time, and other dimensions.

[0025] like Figure 3 As shown, based on the total computing power requirement threshold of the entered training project, a computing power application request is initiated to the intelligent computing center. The request information includes core content such as project identifier, number of groups, total computing power requirement threshold, and training period. After receiving the request, the intelligent computing center allocates quotas according to the current remaining computing power resources. After the allocation is completed, a confirmation message is returned to the system. After receiving the confirmation message, the system locks the total computing power quota of the group to ensure that the quota is used exclusively during the training period. The locked status is synchronized to the teacher management terminal and the student terminal in real time, so that users can view the computing power quota usage in real time.

[0026] By obtaining the weight configurations confirmed by the teachers, including weights for member task allocation, effective code contribution, model iteration, data processing, and computing power utilization efficiency, teachers configure these weight values ​​through the management interface. The system automatically verifies the compliance of the weight configurations; if the verification fails, the teacher is prompted to adjust them until compliance is achieved. Standardized indicator values ​​for five dimensions are extracted from the training dataset for each member, and each dimension indicator value is calculated by weighting according to the preset sub-indicator weight ratios. Task completion index value: The standardized values ​​of sub-indicators such as sub-task completion progress, submission timeliness, and quality verification results are calculated according to preset weight ratios; Effective code submission metric: The standardized values ​​of sub-metrics such as the number of effective code submissions, the number of effective lines of code, and the code merge pass rate are calculated according to preset weight ratios. Model iteration index value: The standardized value of sub-indicators such as the number of effective iterations, the improvement in accuracy, and the effect of hyperparameter tuning is calculated according to the preset weight ratio; Data processing metric values: These are standardized values ​​of sub-indicators such as the scale of the comprehensive dataset processing, the amount of data cleaning completed, the amount of data labeled, and the efficiency of preprocessing, calculated according to preset weight ratios. The computing power utilization efficiency index is calculated by combining the standardized values ​​of sub-indicators such as unit computing power output and computing power idle rate (the reverse index is converted according to preset rules) and the preset weight ratio.

[0027] Substitute into the calculation formula for computing power allocation coefficient The computing power allocation coefficient for each member is calculated. This coefficient is used to characterize the proportion of computing power that should be allocated to each member. The system calculates this based on the individual dynamic computing power quota allocation formula. Calculate the individual dynamic computing power quota for each member. ,in, It is the group's total computing power quota. It is the total number of group members; It is the sum of the computing power allocation coefficients of all members in the group. After the calculation is completed, the system will synchronize the individual dynamic computing power quota to the personal account of each member in real time. Members can view their own quota and usage through the student terminal interface.

[0028] The system is equipped with a dynamic adjustment mechanism for computing power quotas. The adjustment period can be customized by the teacher. During the adjustment period, the system collects the training behavior data of the members in real time and recalculates the computing power allocation coefficient for each member. The system will then redistribute individual dynamic computing power quotas according to the formula described above. If a member’s training behavior changes significantly, the system can trigger an immediate adjustment to ensure that the computing power allocation is accurately matched with the member’s actual training needs.

[0029] The system continuously tracks the computing power usage status of each member through a real-time monitoring engine: preset computing power over-limit threshold and inefficient usage judgment criteria. When a member meets both the conditions of over-using computing power and inefficient usage, the system automatically reduces their quota according to a preset ratio; preset computing power idle judgment criteria. For members who meet the conditions of long-term idleness, the system automatically reclaims the idle computing power and redistributes it to members in the group with tight computing power needs.

[0030] The system pushes real-time warnings of abnormal computing power usage to teachers (such as excessive or inefficient use, long-term idleness, insufficient computing power, etc.). Teachers can view detailed abnormal data and manually intervene in computing power scheduling according to the actual situation.

[0031] The system builds a comprehensive calculation model for individual contributions based on machine learning algorithms. The model input consists of standardized multi-dimensional assessment indicators, including task allocation matching, task completion quality, code contribution, model iteration contribution, data processing contribution, computing power utilization efficiency, and teamwork. The raw data for each indicator undergoes secondary standardization to ensure complete uniformity of measurement dimensions, converting it into a compliant score. Then, according to the pre-configured and confirmed assessment weights for each dimension by the teacher, all standardized scores are weighted and summed to obtain each member's individual contribution score. The formula is as follows: ,in For the first The individual contribution score of each member , , , , , , Assess the weighting of each dimension in the evaluation. , , , , , , Standardized scores for each dimension.

[0032] Teachers evaluate the group's overall score through the management interface. The scoring criteria include project completion, overall group performance, and achievement of training objectives. This overall score serves as the base for calculating the percentage of each member's individual contribution score within the total contribution score of all group members. ,in For the first The percentage of each member's contribution. The total contribution score of all group members is used as the basis for calculating the group's overall score. This score is then used to further divide the group's final grade into individual training scores for each member, using the following formula: ,in For the first Individual training scores of each member The system automatically synchronizes the group's overall score to the students' devices after the scores are broken down, allowing students to view their individual scores. Teachers can also manually adjust the scores.

[0033] The system uses template engine technology to automatically integrate training data from the entire process and generate targeted individual process assessment reports. The report content includes basic information of members, detailed scores of assessment indicators in each dimension, individual contribution scores and related ranking information, statistics on computing power usage, task completion details, existing problems and improvement suggestions. At the same time, the system summarizes the overall progress of the group's training, the distribution of members' contributions, the effectiveness of computing power scheduling, and the highlights and shortcomings in project implementation, forming a complete group training review report. The report supports exporting to various commonly used formats, and teachers can use the report for teaching comments and subsequent training optimization references.

[0034] The system uses blockchain technology to ensure the immutability of data throughout the entire process. All data from the creation to the completion of the training project (including basic project information, group composition information, task division data, multi-dimensional training datasets, computing power scheduling records, contribution scores, individual and group scores, assessment review reports, etc.) are generated into blocks in time stamp order. Each block contains the hash value of the previous block, the hash value of the current block, and the data content. An immutable lifecycle ledger is formed through chain storage. Archived data is stored simultaneously in a distributed data warehouse and blockchain nodes, supporting multiple replica backups to ensure data security and recoverability. After data archiving is completed, the system generates a unique archive identifier for each project, linking the project identifier with the lifecycle ledger for easy subsequent retrieval.

[0035] The system provides teachers with multi-dimensional data query functions, supporting queries by combination of filtering conditions such as project name, implementation period, group number, student identity code, and data type. The query results are displayed visually in the form of lists, charts, etc. The system also supports data export, allowing teachers to select some or all of the data in the query results and export them to various commonly used formats for offline analysis or retention in teaching archives.

[0036] The system incorporates a big data analytics framework to perform in-depth statistical analysis on archived data, including horizontal comparative analysis (between different projects, groups, and members), vertical comparative analysis (the same member at different training stages), and correlation analysis (computing power allocation and utilization efficiency, assessment indicators and training results, etc.). The analysis results automatically generate analysis reports and push them to the teacher management terminal.

[0037] Based on the statistical analysis results, teachers adjust the relevant parameters and teaching rules for practical training: for example, adjusting the weight of computing power allocation in each dimension based on the analysis of computing power utilization efficiency; adjusting the assessment weight of each dimension of individual contribution calculation based on the correlation analysis of assessment indicators; adjusting the total computing power requirement threshold of the project based on the analysis of computing power threshold adaptability; and optimizing the design of practical training tasks and teaching guidance schemes based on the analysis of students' performance shortcomings.

[0038] The system stores the adjusted parameters and rules as a new training configuration template. When creating new training projects in the future, teachers can directly select this template or make minor adjustments based on it to achieve closed-loop management and continuously improve the quality and effectiveness of AI training.

[0039] like Figure 2 As shown, the AI ​​training end-to-end integrated management system provided by this invention includes a training foundation and group management module, a multi-dimensional data acquisition and processing module, a computing power coordination and dynamic scheduling module, a contribution calculation and grade splitting module, and a data archiving and teaching iteration module. The core modules collaborate with each other through standardized data interfaces and message queues. The Practical Training Foundation and Group Management module synchronizes project information, grouping information, and task assignment data to the multi-dimensional data acquisition and processing module and the computing power coordination and dynamic scheduling module. The multi-dimensional data acquisition and processing module pushes the standardized training dataset to the computing power coordination and dynamic scheduling module (for computing power allocation coefficient calculation), the contribution accounting and grade splitting module (for individual contribution accounting), and the data archiving and teaching iteration module (for data archiving and analysis). The computing power coordination and dynamic scheduling module synchronizes the computing power quota allocation and adjustment records to the data archiving and teaching iteration module, and at the same time feeds back computing power usage status data to the multi-dimensional data acquisition and processing module; The contribution calculation and performance breakdown module synchronizes individual contribution scores, individual training results, and assessment review reports to the data archiving and teaching iteration module for archiving and analysis. The data archiving and teaching iteration module feeds back the optimized training parameters and teaching rules to the training foundation and group management module for the configuration of the new round of training projects.

[0040] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for integrated management of the entire AI training process empowered by intelligent computing, characterized in that, The method includes: S1: Teachers enter basic information about the training project through the platform and complete the group formation, constructing a three-level mapping relationship between project, group, and member. The group completes the internal task breakdown and standardized entry, and the task is stored after being reviewed by the teacher. S2: The system connects to the underlying computing power nodes, code repository, model training module and data processing module of intelligent computing, collects training behavior and resource usage data in real time in terms of tasks, code, models, data and computing power, and cleans, deduplicates and standardizes the raw data to form training datasets; S3: Apply to the Intelligent Computing Center for and lock the total computing power quota of the group based on the parameters of the training project. Combine the standardized multi-dimensional data and task division weight to calculate the computing power allocation coefficient of the members and split the individual dynamic computing power quota according to the proportion. At the same time, monitor the computing power usage status in real time, reduce the quota of members who use computing power in excess and inefficiently, reclaim long-term idle computing power and redistribute it to members in need within the group. S4: Based on standardized data collected throughout the entire cycle, a comprehensive accounting model for individual contributions is constructed. Multi-dimensional assessment indicators are scored in a standardized manner and weighted and summed according to the teacher's preset weights to obtain the individual contribution score. The group's overall score is split into individual training scores according to the proportion of individual contributions in the group's total contribution, and individual process assessment reports and group training review reports are automatically generated. S5: Unifies and archives all process data to form an immutable lifecycle ledger, supports teachers to query, export, and perform statistical analysis from multiple dimensions, and adjusts training weights, computing power thresholds, and teaching rules based on the analysis results, forming a closed-loop management from training execution to teaching optimization.

2. The integrated management method for the entire AI training process empowered by intelligent computing as described in claim 1, characterized in that, In S1, teachers input the project type, implementation cycle, total computing power requirement threshold, assessment weight rules, and project acceptance standards of the corresponding AI training project through the platform management terminal. The platform performs legality verification and structured storage on the input project information. Teachers select designated grouping or open student self-grouping to complete the division of training groups according to the teaching arrangement. After the grouping is confirmed, the platform assigns a globally unique project identifier to each training project, assigns an independent group number to each completed group, and binds a unique identity code to each student participating in the training. By linking and binding the project identifier, group number, and student identity code level by level, a three-level one-to-one correspondence mapping model of project-group-member is established.

3. The integrated management method for the entire AI training process empowered by intelligent computing as described in claim 1, characterized in that, In S2, the task dimension includes subtask completion progress, task node submission timeliness, and task quality verification results; the code dimension includes the number of valid code submissions, the number of valid lines of code, the code merging pass rate, and version iteration records; the model dimension includes the number of model training initiations, the number of valid iterations, the improvement in accuracy of iterated models, and hyperparameter tuning records; the data dimension includes the dataset processing scale, the amount of data cleaning completed, the amount of data annotation, and the efficiency of data preprocessing; and the computing power dimension includes the amount of computing power requested by a single member, the actual amount used, the duration of computing power usage, the output per unit of computing power, and the computing power idle rate.

4. The integrated management method for the entire AI training process empowered by intelligent computing as described in claim 1, characterized in that, In step S3, the collected multi-dimensional training data is uniformly normalized and organized. Then, combined with the task assignment weights of the members confirmed by the teacher, the computing power allocation coefficient of each member is obtained through the computing power allocation coefficient calculation formula. Subsequently, the total computing power quota of the group is applied for and locked from the intelligent computing center. The individual dynamic computing power quota is accurately divided according to the proportion of each member's computing power allocation coefficient in the total coefficient of all members in the group through the individual dynamic computing power quota splitting calculation formula. The split quota is synchronized to each member's personal account in real time. Furthermore, the coefficient will be recalculated and the individual quota will be adjusted in conjunction with the update of the members' training behavior data to ensure that the computing power allocation is accurately matched with the actual training needs of the members.

5. The integrated management method for the entire AI training process empowered by intelligent computing as described in claim 4, characterized in that, In S3, the computing power allocation coefficient for each member is obtained through the computing power allocation coefficient calculation formula, which is as follows: ,in, It is the first The computing power allocation coefficient of each member is used to characterize the proportion of computing power that should be allocated to each member. The larger the coefficient, the more computing power can be allocated. The task allocation weights are customized by the teacher according to the needs of the practical training project. It is the first The task completion index value for each member is a standardized comprehensive value of subtask completion progress, submission timeliness, and quality verification results. It is the effective code contribution weight, a teacher-defined configuration that represents the significant proportion of code submission dimensions in computing power allocation; It is the first The effective code commit metric for each member is a standardized composite value of the number of effective code commits, lines of code, and merge pass rate. It is the model iteration weight, which is a teacher-defined configuration and represents the important proportion of the model training iteration dimension in the computing power allocation; It is the first The model iteration index value of each member is a standardized comprehensive value of the effective number of model iterations, the improvement in accuracy, and the effect of hyperparameter tuning; It is the data processing weight, which is a teacher-defined configuration that represents the important proportion of data processing dimensions in the allocation of computing power; It is the first The data processing index value for each member is a standardized comprehensive value of the dataset processing scale, the amount of cleaning completed, the amount of annotation, and the preprocessing efficiency. It is a weight for computing power utilization efficiency, which is a teacher-defined configuration that represents the important proportion of the rationality of computing power utilization in computing power allocation; It is the first The computing power utilization efficiency index value of each member is a standardized comprehensive value of unit computing power output and computing power idle rate.

6. The integrated management method for the entire AI training process empowered by intelligent computing as described in claim 5, characterized in that, In S3, the calculation formula for splitting individual dynamic computing power quotas is based on the proportion of each member's computing power allocation coefficient in the total coefficients of all members in the group. The formula is as follows: ,in, It is the first The individual dynamic computing power quota of each member, that is, the amount of computing power resources ultimately allocated to that member; It is the group's total computing power quota. It is the total number of group members; It is the sum of the computing power allocation coefficients of all members in the group.

7. The integrated management method for the entire AI training process empowered by intelligent computing as described in claim 1, characterized in that, In S4, based on the training data collected and standardized throughout the entire cycle, a comprehensive accounting model for individual contributions is built. Standardized scores are applied to multi-dimensional assessment indicators, including task division matching degree, task completion quality, code contribution, model iteration contribution, data processing contribution, computing power utilization efficiency, and team collaboration degree. The raw data of each indicator is converted into compliant scores with unified dimensions. Then, according to the assessment weights of each dimension pre-configured and confirmed by the teacher, all standardized scores are weighted and summed to obtain the individual contribution score of each member.

8. The integrated management method for the entire AI training process empowered by intelligent computing as described in claim 1, characterized in that, In S4, the overall group evaluation score assessed by the teacher is used as the calculation base. The proportion of each member's individual contribution score in the total contribution score of all members in the group is calculated. The overall group evaluation score is then reasonably divided into individual training scores for each member according to this proportion. At the same time, the entire training process data is automatically integrated to generate a targeted individual process assessment report. The report also summarizes the overall progress of the group's training, the distribution of member contributions, the effectiveness of computing power scheduling, and the highlights and shortcomings in the project implementation, forming a complete group training review report.

9. A computing-enabled AI training end-to-end integrated management system, the method being applicable to the computing-enabled AI training end-to-end integrated management method described in any one of claims 1-8, characterized in that, The system includes: Practical Training Basics and Group Management Module: Responsible for creating practical training projects, grouping groups, establishing three-level mapping relationships, and inputting and approving task assignments; Multi-dimensional data acquisition and processing module: used to connect with intelligent computing power nodes and training tools to complete the real-time acquisition, cleaning, standardization processing and storage of data in five dimensions; The computing power coordination and dynamic scheduling module enables the group's total computing power application and locking, individual computing power coefficient calculation, quota dynamic splitting and adjustment, as well as computing power anomaly monitoring, early warning and idle recovery; Contribution Calculation and Performance Breakdown Module: Equipped with a multi-dimensional calculation model to complete the quantitative scoring of contributions, break down individual performance according to the contribution percentage and generate an assessment review report; Data archiving and teaching iteration module: Provides full-process data archiving, retrieval, export and analysis functions to support the optimization and iteration of training parameters and teaching plans.

Images