Task execution method and device based on synthetic data quality control, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, and provides a task execution method and device based on synthetic data quality control, equipment and a medium, which can perform data extraction on initial synthetic data based on an answer extraction rule to reduce data noise. Calling a basic model to carry out single reasoning evaluation on the initial data set, and carrying out difficulty self-adaptive rewriting on the initial data set according to a reasoning result so as to improve the proportion of learnable samples; the basic model is called to perform multiple reasoning evaluation on the rewriting sample set, and the rewriting samples are graded and exported according to the reasoning result, so that the output data are more concentrated in a learnable interval, the training effectiveness is improved, and the waste of training resources is reduced; performing post-model training on the pre-training model by using the training set, and calling the baseline model and the preset evaluation set to evaluate the task execution model, so as to form end-to-end closed-loop iteration; and the task is effectively executed in combination with the high-quality synthetic data, so that the task execution effect is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a task execution method, apparatus, device, and medium based on synthetic data quality control. Background Technology

[0002] Currently, due to factors such as data acquisition costs, copyright and privacy compliance restrictions, insufficient data coverage, and delayed updates, it is difficult to meet the data scale and diversity requirements of model iteration in many industry scenarios.

[0003] To address the aforementioned issues, synthetic data technology has been widely researched and applied. However, the following problems still exist when performing tasks based on synthetic data: (1) The quality control and screening of synthetic data mostly remain at the level of rules or sampling inspections, lacking an evaluation-driven closed-loop control mechanism.

[0004] Current practices often use rule filtering, deduplication, or a small amount of manual sampling to control quality, but these methods struggle to provide a quantitative characterization of difficulty distribution, learnable range, and error types. Furthermore, the proportion of overly easy, overly difficult, or unlearnable samples in synthetic data is too high, resulting in low utilization of training resources.

[0005] (2) The links of synthetic data, model training and post-training evaluation are disconnected, and the versions and parameters are not traceable, making it difficult to form an end-to-end closed-loop iteration.

[0006] In existing engineering processes, data synthesis, training task execution, and post-training evaluation are often completed by different scripts or systems, lacking a unified task orchestration and metadata recording mechanism. This makes it difficult to establish a stable correlation between data versions, generation parameters, training configurations, and evaluation results. When model performance changes, it is difficult to quickly locate the cause and to automatically drive the iterative optimization of the next round of data synthesis and training strategies. Summary of the Invention

[0007] In view of the above, it is necessary to provide a task execution method, apparatus, equipment and medium based on synthetic data quality control, which can improve the task execution effect by controlling the quality of synthetic data.

[0008] A task execution method based on synthetic data quality control, the task execution method based on synthetic data quality control includes: In response to a task execution instruction based on initial synthesized data, data is extracted from the initial synthesized data based on answer extraction rules to obtain an initial dataset; The base model is invoked to perform a single inference evaluation on the initial dataset to obtain a first inference result, and the initial dataset is rewritten in an adaptive difficulty manner based on the first inference result to obtain a rewritten sample set; The base model is invoked to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and the rewritten samples are then graded and derived based on the second inference result to obtain a training set; The pre-trained model is post-trained using the training set to obtain the task execution model. The task execution model is evaluated using a baseline model and a preset evaluation set to obtain evaluation results; When the evaluation result indicates that the task execution model meets the standard, the task execution model is used to respond to the task execution instruction.

[0009] A task execution device based on synthetic data quality control, the task execution device based on synthetic data quality control includes: An extraction unit is configured to, in response to a task execution instruction based on initial synthetic data, extract data from the initial synthetic data based on answer extraction rules to obtain an initial dataset. The rewriting unit is used to call the base model to perform a single inference evaluation on the initial dataset to obtain a first inference result, and to perform difficulty adaptive rewriting on the initial dataset based on the first inference result to obtain a rewritten sample set. The export unit is used to call the base model to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and to export the rewritten samples in a hierarchical manner based on the second inference result to obtain a training set. The post-training unit is used to perform post-training on the pre-trained model using the training set to obtain the task execution model. The evaluation unit is used to call the baseline model and the preset evaluation set to evaluate the task execution model and obtain the evaluation results; A response unit is used to respond to the task execution instruction using the task execution model when the evaluation result indicates that the task execution model meets the standard.

[0010] A computer device, the computer device comprising: A memory for storing at least one instruction; and a processor for executing the instructions stored in the memory to implement the task execution method based on synthetic data quality control.

[0011] A computer-readable storage medium storing at least one instruction, which is executed by a processor in a computer device to implement the task execution method based on synthetic data quality control.

[0012] As can be seen from the above technical solutions, this invention can extract data from the initial synthetic data based on answer extraction rules to reduce data noise; it calls the basic model to perform a single inference evaluation on the initial dataset to obtain a first inference result, and then rewrites the initial dataset with adaptive difficulty based on the first inference result to increase the proportion of learnable samples; it calls the basic model to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and then categorizes the rewritten samples based on the second inference result to make the output data more concentrated in the learnable range, thereby improving training effectiveness and reducing training resource waste; it uses the training set to perform post-training on the pre-trained model, and calls the baseline model and preset evaluation set to evaluate the task execution model, thus forming an end-to-end closed-loop iteration; when the evaluation result shows that the task execution model meets the standard, it uses the task execution model to respond to the task execution instructions, thereby effectively executing the task in combination with high-quality synthetic data and improving the task execution effect. Attached Figure Description

[0013] Figure 1 This is a flowchart of a preferred embodiment of the task execution method based on synthetic data quality control of the present invention.

[0014] Figure 2 This is a functional block diagram of a preferred embodiment of the task execution device based on synthetic data quality control of the present invention.

[0015] Figure 3 This is a schematic diagram of the structure of a computer device that implements a preferred embodiment of the task execution method based on synthetic data quality control according to the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1 The diagram shown is a flowchart of a preferred embodiment of the task execution method based on synthetic data quality control according to the present invention. The order of the steps in this flowchart can be changed, and some steps can be omitted, depending on different requirements.

[0018] The task execution method based on synthetic data quality control is applied to one or more computer devices. The computer device is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0019] The computer device can be any electronic product that can interact with the user, such as a personal computer, tablet computer, smartphone, personal digital assistant (PDA), game console, interactive network television (IPTV), smart wearable device, etc.

[0020] The computer equipment may also include network equipment and / or user equipment. The network equipment includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of hosts or network servers.

[0021] The server can be a standalone server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0022] Artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0023] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0024] The network in which the computer device is located includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, and virtual private network (VPN).

[0025] S10, in response to the task execution instruction based on the initial synthetic data, data extraction is performed on the initial synthetic data based on the answer extraction rules to obtain the initial dataset.

[0026] In this embodiment, the initial synthetic data can be data generated through any synthetic method.

[0027] In this embodiment, the task execution instruction can be automatically triggered when the generation of the initial synthetic data is detected.

[0028] In this embodiment, the step of extracting data from the initial synthetic data based on the answer extraction rules to obtain the initial dataset includes: The initial synthesized data is subjected to structured parsing and format verification to extract candidate structured data; Intermediate data is extracted from the candidate structured data according to the preset identifier, and multiple data synthesis operations are performed on the original data corresponding to the intermediate data. Data is selected from the multiple synthesized data based on the voting mechanism and added to the initial dataset. Specifically, when intermediate data fails to vote, the intermediate data that failed to vote is marked, or a rollback process is performed based on the intermediate data that failed to vote; when the amount of data in the initial dataset is less than the data amount threshold, the intermediate data that failed to vote is directly retained in the initial dataset.

[0029] In this process, by performing structured parsing and format verification on the initial synthesized data, fields that meet the sample specifications can be extracted, and outputs that cannot be parsed or have missing fields can be filtered out.

[0030] The preset identifier may include a preset mark or label, which can be used to uniquely identify the data.

[0031] Based on the voting mechanism, stable output data can be selected from multiple synthesized data.

[0032] The data volume threshold can be the minimum sample size required to meet training needs.

[0033] This embodiment extracts the final answer from the synthetic data based on preset answer marking or labeling rules; it generates multiple times for the same input and selects a stable output through majority voting to reduce generation fluctuations and format deviations; it triggers failure processing or retains the original output when voting fails or there are insufficient valid candidates, thereby improving sample availability and consistency, and reducing the probability of noisy samples entering the training set.

[0034] S11, the basic model is called to perform a single inference evaluation on the initial dataset to obtain a first inference result, and the initial dataset is rewritten in an adaptive difficulty manner based on the first inference result to obtain a rewritten sample set.

[0035] In this embodiment, the base model can be a benchmark AI model used for reasoning and evaluating samples.

[0036] In this embodiment, the step of adaptively rewriting the initial dataset based on the first inference result to obtain a rewritten sample set includes: Based on the first inference result, data that can be solved by the basic model is obtained from the initial dataset as easy samples, and data that cannot be solved by the basic model is obtained from the initial dataset as hard samples; Perform increased difficulty rewriting on the easy samples and decreased difficulty rewriting on the difficult samples; The rewritten data is then used to update the initial dataset to obtain the rewritten sample set.

[0037] The easy samples refer to the samples that the basic model can easily solve. The core of increasing the difficulty is to increase the complexity, ambiguity or reasoning depth of the problem, making it difficult for the basic model to directly output the answer.

[0038] The difficult samples refer to those that the basic model cannot solve. The core of difficulty reduction is to reduce the complexity of the problem and break down the task steps so that the basic model can process and output the answer step by step.

[0039] By performing difficulty-adaptive rewriting, the effectiveness of the dataset can be improved, invalid data redundancy can be avoided, and a high-quality training dataset can be constructed.

[0040] S12, the basic model is called to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and the rewritten samples are graded and exported according to the second inference result to obtain a training set.

[0041] In this embodiment, the step of grading the rewritten samples according to the second inference result to obtain the training set includes: Calculate the pass rate score for each sample in the rewritten sample set based on the second inference result; The samples with a pass rate score of 0 are obtained from the rewritten sample set as unsolved samples, the samples with a pass rate score of 1 are obtained from the rewritten sample set as solved samples, and the samples with a pass rate score greater than 0 and less than 1 are obtained from the rewritten sample set as learnable samples. The training set is obtained by classifying and categorizing the unsolved samples, solved samples, and learnable samples.

[0042] The pass rate score is the percentage of successful attempts across multiple reasoning evaluations.

[0043] In this process, after performing multiple inference evaluations on the rewritten sample set, multiple rounds of difficulty-adaptive rewriting can be performed on the rewritten sample set based on the obtained second inference results, so as to further increase the number of learnable samples. The specific number of rounds can be configured according to the actual sample size requirements.

[0044] In the above embodiments, by exporting training data by category, it is possible to achieve refined and differentiated use of training data. Samples of different categories correspond to different optimization directions of the model. After exporting by category, training data can be selected as needed, avoiding one-size-fits-all training, improving the efficiency and targeting of model training, and reducing the cost of ineffective training.

[0045] In this embodiment, after obtaining the training set, the method further includes: The training set is subjected to deduplication, similarity threshold filtering, length or format rule filtering, etc., to reduce homogenization and noise.

[0046] In this embodiment, after obtaining the training set, the method further includes: Synchronously record relevant metadata such as scores, categories, and strategy parameters for easy traceability.

[0047] In this embodiment, after obtaining the training set, the method further includes: When the configured language type is not the current default language, the translation model is invoked to translate the input and output fields of each training sample in the training set; During the translation process, the structure and identifiers of each field in each training sample are kept unchanged to ensure structural consistency and trainability of the multilingual output. After translation, the translated training set and corresponding metadata are associated and archived for traceability.

[0048] The corresponding metadata may include the configured language type, model identifier, and parameters.

[0049] Without altering the main pipeline structure, the system can be expanded to more language tasks by adding language adaptation configurations. For example, multilingual adaptation can be achieved by configuring different language generation, rewriting, and translation strategies, enabling low-cost multilingual migration and consistent output, thus improving the efficiency and trainability of cross-language data construction. Specifically, by introducing target language configuration and translation / rewriting components under a unified sample specification and closed-loop process, the synthesized training samples can be converted into target language versions of the dataset while maintaining structural constraints (at least including input and output fields, and protecting the post-labeling of answer tags), and seamlessly integrated with data evaluation, rewriting, hierarchical export, and post-training processes. Compared to existing methods that require designing separate prompts, collecting or cleaning data separately for different languages, this significantly reduces the cost of cross-language data construction and iteration, improves data supply capacity and training efficiency in low-resource language scenarios, and ensures consistency in format, quality control, and process reproducibility across different language versions of the data.

[0050] In this embodiment, while exporting the training set, optional subsets that can be selected and used as needed can also be exported simultaneously based on additional filtering conditions.

[0051] For example, the optional subset may include a high-confidence learnable subset for optimizing the model's reasoning ability for medium-difficulty tasks, a subset of rewriting strategies for verifying the impact of different rewriting strategies on model training, a subset of task types for specialized training on a single task, a subset of validation for verifying the model's basic capabilities, and a lightweight subset for rapid small-batch training and validation, etc.

[0052] This embodiment employs a closed-loop screening method based on data evaluation, adaptive difficulty rewriting, re-evaluation, and tiered derivation. It uses a base model to perform inference evaluation on the initial dataset to obtain scores. Based on these scores, samples are categorized as too easy, learnable, or too difficult, and a "difficulty-increasing, maintenance, and difficulty-reducing" rewriting strategy is implemented. The rewriting results are re-evaluated (including but not limited to the pass@n method), and different subsets are derived according to the scores, thus concentrating the output data within the learnable range, improving training effectiveness, and reducing training resource waste. This addresses the problems of existing synthetic data lacking quantitative evaluation and difficulty control, sample homogeneity or imbalanced difficulty distribution, and low training resource utilization efficiency. It makes the output data more concentrated within the learnable range and facilitates tiered use according to purpose. Compared to methods that only perform rule filtering or sampling, it can more systematically control the difficulty distribution and improve the efficiency of training resource utilization.

[0053] This embodiment provides a quality control and consistency stabilization mechanism for synthetic data, which solves problems such as non-compliant format, inconsistent answers, and noise sample mixing caused by random fluctuations in the generated output, thereby improving the structured usability and consistency of synthetic data.

[0054] S13, use the training set to perform post-training on the pre-trained model to obtain the task execution model.

[0055] In this embodiment, the step of using the training set to perform post-training on the pre-trained model to obtain the task execution model includes: Get the current training configuration data, training method, and current training framework; The training set is converted into the format required by the current training framework based on the current training configuration data. Based on the training method, the pre-trained model is trained on the current training framework using the transformed training set to obtain the task execution model; The system records the version number of the task execution model and the training log.

[0056] The current training configuration data may include training parameters, data version, baseline model version, and environment information, among other training-related data.

[0057] For example, training methods such as SFT (Supervised Fine-Tuning) or GRPO (Group Relative Policy Optimization) can be selected. The training set is converted into the format required by the current training framework according to the current training configuration data. Training is performed on the GPU (Graphics Processing Unit) and the trained model version and training logs are generated. For example, SFT and / or GRPO training can be performed based on the VERL (Volcano Engine Reinforcement Learning for LLMs, an open-source reinforcement learning library released by the ByteDance Seed team) framework, or an equivalent post-training process can be implemented based on other training frameworks.

[0058] In addition to recording the version number and training log of the task execution model, it can also record checkpoints and optional release artifacts (such as export or upload) and other related data.

[0059] In the above embodiments, high-quality synthetic data directly benefits post-training of the model, improving the performance of the target task and the domain adaptation effect. Training sets that have undergone quality control, adaptive difficulty adjustment, and tiered selection allow the data distribution to more closely match the model's learnable range and cover key knowledge points, task structures, and reasoning styles in the target domain (e.g., CoT (Chain of Thought) stepwise reasoning, answer labeling, etc.), thereby increasing the benefits of post-training with the same or lower manual annotation costs. Compared to directly using unfiltered public data or weakly constrained generated data, the synthetic data in this embodiment can improve training stability and the proportion of effective samples, reduce the negative impact of noisy samples on training, and in some embodiments, this manifests as a significant improvement in the accuracy, pass rate, or overall performance of the trained model on the target domain evaluation set.

[0060] S14, the baseline model and preset evaluation set are called to evaluate the task execution model and obtain the evaluation results.

[0061] In this embodiment, the evaluation of the task execution model by invoking the baseline model and the preset evaluation set, and obtaining the evaluation results include: The baseline model is used to process the preset evaluation set to obtain baseline data; The preset evaluation set is processed using the task execution model to obtain the data to be evaluated; The baseline data is scored according to a preset indicator evaluation system to obtain a benchmark score; The data to be evaluated is scored according to the preset indicator evaluation system to obtain the evaluation score; Calculate the improvement of the score to be evaluated relative to the benchmark score; The evaluation result is generated based on the stated improvement.

[0062] For example, when the improvement is greater than or equal to a preset range, it can be determined that the task execution model meets the standard; when the improvement is less than the preset range, it can be determined that the task execution model does not meet the standard.

[0063] An evaluation report can be generated based on the evaluation results. The evaluation report may include data such as model version comparison, key indicators, representative failure cases and optional error statistics, and can be archived in association with training configuration and data version information.

[0064] In this embodiment, the evaluation results and the evaluation report can serve as the basis for the next round of data synthesis and training parameter adjustment, thereby driving the formation of an integrated closed loop of data synthesis, post-model training, and model evaluation. This closed loop uniformly records and manages the dataset version, training configuration, and model output during execution, supporting traceable and reproducible batch execution. It solves the problems of fragmented synthesis, training, and evaluation links, inconsistent configurations, and difficulty in reproducing results in traditional solutions. Compared with existing methods where synthesis, training, and evaluation are independent and it is difficult to align versions and parameters, this significantly reduces the cost of reproducing experiments and attributing results.

[0065] S15, when the evaluation result indicates that the task execution model meets the standard, the task execution model is used to respond to the task execution instruction.

[0066] When the evaluation result indicates that the task execution model meets the standard, it can be determined that the performance of the task execution model is good. Therefore, using the task execution model to respond to the task execution instructions can improve the task execution effect.

[0067] For example, the task execution model can be used to perform inference tasks, classification tasks, prediction tasks, etc.

[0068] In this embodiment, the modality adapter and unified sample specification can be extended to modalities such as images or audio: define the interface for modality acquisition, preprocessing, context binding (such as surrounding text, subtitles or transcription, metadata) and model input organization, and reuse the same set of quality control, data closed-loop screening, training and evaluation main processes.

[0069] By standardizing sample specifications and modality adaptation interfaces, text, image, and audio modal data, along with their contextual information (such as surrounding text, subtitles or transcriptions, and metadata), can reuse the same set of quality control, data closed-loop screening, post-training, and evaluation processes. Compared to existing methods that customize processes separately for each modality, extending to new modalities has lower engineering costs and stronger process consistency.

[0070] As can be seen from the above technical solutions, this invention can extract data from the initial synthetic data based on answer extraction rules to reduce data noise; it calls the basic model to perform a single inference evaluation on the initial dataset to obtain a first inference result, and then rewrites the initial dataset with adaptive difficulty based on the first inference result to increase the proportion of learnable samples; it calls the basic model to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and then categorizes the rewritten samples based on the second inference result to make the output data more concentrated in the learnable range, thereby improving training effectiveness and reducing training resource waste; it uses the training set to perform post-training on the pre-trained model, and calls the baseline model and preset evaluation set to evaluate the task execution model, thus forming an end-to-end closed-loop iteration; when the evaluation result shows that the task execution model meets the standard, it uses the task execution model to respond to the task execution instructions, thereby effectively executing the task in combination with high-quality synthetic data and improving the task execution effect.

[0071] like Figure 2 The diagram shown is a functional block diagram of a preferred embodiment of the task execution device 11 based on synthetic data quality control of the present invention. The task execution device 11 based on synthetic data quality control includes an extraction unit 110, a rewriting unit 111, an export unit 112, a post-training unit 113, an evaluation unit 114, and a response unit 115. The module / unit referred to in this invention refers to a series of computer program segments that can be executed by a processor and perform a fixed function, and are stored in memory. In this embodiment, the functions of each module / unit will be described in detail in subsequent embodiments.

[0072] The extraction unit 110 is configured to extract data from the initial synthetic data based on the answer extraction rules in response to a task execution instruction based on the initial synthetic data, thereby obtaining an initial dataset. The rewriting unit 111 is used to call the base model to perform a single inference evaluation on the initial dataset to obtain a first inference result, and to perform difficulty adaptive rewriting on the initial dataset based on the first inference result to obtain a rewritten sample set. The export unit 112 is used to call the base model to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and to export the rewritten samples in a hierarchical manner according to the second inference result to obtain a training set. The post-training unit 113 is used to perform post-training on the pre-trained model using the training set to obtain the task execution model. The evaluation unit 114 is used to call the baseline model and the preset evaluation set to evaluate the task execution model and obtain the evaluation result; The response unit 115 is used to respond to the task execution instruction using the task execution model when the evaluation result indicates that the task execution model meets the standard.

[0073] As can be seen from the above technical solutions, this invention can extract data from the initial synthetic data based on answer extraction rules to reduce data noise; it calls the basic model to perform a single inference evaluation on the initial dataset to obtain a first inference result, and then rewrites the initial dataset with adaptive difficulty based on the first inference result to increase the proportion of learnable samples; it calls the basic model to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and then categorizes the rewritten samples based on the second inference result to make the output data more concentrated in the learnable range, thereby improving training effectiveness and reducing training resource waste; it uses the training set to perform post-training on the pre-trained model, and calls the baseline model and preset evaluation set to evaluate the task execution model, thus forming an end-to-end closed-loop iteration; when the evaluation result shows that the task execution model meets the standard, it uses the task execution model to respond to the task execution instructions, thereby effectively executing the task in combination with high-quality synthetic data and improving the task execution effect.

[0074] like Figure 3 The diagram shown is a structural schematic of a computer device that implements a preferred embodiment of the task execution method based on synthetic data quality control according to the present invention.

[0075] The computer device 1 may include a memory 12, a processor 13, and a bus (the arrow in the figure represents the bus), and may also include a computer program stored in the memory 12 and executable on the processor 13, such as a task execution program based on synthetic data quality control.

[0076] Those skilled in the art will understand that the schematic diagram is merely an example of computer device 1 and does not constitute a limitation on computer device 1. Computer device 1 can be either a bus topology or a star topology. Computer device 1 may also include more or fewer other hardware or software than shown in the diagram, or different component arrangements. For example, computer device 1 may also include input / output devices, network access devices, etc.

[0077] It should be noted that the computer device 1 described is merely an example. Other existing or future electronic products that are adaptable to this invention should also be included within the scope of protection of this invention and are incorporated herein by reference.

[0078] The memory 12 includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 12 can be an internal storage unit of the computer device 1, such as a portable hard drive of the computer device 1. In other embodiments, the memory 12 can be an external storage device of the computer device 1, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the computer device 1. Furthermore, the memory 12 can include both internal and external storage units of the computer device 1. The memory 12 can be used not only to store application software and various types of data installed on the computer device 1, such as the code of a task execution program based on synthetic data quality control, but also to temporarily store data that has been output or will be output.

[0079] In some embodiments, the processor 13 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 13 is the control unit of the computer device 1, connecting various components of the computer device 1 via various interfaces and lines. It executes programs or modules stored in the memory 12 (e.g., executing task execution programs based on synthetic data quality control) and calls data stored in the memory 12 to perform various functions of the computer device 1 and process data.

[0080] The processor 13 executes the operating system of the computer device 1 and various installed applications. The processor 13 executes the applications to implement the steps in the above-described embodiments of the task execution method based on synthetic data quality control, for example... Figure 1 The steps are shown.

[0081] For example, the computer program may be divided into one or more modules / units, which are stored in the memory 12 and executed by the processor 13 to complete the present invention. The one or more modules / units may be a series of computer-readable instruction segments capable of performing specific functions, which describe the execution process of the computer program in the computer device 1. For example, the computer program may be divided into an extraction unit 110, a rewriting unit 111, an export unit 112, a post-training unit 113, an evaluation unit 114, and a response unit 115.

[0082] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute portions of the task execution method based on synthetic data quality control described in the various embodiments of this invention.

[0083] If the modules / units integrated in the computer device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware devices. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above.

[0084] The computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory, etc.

[0085] Furthermore, the computer-readable storage medium may primarily include a stored program area and a stored data area, wherein the stored program area may store the operating system, an application program required for at least one function, etc.; and the stored data area may store data created based on the use of blockchain nodes, etc.

[0086] The blockchain referred to in this invention is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0087] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, in... Figure 3 The bus is represented by only one straight line, but this does not mean that there is only one bus or one type of bus. The bus is configured to enable communication between the memory 12 and at least one processor 13, etc.

[0088] Although not shown, the computer device 1 may also include a power supply (such as a battery) to power various components. Preferably, the power supply can be logically connected to the at least one processor 13 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The computer device 1 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0089] Furthermore, the computer device 1 may also include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), which is typically used to establish communication connections between the computer device 1 and other computer devices.

[0090] Optionally, the computer device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the computer device 1 and to display a visual user interface.

[0091] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0092] It will be understood by those skilled in the art that Figure 3 The structure shown does not constitute a limitation on the computer device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0093] Combination Figure 1 The memory 12 in the computer device 1 stores multiple instructions to implement a task execution method based on synthetic data quality control, and the processor 13 can execute the multiple instructions to achieve: In response to a task execution instruction based on initial synthesized data, data is extracted from the initial synthesized data based on answer extraction rules to obtain an initial dataset; The base model is invoked to perform a single inference evaluation on the initial dataset to obtain a first inference result, and the initial dataset is rewritten in an adaptive difficulty manner based on the first inference result to obtain a rewritten sample set; The base model is invoked to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and the rewritten samples are then graded and derived based on the second inference result to obtain a training set; The pre-trained model is post-trained using the training set to obtain the task execution model. The task execution model is evaluated using a baseline model and a preset evaluation set to obtain evaluation results; When the evaluation result indicates that the task execution model meets the standard, the task execution model is used to respond to the task execution instruction.

[0094] Specifically, the processor 13's implementation method for the above instructions can be found in [reference needed]. Figure 1 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0095] It should be noted that all the data involved in this case was legally obtained.

[0096] If any AI models, software tools, or components not belonging to this company appear in the embodiments of this invention, they are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this invention has been obtained by an entity authorized (with the knowledge and consent) or fully authorized by all parties through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0097] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0098] This invention can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0099] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0100] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0101] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0102] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0103] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices described in this invention can also be implemented by a single unit or device through software or hardware. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A task execution method based on synthetic data quality control, characterized in that, The task execution method based on synthetic data quality control includes: In response to a task execution instruction based on initial synthesized data, data is extracted from the initial synthesized data based on answer extraction rules to obtain an initial dataset; The base model is invoked to perform a single inference evaluation on the initial dataset to obtain a first inference result, and the initial dataset is rewritten in an adaptive difficulty manner based on the first inference result to obtain a rewritten sample set; The base model is invoked to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and the rewritten samples are then graded and derived based on the second inference result to obtain a training set; The pre-trained model is post-trained using the training set to obtain the task execution model. The task execution model is evaluated using a baseline model and a preset evaluation set to obtain evaluation results; When the evaluation result indicates that the task execution model meets the standard, the task execution model is used to respond to the task execution instruction.

2. The task execution method based on synthetic data quality control as described in claim 1, characterized in that, The initial dataset obtained by extracting data from the initial synthetic data based on the answer extraction rules includes: The initial synthesized data is subjected to structured parsing and format verification to extract candidate structured data; Intermediate data is extracted from the candidate structured data according to the preset identifier, and multiple data synthesis operations are performed on the original data corresponding to the intermediate data. Data is selected from the multiple synthesized data based on the voting mechanism and added to the initial dataset. Specifically, when intermediate data fails to vote, the intermediate data that failed to vote is marked, or a rollback process is performed based on the intermediate data that failed to vote; when the amount of data in the initial dataset is less than the data amount threshold, the intermediate data that failed to vote is directly retained in the initial dataset.

3. The task execution method based on synthetic data quality control as described in claim 1, characterized in that, The step of adaptively rewriting the initial dataset based on the first inference result to obtain the rewritten sample set includes: Based on the first inference result, data that can be solved by the basic model is obtained from the initial dataset as easy samples, and data that cannot be solved by the basic model is obtained from the initial dataset as hard samples; Perform increased difficulty rewriting on the easy samples and decreased difficulty rewriting on the difficult samples; The rewritten data is then used to update the initial dataset to obtain the rewritten sample set.

4. The task execution method based on synthetic data quality control as described in claim 1, characterized in that, The step of grading the rewritten samples based on the second inference result to obtain the training set includes: Calculate the pass rate score for each sample in the rewritten sample set based on the second inference result; The samples with a pass rate score of 0 are obtained from the rewritten sample set as unsolved samples, the samples with a pass rate score of 1 are obtained from the rewritten sample set as solved samples, and the samples with a pass rate score greater than 0 and less than 1 are obtained from the rewritten sample set as learnable samples. The training set is obtained by classifying and categorizing the unsolved samples, solved samples, and learnable samples.

5. The task execution method based on synthetic data quality control as described in claim 4, characterized in that, After obtaining the training set, the method further includes: When the configured language type is not the current default language, the translation model is invoked to translate the input and output fields of each training sample in the training set; During the translation process, the structure and identifiers of each field in each training sample remain unchanged, and after the translation is completed, the translated training set is associated with and archived with the corresponding metadata.

6. The task execution method based on synthetic data quality control as described in claim 1, characterized in that, The step of using the training set to train the pre-trained model to obtain the task execution model includes: Get the current training configuration data, training method, and current training framework; The training set is converted into the format required by the current training framework based on the current training configuration data. Based on the training method, the pre-trained model is trained on the current training framework using the transformed training set to obtain the task execution model; The system records the version number of the task execution model and the training log.

7. The task execution method based on synthetic data quality control as described in claim 1, characterized in that, The evaluation of the task execution model using the baseline model and the preset evaluation set yields the following results: The baseline model is used to process the preset evaluation set to obtain baseline data; The preset evaluation set is processed using the task execution model to obtain the data to be evaluated; The baseline data is scored according to a preset indicator evaluation system to obtain a benchmark score; The data to be evaluated is scored according to the preset indicator evaluation system to obtain the evaluation score; Calculate the improvement of the score to be evaluated relative to the benchmark score; The evaluation result is generated based on the stated improvement.

8. A task execution device based on synthetic data quality control, characterized in that, The task execution device based on synthetic data quality control includes: An extraction unit is configured to, in response to a task execution instruction based on initial synthetic data, extract data from the initial synthetic data based on answer extraction rules to obtain an initial dataset. The rewriting unit is used to call the base model to perform a single inference evaluation on the initial dataset to obtain a first inference result, and to perform difficulty adaptive rewriting on the initial dataset based on the first inference result to obtain a rewritten sample set. The export unit is used to call the base model to perform multiple inference evaluations on the rewritten sample set to obtain a second inference result, and to export the rewritten samples in a hierarchical manner based on the second inference result to obtain a training set. The post-training unit is used to perform post-training on the pre-trained model using the training set to obtain the task execution model. The evaluation unit is used to call the baseline model and the preset evaluation set to evaluate the task execution model and obtain the evaluation results; A response unit is used to respond to the task execution instruction using the task execution model when the evaluation result indicates that the task execution model meets the standard.

9. A computer device, characterized in that, The computer device includes: A memory for storing at least one instruction; and a processor for executing the instructions stored in the memory to implement the task execution method based on synthetic data quality control as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, which is executed by a processor in a computer device to implement the task execution method based on synthetic data quality control as described in any one of claims 1 to 7.

Images