End-cloud collaboration-based large model ecological online evolution learning method and application

Abstract

The invention provides a large model ecological online evolution learning method based on end-cloud collaboration and an application thereof, and the method is applied to the field of e-commerce, so that efficient training and deployment of an e-commerce model are realized. According to the method, an end-cloud collaborative architecture is adopted, a cloud end deploys a large language model, an end side uses a small and medium-sized language model, and efficient training and reasoning of large model ecology are realized through collaborative work of the cloud end and the end side. In the training process, the cloud model serves as a teacher model, a pseudo tag is generated to supplement a data set, and optimization training of the end side model is guided. Meanwhile, a multi-model cooperation self-training mechanism is introduced, multiple models play different roles and mutually discuss, training labels are autonomously generated, and the end side model is further optimized. And meanwhile, user feedback information is brought into cue words of the cloud model for situational learning, so that the model performance is continuously optimized. The method is applied to an e-commerce scene, the e-commerce model ecology is successfully constructed and deployed, and intelligent development of the field is realized.

Description

Technical Field

[0001] The present invention relates to the field of e-commerce technology, and specifically to a large-model ecological online evolutionary learning method and application based on end-cloud collaboration. Background Art

[0002] With the rapid development of artificial intelligence technology, language models are increasingly used in the field of natural language processing, especially in customer service, content generation, data analysis and other tasks. However, existing customer service systems still face many challenges, especially in how to efficiently respond to changing user needs and ensure data security. Take the e-commerce field as an example. In e-commerce platforms, customer service systems undertake multiple tasks such as providing consultation and services to customers, answering questions, and handling complaints. These tasks require not only rapid responses, but also a high degree of situational understanding and personalized service capabilities.

[0003] At present, mainstream customer service solutions are mainly divided into the following categories: Manual customer service: Relying on a large number of manpower to handle customer inquiries and complaints, it can provide personalized services, but its scalability is limited due to high costs and difficulty in coping with a large number of inquiries during peak periods.

[0004] Rule-based automatic customer service system: uses preset rules and keywords to match and answer questions. Although it can handle some simple customer questions, it is difficult to handle complex and diverse customer needs, and its intelligence and accuracy are still limited.

[0005] Systems based on traditional machine learning or simple deep learning: Some systems use machine learning algorithms such as decision trees and support vector machines, or use models such as recurrent neural networks. Although these systems have improved the level of intelligence to a certain extent, they still perform poorly in complex scenarios.

[0006] In recent years, the emergence of large models (such as GPT-3, Gemini, etc.) has brought great progress in the field of natural language processing. Large language models have the ability to handle complex language tasks by pre-training on massive data. According to the expansion law of large models, the final performance of the model has a certain relationship with the amount of calculation, the number of model parameters, and the size of the data. When not restricted by the other two factors, the model performance shows a power law relationship with each factor. As the scale of model parameters increases, the performance of large models gradually improves, but the computing requirements also increase significantly, making it difficult to run directly on end-side devices with limited computing power.

[0007] To solve this problem, cloud computing was introduced. For example, in e-commerce, large models can be deployed in the cloud, and the end-side devices can process e-commerce customers' inquiries by calling the inference capabilities of the cloud model. However, although this cloud computing-based solution improves the availability of large models, it also brings risks of communication delay and privacy leakage, especially when processing sensitive data of e-commerce customers, which faces huge security risks.

[0008] In this context, end-cloud collaboration technology has gradually become a major technical development trend of customer service systems. End-cloud collaboration is a way to collaboratively complete tasks in a distributed computing environment by coordinating cloud and edge computing resources. This technology combines the advantages of cloud computing and edge computing, allowing models of different sizes to be deployed on the cloud and end for collaborative reasoning and training. In e-commerce scenarios, end-side devices can deploy lightweight models to process simple or privacy-related data, while the cloud handles complex tasks. Through this collaborative approach, the computing resources of the cloud and end can be fully utilized to improve the response speed of customer service and reduce the risk of privacy leakage.

[0009] However, the existing end-cloud collaborative system still has some shortcomings. For example, in e-commerce customer service, the collaborative optimization of cloud models and end-side models is often not achieved, the cloud model cannot fully guide the optimization training of the end-side model, and the end-side model cannot provide feedback to the cloud model based on user feedback. At the same time, when training the end-side model, it still needs to rely on a large amount of manually labeled data for supervision and guidance, while labeled data in specific fields is scarce, which makes model training difficult and costly. In addition, the update of most large models relies on offline methods, which makes it difficult to quickly adapt to the rapidly changing customer needs and market environment in e-commerce. These problems limit the response speed and accuracy of the customer service system of the e-commerce platform when facing complex and changing customer needs.

[0010] As an adaptive learning method, online evolutionary learning can continuously optimize the model and enhance its personalized service capabilities based on user feedback and environmental changes. For example, in e-commerce scenarios, online evolutionary learning can help large language models continuously evolve based on real-time user feedback, thereby providing more accurate and personalized responses, which helps improve the e-commerce customer service experience.

[0011] In summary, the existing customer service system and large model deployment solutions are still insufficient in dealing with complex scenarios. The end-cloud collaborative architecture can not only efficiently coordinate cloud and end-side resources, but also introduce multi-model self-training mechanisms and online evolutionary learning to continuously optimize the model in a dynamic environment and provide more personalized services. Applying this technical solution to the e-commerce field is expected to improve the intelligence level and service quality of e-commerce customer service systems. Summary of the invention

[0012] In view of the shortcomings of the existing technology, the present invention provides a large model ecological online evolutionary learning method and application based on end-cloud collaboration, which is applied to the e-commerce field. By designing a multi-model collaborative discussion mechanism and introducing multiple role models, the human team collaborative discussion process is effectively simulated, so that the model can autonomously generate training data during writing, reducing the dependence on manually labeled data. At the same time, by introducing human feedback and online fine-tuning, situational learning and other strategies, the online evolution of the models on both sides is effectively realized.

[0013] To achieve the above objectives, the present invention is implemented through the following technical solutions: a large model ecological online evolutionary learning method and application based on end-cloud collaboration, specifically including the following steps S1. Multi-model role allocation Multiple small models on the client side and one large cloud model are set to different roles to facilitate effective discussion and collaboration during the training process. The large cloud model is the teacher model, providing prior input and constraints for multi-model discussions. The role assignments of the small models on the client side include: Questioner model: Focus on asking industry-related questions and simulate the way real users ask questions.

[0014] Respondent model: responsible for generating responses based on industry knowledge and task objectives to improve the accuracy and professionalism of responses.

[0015] S2. Prior input and constraint support for large cloud models In each round of collaborative discussion, the cloud-based big model supports the training of the small model by providing prior input and constraints. Specifically, in each round of dialogue, the cloud-based big model has two functions: scene prior generation and generation quality assessment.

[0016] Scenario prior generation: Provides core industry knowledge, such as the best answer model in professional services, personalized service techniques, privacy data processing requirements, etc., so that small models have a unified standard reference during training.

[0017] Generation quality assessment: When the small model generates answers, the large model conducts quality assessment and feedback on these answers, and uses this constraint to guide the small model to generate content that better meets industry needs, ensuring the consistency and accuracy of the model.

[0018] S3. Discussion on multi-model collaboration During the collaboration process, each end-side model discusses by generating questions and answers to each other, forming multiple rounds of dialogue. This discussion mechanism enables the model to learn and understand various scenario requirements during interaction: Prior input: Before the questioner model and the answerer model generate questions and answers, the scenario prior of the large model is input to the questioner and answerer models to provide a prior knowledge base for the training of the small model on the terminal side, thereby improving the quality of question and answer generation.

[0019] Question generation: The questioner model asks questions based on industry needs and task scenarios.

[0020] Answer generation: The respondent model answers questions generated by the questioner model by providing detailed answers, analyzing scenario requirements, and improving service quality.

[0021] Quality verification: After each round of question and answer, the big model is used to evaluate and correct the dialogue generated by the end-side model to accelerate the convergence of end-side multi-model training.

[0022] S4. Generation and online evolution of collaborative training data During the discussion and collaboration process, through multiple rounds of dialogue, the answers generated between the small models are labeled as high-quality training data and added to the new data set after being constrained and verified by the large model. The small model is trained using the generated collaborative data and efficient parameter fine-tuning techniques (such as low-rank adaptive fine-tuning or prefix fine-tuning) to achieve optimization effects on specific tasks. Through regular end-cloud collaborative multi-model discussion and training mechanisms, the cloud and end-side models are continuously optimized online. The end-side models and cloud-side models continuously improve their ability to respond to industry scenarios through mutual learning and collaboration, and achieve adaptive learning and performance optimization. The online evolving end-side small model will have a stronger response capability, and will be able to adaptively process service tasks for specific industries based on collaborative training data, and achieve higher-quality responses.

[0023] Preferably, in S2, the customer service conversation dataset in a specific field is contextually modeled through the adaptive ability and contextual learning ability of the large model, so that it can handle specific complex problems in the scenario of the field and assume the collaborative a priori and constraint role for the small model.

[0024] Preferably, in S3, one model in the end-side model acts as the questioner, mainly responsible for raising specific questions in the industry, and the other model is responsible for answering the questions raised by the questioner. Information is exchanged through a discussion method of mutually generating questions and answers, autonomously mining the characteristics of conversation data in the field, and generating high-quality pseudo labels to guide the optimization training of the end-side model. The cloud model participates in and evaluates the discussion between small and medium models, and corrects the pseudo labels it generates as constraints, further improving the quality of information. Preferably, in S4, multiple small models on the end realize online evolutionary learning through collaborative discussion. The system automatically collects multi-model interaction feedback and integrates the feedback information into the training data of the end-side model to form a continuously updated data source. The small model on the end-side makes adaptive adjustments based on the prior knowledge and quality assessment of the large model during collaboration, and achieves self-optimization without manual intervention through multiple rounds of discussion and pseudo-label generation. The large cloud model uses the data set collected during the discussion as contextual prompts for contextual learning, continuously optimizes its generation capabilities, and achieves continuous adaptive optimization.

[0025] The present invention provides a large-model ecological online evolutionary learning method and application based on end-cloud collaboration. It has the following beneficial effects: 1. The present invention provides a large-model ecological online evolutionary learning method and application based on end-cloud collaboration. Through the end-cloud collaborative architecture, it innovatively solves the problem of insufficient computing power of end-side devices and inability to process complex tasks. In the traditional end-side model training process, a large amount of high-quality manually labeled data is often required as training corpus, but the cost of obtaining these labeled data is high and the efficiency is low. The present invention introduces the technology of generating pseudo labels by cloud models to automatically generate high-quality pseudo labels based on existing data, which not only greatly expands the training corpus of the end-side model, but also significantly reduces the dependence on manually labeled data and improves the training efficiency. In the reasoning process of traditional end-side models, the end-side model is limited by limited computing power and storage resources and cannot process complex tasks. The present invention proposes a new reasoning mechanism through the collaborative work of cloud models and end-side models. During the reasoning process, the cloud model can provide high-quality supplementary reasoning information for the end-side model, so that the end-side model not only integrates the output information of the cloud model during reasoning, but also combines its own local knowledge to provide users with more accurate and personalized responses, significantly improving the reasoning ability and accuracy.

[0026] 2. The present invention provides a large-model ecological online evolutionary learning method and application based on end-cloud collaboration. By introducing a multi-model collaborative discussion mechanism and realizing online evolutionary learning of cloud and end-side models, the system's adaptability and self-optimization capabilities are significantly improved. In traditional model training, end-side models often rely on a single data source and training method. In addition, traditional large model updates mostly rely on offline data and cannot respond to users' personalized needs in a timely manner. Based on end-cloud collaboration, the present invention generates new training labels through collaborative discussion and information feedback between the cloud model and several end-side models to achieve self-training and optimization. With the help of real-time feedback from users, the model can evolve online, continuously adjust weights or parameters, and gradually adapt to users' personalized needs and changing usage habits, forming a dynamically self-evolving intelligent customer service system. This mechanism greatly reduces the dependence on manually labeled data and improves the adaptability and intelligence of the system in complex scenarios.

[0027] 3. The present invention provides a large-model ecological online evolutionary learning method and application based on end-cloud collaboration. In view of the strict requirements on user privacy and data security in e-commerce scenarios, through the local privacy discrimination mechanism, when processing user requests, the end-side model can autonomously identify sensitive information and complete reasoning locally, avoiding uploading private data to the cloud and ensuring data security. This mechanism not only effectively enhances data security, but also meets the high standards for data privacy protection in e-commerce scenarios, further enhancing users' trust in the system and usage experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the large-model ecological online evolutionary learning method based on end-cloud collaboration of the present invention; Figure 2 A schematic diagram of the process flow of the multi-model collaborative discussion of the present invention; Figure 3 A schematic diagram of the process of collaborative reasoning and online evolutionary learning of the model of the present invention; Figure 4 The present invention is a flowchart of the implementation of the large model ecosystem based on end-cloud collaboration in the e-commerce field.

[0029] in, Figure 1 In the diagram, the icon on the left represents the database, the icon on the right represents the user, and the arrows represent the process of information flow and data transmission; Figure 2 and Figure 3 In the figure, each arrow represents the relationship and interaction process of each module; Figure 4 In the figure, each arrow represents the process of information flow and data transmission. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] like Figure 1-4 As shown, the embodiment of the present invention provides a large model ecological online evolutionary learning method and application based on end-cloud collaboration, which specifically includes the following steps S1. Cloud-based large model situational learning and adjustment By leveraging the powerful contextual learning capabilities of the cloud-based big model and using prompt word engineering, we can adjust it to a big model suitable for specific fields. Through the adaptive capabilities of the big model, we can perform contextual modeling on customer service conversation datasets in specific fields, enabling it to handle specific complex problems in scenarios in that field. Specifically, taking the e-commerce industry as an example, the specific process of applying the online evolutionary learning method of the large model ecosystem based on end-cloud collaboration to the actual industry is demonstrated. In the experiment, the Gemini 1.5 Pro model was selected as the cloud-side large model, and the ChatGLM3-6B model was selected as the end-side model. Combined with the steps in S1, the following prompt words were used as input in the cloud-side large model: "Suppose you are an e-commerce customer service, with rich e-commerce expertise, good service awareness, and can properly handle customer needs. Next, please play the role of an e-commerce customer service and handle customer problems. The following are examples of user questions and customer service answers for your reference. Please reply according to the customer service model in the example." Then, 500 actual multi-round conversations from the Taobao e-commerce platform were selected as supplements to the prompt words to further enrich the model's scenarios. For example: "User: 'Send me a clip.' Customer service: "I'm sorry, sir, we don't have clamps here, we only have openers." User: 'Are there any more discounts?' Customer Service: "Sir, if you buy more, you can use the coupon. Please check your delivery address." User: 'Yes.' Customer Service: "Okay, sir. We will ship it to you tomorrow. The courier has received it at 4 o'clock and will ship it tomorrow." By collecting real conversations on e-commerce platforms, combining user data with input into a large cloud model, and simulating multiple rounds of conversations, a hybrid dataset consisting of real e-commerce conversations and model-generated conversations was generated.

[0032] S2. Supervision Information Generation and Dataset Construction After the scenario learning of the cloud-based large model is completed, the cloud-based model generates supervision information for training the client-side model. At the same time, the supervision information generated by the cloud-based large model is used as a generated data set and mixed with the natural data set for subsequent model training. Specifically, the ChatGLM3-6B model is used as the client-side model and fine-tuned using a mixed dataset. The client-side model is trained using methods such as Low Rank Adaptive Fine-tuning (LoRA) and gradually adjusted to a model specifically designed for e-commerce customer service tasks.

[0033] S3. On-device model training and multi-model collaboration Based on high-quality supervision information generated in the cloud and supplementary industry dialogue data sets, the end-side model is trained using efficient parameter fine-tuning methods. At the same time, a multi-model collaborative discussion mechanism is introduced. The end-side model and other small and medium-sized models play different roles in the collaborative process and discuss with each other. Specifically, one model acts as a questioner, mainly responsible for raising specific questions in the industry, and the other model is responsible for answering questions raised by the questioner. Information is exchanged through a discussion method of mutually generating questions and answers, autonomously mining the characteristics of dialogue data in the field, and generating high-quality pseudo-labels to guide the optimization training of the end-side model. The cloud-side model participates in and evaluates the discussion between small and medium-sized models, and corrects the pseudo-labels they generate as constraints to further improve the quality of information. Specifically, in order to further improve the training effect of the model, a multi-model collaborative discussion method is introduced. In terms of specific implementation, the two pre-trained models on the end side play the roles of customers and e-commerce customer service respectively, and communicate with each other to generate pseudo labels. Among them, the end-side e-commerce model is the main trained model, and it is also the model that is subsequently deployed to the actual environment to interact with people. The end-side customer model is a model obtained by contextual learning based on the pre-trained model with the guidance of certain prompt words. The prompt words are: "Please play the role of a consumer (customer) of an e-commerce store, and I play the role of an e-commerce customer service. You are having multiple rounds of e-commerce customer service conversations with me. You can ask me all kinds of questions in the e-commerce field, including but not limited to product information, logistics, purchase return and exchange services, etc., and continue to complete the conversation based on the replies. Please imitate human consumers, talk naturally and colloquially, and output the command 'end of conversation' when you think the topic should end." S4. Collaboration between client-side models during the inference phase The trained client-side model is applied to actual customer service scenarios. When the client-side model receives a user's question, it first uses the local discriminant model to determine whether the question involves privacy. If the question involves privacy, the client-side model will answer it independently. If the question does not involve privacy, the question will be uploaded to the cloud model. The cloud model and client-side model will discuss and generate a reply together. The client-side model will combine the cloud model's answer to generate the final response. Specifically, for the cloud-based e-commerce model, additional prompts are required: "Please evaluate the quality of the e-commerce customer service response from the customer's perspective. Generally speaking, the criteria for judging the quality of e-commerce customer service responses include: whether the response is timely and effectively solves the customer's problem; whether the response is accurate, polite and professional; whether the communication is clear and friendly; whether personalized services are provided, etc. In addition, it is also necessary to consider whether the response conforms to the characteristics of colloquial language, and whether it is concise and efficient, reflecting the professional characteristics of the e-commerce customer service." The cloud-based large model learned through the above scenarios acts as a discriminant model to evaluate the quality of the conversation. If the conversation quality is judged to be high, these conversations are retained in the training corpus and used to fine-tune the client-side model.

[0034] S5. Real-time feedback and online fine-tuning When the client-side model performs reasoning tasks, it generates two different response plans at a certain frequency. After the user selects the better response, the answer is matched with the user's question and added to the new training set for online fine-tuning and optimization of the client-side model. At the same time, the collected human feedback is used as contextual prompts for the cloud-based large model to learn contextual situations and continuously optimize its generation capabilities. Specifically, when fine-tuning the client-side model, select mainstream model efficient parameter fine-tuning methods such as low-rank adaptive fine-tuning (LoRA) and prefix fine-tuning.

[0035] The fine-tuned end-side e-commerce model is deployed in actual e-commerce scenarios to handle user questions. When the end-side model receives a user's question, it first uses the local discriminant model to determine whether the question involves private information. If it involves privacy, the end-side model completes the reply independently; otherwise, the question is uploaded to the cloud-based large model. After the cloud-based large model generates a reply, the end-side model combines the output of the cloud-based model with the user's input to generate the final answer.

[0036] S6. Online evolution and model iteration Through regular human feedback and end-cloud collaborative training, the cloud and end-side models are continuously optimized online. The end-side and cloud-side models continuously improve their ability to respond to e-commerce customer service scenarios through mutual learning and collaboration, achieving adaptive learning and performance optimization.

[0037] Specifically, in order to achieve personalized services, the system generates one or two different replies with a certain probability when generating replies. If two replies are generated, the user can choose a more satisfactory reply from them. The result selected by the user will be paired with the question and added to the training corpus to perform online fine-tuning on the end-side model. At the same time, these high-quality dialogues selected by user feedback will also be used to update the prompt words of the cloud-based large model, further improving the supervision and generation capabilities of the cloud-based model.

[0038] By repeating this process, the client-side model and the cloud-side model continue to evolve online during continuous interaction and feedback with users, thereby providing more personalized and accurate e-commerce customer service.

[0039] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large-model ecological online evolutionary learning method and application based on end-cloud collaboration, characterized in that: The specific steps include: S1. Multi-model role allocation Multiple small models on the client side and one large model on the cloud side are set to different roles to facilitate effective discussion and collaboration during the training process. The large cloud model is the teacher model, providing prior input and constraints for multi-model discussions. The role allocation of the small models on the client side includes: Questioner model: Focus on asking industry-related questions and simulate the way real users ask questions. Respondent model: responsible for generating responses based on industry knowledge and task objectives to improve the accuracy and professionalism of responses. S2. Prior input and constraint support for large cloud models In each round of collaborative discussion, the cloud-based big model supports the training of the small model by providing prior input and constraints. Specifically, in each round of dialogue, the cloud-based big model has two functions: scene prior generation and generation quality assessment. Scenario prior generation: Provides core industry knowledge, such as the best answer model in professional services, personalized service techniques, privacy data processing requirements, etc., so that small models have a unified standard reference during training. Generation quality assessment: When the small model generates answers, the large model conducts quality assessment and feedback on these answers, and uses this constraint to guide the small model to generate content that better meets industry needs, ensuring the consistency and accuracy of the model. S3. Discussion on multi-model collaboration During the collaboration process, each end-side model discusses by generating questions and answers to each other, forming multiple rounds of dialogue. This discussion mechanism enables the model to learn and understand various scenario requirements during interaction: Prior input: Before the questioner model and the answerer model generate questions and answers, the scenario prior of the large model is input to the questioner and answerer models to provide a prior knowledge base for the training of the small model on the terminal side, thereby improving the quality of question and answer generation. Question generation: The questioner model asks questions based on industry needs and task scenarios. Answer generation: The respondent model answers questions generated by the questioner model by providing detailed answers, analyzing scenario requirements, and improving service quality. Quality verification: After each round of question and answer, the big model is used to evaluate and correct the dialogue generated by the end-side model to accelerate the convergence of end-side multi-model training. S4. Generation and online evolution of collaborative training data During the discussion and collaboration process, through multiple rounds of dialogue, the answers generated between the small models are labeled as high-quality training data and added to the new data set after being constrained and verified by the large model. The small model is trained using the generated collaborative data and efficient parameter fine-tuning techniques (such as low-rank adaptive fine-tuning or prefix fine-tuning) to achieve optimization effects on specific tasks. Through regular end-cloud collaborative multi-model discussion and training mechanisms, the cloud and end-side models are continuously optimized online. The end-side models and cloud-side models continuously improve their ability to respond to industry scenarios through mutual learning and collaboration, and achieve adaptive learning and performance optimization. The online evolving end-side small model will have a stronger response capability, and will be able to adaptively process service tasks for specific industries based on collaborative training data, and achieve higher-quality responses.

2. According to claim 1, a large model ecological online evolutionary learning method and application based on end-cloud collaboration is characterized by: In S2, the adaptive ability and situational learning ability of the large model are used to perform situational modeling on the dialogue dataset in a specific field, so that it can handle specific complex problems in the scenario of the field and assume the collaborative prior and constraint role of the small model.

3. According to claim 1, a large model ecological online evolutionary learning method and application based on end-cloud collaboration is characterized by: In S3, one model in the client-side model acts as the questioner, mainly responsible for raising industry-specific questions, while the other model is responsible for answering the questions raised by the questioner. Information is exchanged through a discussion method of mutually generating questions and answers, autonomously mining the characteristics of conversation data in the field, and generating high-quality pseudo labels to guide the optimization training of the client-side model. The cloud-side model participates in and evaluates the discussion between small and medium-sized models, and corrects the pseudo labels it generates as constraints, further improving the quality of information.

4. According to claim 1, a large model ecological online evolutionary learning method and application based on end-cloud collaboration is characterized by: In S4, multiple small models on the end realize online evolutionary learning through collaborative discussion. The system automatically collects multi-model interaction feedback and integrates the feedback information into the training data of the end-side model to form a continuously updated data source. The small model on the end makes adaptive adjustments based on the prior knowledge and quality assessment of the large model during collaboration, and achieves self-optimization without manual intervention through multiple rounds of discussion and pseudo-label generation. The large cloud model uses the data set collected during the discussion as contextual prompts for contextual learning, continuously optimizes its generation capabilities, and achieves continuous adaptive optimization.

Images