A large model data purification iteration method based on user incentive interaction and multi-AI collaborative determination

Abstract

This invention discloses a method for refining and iteratively improving large-scale model data based on user-incentivized interaction and multi-AI collaborative judgment. It guides user interaction with the large-scale model through diverse user incentive strategies, employing at least two independent AI models working collaboratively. The primary AI handles the interaction, while independent AIs uniformly judge data quality and behavioral authenticity. The comprehensive judgment results can be used for data filtering and incentive distribution decisions. Finally, high-value data is fed back into the large-scale model for training iteration. This invention effectively filters spam data and malicious behavior, reduces incentive costs, improves data quality and model iteration efficiency, and is applicable to various large-scale artificial intelligence model systems, possessing high industry value and versatility.

Description

Technical Field This invention relates to the field of artificial intelligence large model technology, specifically to a method for data purification and iteration of large models based on user-incentivized interaction and multi-AI collaborative judgment. It is applicable to data collection, risk control filtering, high-quality data purification, and continuous model optimization and upgrading of various large models and intelligent interactive systems. Background Technology The training of current large-scale artificial intelligence models heavily relies on real, high-quality, and large-scale user interaction data. While using incentives to guide user interaction can rapidly expand data sources, it also introduces a large amount of low-quality content, bot-generated traffic, and malicious manipulation by "wool party" members, leading to wasted incentive costs, training data pollution, and a decline in model training effectiveness.

[0001] Existing technologies typically implement incentive distribution, data filtering, and model training independently, failing to use the results of multiple AI judgments simultaneously for data screening and incentive decisions. They lack a complete, efficient, and low-cost closed-loop solution, which cannot meet the industry's demand for high-quality, low-cost, and sustainable iteration of large models. Summary of the Invention The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for purifying and iterating large model data based on user-incentivized interaction and multi-AI collaborative judgment. This method guides user interaction through incentive strategies and utilizes at least two different, independent AI models to work collaboratively. The main AI completes the interaction, while the independent AI completes the comprehensive judgment. The judgment results can be used for data filtering and incentive distribution decisions, achieving high-quality data purification and precise control of incentive costs, and significantly improving the training efficiency and iteration effect of large models. This invention is achieved through the following technical solution: User incentive strategies are used to guide users to interact with the target large model, and raw interaction data is collected. At least two independent AI models from different sources or operating entities are used to process the data collaboratively. The first AI model is responsible for user interaction, while the second AI model is responsible for data quality judgment and user behavior authenticity detection. Based on the comprehensive judgment results of multiple AI models, the data is classified, screened, and filtered, and can be used as a basis for incentive distribution decisions. The selected high-value data is directly fed back into the target large model to complete model training, optimization, and iterative upgrades. The beneficial effects of this invention are: 1. Quickly acquire large-scale, authentic, and effective user interaction data through diversified incentives; 2. Employing independent AI collaborative judgment from different entities avoids self-judgment and self-inspection by a single AI, resulting in more objective and reliable judgment results; 3. Data quality and behavior detection are uniformly completed by independent AI, resulting in a simple process that is not easily bypassed; 4. The judgment results can be used for data filtering and incentive decision-making, controlling costs and eliminating waste from the source; 5. High-value data can be directly used for model training, significantly improving the efficiency and effectiveness of large model iterations; 6. The incentive formats cover both regular incentives and advertising interaction rewards, with a wide range of applications and clear boundaries; 7. The overall solution forms a complete technical closed loop, has strong versatility, reasonable protection scope, and high commercial value. Detailed Implementation The incentive interaction module guides users to engage in effective interactions with the target big model, such as dialogue, command execution, content generation, and information feedback, through incentive methods such as red envelopes, points, task rewards, advertising interaction rewards, and membership benefits, and collects raw user interaction data. Multiple AI decision-making modules initiate collaborative processing: The first AI model, as the main interactive AI, directly interacts with users, generates content, and responds to commands. The second AI model, acting as an independent external judgment AI, performs a unified judgment on the interaction data, including the validity, logic, value, and quality of the content. It also identifies risky data such as whether the user is operating a machine, using multiple accounts, engaging in fraudulent activities, maliciously inflating traffic, or exhibiting abnormal behavior. The data filtering and incentive decision-making module classifies, filters, and labels data based on the comprehensive judgment results of multiple AI models, and can make incentive distribution decisions based on the judgment results. High-value data is retained and enters the model training process, while low-value and abnormal data are directly filtered or blocked. The model iteration module automatically imports high-value data through verification into the target large model training dataset to complete the continuous optimization, capability upgrade and effect iteration of the model.

[0002] The above process is executed cyclically, forming a complete closed-loop system of user incentives, multi-AI collaborative judgment, data purification, incentive decision-making, and model iteration.

Claims

1. A method for iterative purification of large model data based on user-incentivized interaction and multi-AI collaborative judgment, characterized in that, Includes the following steps: User incentive strategies are used to guide users to interact with the target large model, and raw interaction data is collected. The interaction data is processed in parallel by at least two independent AI models to make a comprehensive judgment on data quality and the authenticity of user behavior. Based on the comprehensive judgment results of multiple AI models, the interactive data is value-graded and filtered, and can be used for data filtering and incentive distribution decisions; The selected high-value data is fed back into the target large model for training and iterative upgrades.

2. The method according to claim 1, characterized in that, The incentive strategies include, but are not limited to: cash red envelopes, cash subsidies, points, task rewards, virtual currency, membership benefits, redemption benefits, advertising interaction rewards, advertising incentives, advertising cashback, and other incentive methods used to guide user interaction.

3. The method according to claim 1, characterized in that, The at least two independent AI models are independent AI models with different training systems, different algorithm sources, or different operating entities, wherein: The first AI model, as the main interactive AI, is responsible for dialogue and interaction with users, content generation, and information response; The second AI model, as an independent judgment AI, is responsible for judging the validity, logic, information value, and content quality of interactive data, and for detecting the authenticity of user behavior, abnormal operations, and malicious behavior.

4. The method according to claim 1, characterized in that, The comprehensive judgment result is also used to: screen high-value data for model training, and can serve as a basis for whether to issue incentives.

5. The method according to claim 1, characterized in that, The comprehensive judgment method includes any one of the following: voting mechanism, weighted scoring, threshold judgment, and logical combination judgment.

6. The method according to claim 1, characterized in that, Low-value, spam, or anomalous data can be flagged, filtered, or blocked to prevent them from being used in the training of large models.

7. A large-scale model data purification and iterative system based on user-incentivized interaction and multi-AI collaborative judgment, characterized in that, include: The incentive interaction module is used to guide users to interact with the large model and collect raw data through user incentive strategies. The multi-AI judgment module contains at least two independent AI models from different entities, which work together to make a comprehensive judgment on the interactive data. The data filtering and incentive decision-making module is used to complete data filtering and incentive distribution decisions based on the results of multiple AI judgments. The model iteration module is used to feed the selected high-value data back into the target large model to complete model training and iterative upgrades.

8. The system according to claim 7, characterized in that, The system can be deployed independently or integrated into large model platforms, intelligent interaction platforms, user operation platforms, and advertising incentive platforms.