Method and device for constructing marketing vertical class model based on large language model, and storage medium

Abstract

The invention discloses a method for constructing a marketing vertical class model based on a large language model, and the method comprises the following steps: selecting an open-source large model as a base large model, and carrying out the pre-training; the method comprises the following steps: crawling a large amount of marketing-related high-quality copywriting data from the internet by using data crawling, and cleaning and preprocessing the data to obtain a marketing copywriting corpus; performing incremental pre-training on the base large model by utilizing a Chinese case in a corpus; on the basis of the incremental pre-trained base large model, a QLoRA technology is adopted to carry out fine adjustment on data of each subdivided industry on the incremental pre-trained base large model, and each industry obtains a LoRA adapter; connecting the base large model which is finely adjusted according to the data of each subdivided industry with LoRA adapters of different industries to obtain a marketing vertical model; fine adjustment of different subdivided industries is carried out on a large language model by utilizing high-quality data accumulated by previous marketing cases and platforms and data crawling, so that the marketing document generation performance of the model in various specific industries is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of generative artificial intelligence (GenAI) and natural language processing (NLP), and in particular to a method, device and storage medium for building a marketing vertical model based on a large language model. Background Art

[0002] With the rapid advancement of generative artificial intelligence and natural language processing technologies, large language models (such as Wenxin Yiyan) have demonstrated excellent performance in a wide range of application scenarios. However, when it comes to specific segments of the marketing industry, these general-purpose large language models are difficult to fully meet the complex and sophisticated needs in this field. The reason is that most current models are trained based on widely used data sets and lack in-depth mining and training of specific data in the marketing industry. This makes it difficult to accurately capture and reflect industry-specific details and professional terms when generating marketing copy, thus affecting the quality and effectiveness of the copy. Therefore, developing a more accurate and efficient large language model for marketing verticals has become an important issue that needs to be solved urgently.

[0003] The existing large language models in the marketing field are usually based on open source large models and are implemented through incremental pre-training of professional corpora or instruction fine-tuning. However:

[0004] 1. The marketing field involves many sub-industries, each of which has its own unique terminology, expressions, and market characteristics. A single marketing corpus cannot fully cover the knowledge of sub-industries, resulting in poor performance of the model in certain industries.

[0005] 2. Tasks in the marketing field are often highly complex and diverse. For example, different marketing strategies may require customized designs for different target audiences, product features, and market environments. This makes it difficult for a single instruction fine-tuning method to cover all possible scenarios and needs.

[0006] To this end, this application provides a method, device and storage medium for constructing a marketing vertical model by fine-tuning a large language model based on segmented marketing industry data. By utilizing past marketing cases, high-quality data accumulated by the platform and data crawling, the large language model is fine-tuned for different segmented industries to improve the model's marketing copy generation performance in various specific industries. Summary of the invention

[0007] The embodiment of the present application provides a method for constructing a marketing vertical model by fine-tuning a large language model based on segmented marketing industry data. By utilizing past marketing cases, high-quality data accumulated by the platform, and data crawling, the large language model is fine-tuned for different segmented industries to improve the model's performance in generating marketing copy in various specific industries.

[0008] The embodiment of the present application provides a method for building a marketing vertical model based on a large language model, including the following process:

[0009] Establishment of the base model: Select an open source model as the base model for pre-training;

[0010] Constructing a corpus: Using data crawling technology to crawl a large amount of high-quality marketing-related copywriting data from the Internet, and then cleaning and preprocessing the data to obtain a marketing copywriting corpus;

[0011] Incremental pre-training: Using the massive marketing copy collected in the corpus, perform incremental pre-training of all parameters of the base model to obtain the base model after incremental pre-training;

[0012] Build LoRA adapters for different industries: Based on the incremental pre-trained base model, QLoRA (Quantization and Low-Rank Adaptation) technology is used to fine-tune the incremental pre-trained base model with the data of each subdivided industry (during the fine-tuning process, the parameters of the base model are frozen, and only the parameters of the LoRA adapter are trainable), and each industry gets a LoRA adapter.

[0013] Build a marketing vertical model: Connect the base model fine-tuned according to the data of each sub-industry with the LoRA adapters of different industries to obtain a marketing vertical model;

[0014] Use of the marketing vertical model: Integrate LoRA adapters from multiple different industries to form a unified and flexible structure, enabling it to implement task processing in different industries by loading each LoRA layer separately.

[0015] The marketing copy data crawled by data crawling technology comes from websites or social media platforms and covers all walks of life.

[0016] The training task of incremental pre-training of the base large model focuses on language modeling, that is, allowing the model to predict and generate the next most likely word or word fragment in the sequence based on a given text sequence (i.e. context). This not only enhances the model's ability to process Chinese text, but also enables it to inject rich marketing knowledge into the model by continuously learning the unique vocabulary, sentence patterns and logical structures in the marketing field.

[0017] The specific process of using QLoRA technology combined with data from various industry segments to fine-tune the incremental pre-trained base model and build a LoRA adapter is as follows:

[0018] Load the incrementally pre-trained pedestal language model;

[0019] Decompose the weight matrix of the incrementally pre-trained base large language model into two low-rank matrices to reduce the number of parameters that need to be adjusted;

[0020] Use the processed high-quality data from various industries to fine-tune the incremental pre-trained base large language model, and adjust the parameters of the low-rank matrix to obtain the LoRA adapter corresponding to each industry.

[0021] Each LoRA adapter in each industry is assigned a unique task identifier, which is used to select the corresponding LoRA adapter parameters. The input task identifier determines which LoRA adapter will be loaded by the base model, thereby generating marketing copy for the corresponding industry.

[0022] In the process of fine-tuning the incremental pre-trained base language model: For industries with sufficient training data (such as furniture, beauty, and food industries), SFT (Supervised Fine-Tuning) technology can be directly used to fine-tune the model using a large amount of high-quality data to improve the model's generation capabilities;

[0023] For industries with relatively less training data (such as the film and television industry), in order to solve the problem of insufficient training data, a hybrid strategy of SFT (Supervised Fine-Tuning) technology combined with RLHF (Reinforcement Learning from Human Feedback) technology is adopted. Through model self-generated data, human feedback and reinforcement learning optimization, the amount of training data and model performance are gradually increased, thereby improving the large model's ability to generate high-quality copy for different industries.

[0024] Collection and processing of high-quality data from various industries: Using data crawling technology, we extract texts with high reading volume, likes, favorites, and comments from major social media platforms, treat these texts as potential high-quality texts, and use AI technology to mark marketing needs.

[0025] For industries with less training data, the specific implementation process of combining SFT and RLHF to enhance the amount of training data and model performance is as follows:

[0026] Step 21: Preliminary fine-tuning (SFT): First, use high-scoring and high-quality marketing copy to perform preliminary supervised fine-tuning on the model, so that the model has basic marketing copy generation capabilities;

[0027] Step 22: Reward model training: The reward model is trained based on the human feedback data obtained in the above data collection phase, so that it can score the generated copy;

[0028] Step 23: Reinforcement learning optimization: Use reinforcement learning algorithms to optimize the copywriting generation model for a specific industry based on the score of the reward model;

[0029] Step 24: Use the reward model to score the generated new copy. The higher the score, the better the quality of the copy.

[0030] Step 25: Calculate the reward value of each generated copy based on the score of the reward model, and then use a reinforcement learning algorithm (such as PPO) to optimize the strategy, so that the generated copy can get a higher reward by optimizing the model parameters;

[0031] Step 26, Data Enhancement: By generating new high-quality copywriting and collecting human feedback, the amount of training data is gradually increased to further improve the model performance. Through cyclic iteration, the model is gradually optimized so that the copywriting it generates increasingly meets the expected standards, thereby obtaining the LoRA adapter for the corresponding industry.

[0032] The call of the marketing vertical model integrates LoRA adapters of multiple industries to achieve simultaneous deployment and call of models of different industries to meet the specific marketing needs of different sub-industries. Specifically, it includes the following processes:

[0033] Marketing vertical model deployment: Deploy the combined marketing vertical model within the enterprise or in the cloud to ensure its efficient operation;

[0034] API interface: Develop and deploy API interfaces to facilitate external systems to call specific models of different industries to generate marketing copy.

[0035] A device for building a marketing vertical model based on a large language model, including a processor, a memory, and a user terminal device;

[0036] The user-end device is connected to the processor via an API interface, and the processor and the memory are integrated on a bus.

[0037] A storage medium for building a marketing vertical model based on a large language model. The storage medium is a hardware device or virtual device used to store model data.

[0038] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0039] 1. Through the construction method of the marketing vertical model, the present invention significantly improves the performance of the large language model in marketing copy generation, especially the application effect in segmented industries; a large amount of marketing knowledge is injected into the incremental pre-training process of the large language model, so that the model performs better in Chinese processing and marketing copy generation.

[0040] 2. Combining QLoRA technology and RLHF technology to fine-tune the large model further improves the accuracy and practicality of the model in specific industries.

[0041] 3. The fine-tuned base language model is inserted into LoRA adapters of different industries to obtain the marketing vertical model, which realizes the deployment and calling of models of different industries, meets the specific needs of segmented industries, and improves the efficiency and quality of marketing copy generation.

[0042] 4. Use RLSF technology to continuously track the published copywriting, use reinforcement learning to adjust the strategy, improve the large model's ability to generate text, and make the output of the large model more inclined to popular copywriting. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the construction process of the marketing vertical model;

[0044] Figure 2 for Figure 1 Get the LoRA adapter process diagram;

[0045] Figure 3 A schematic diagram of the localized deployment process for marketing vertical models;

[0046] Figure 4 Schematic diagram of the device for building a marketing vertical model.

[0047] In the figure: 10 bus, 11 processor, 12 user terminal device, 13 memory, 14 API interface. DETAILED DESCRIPTION

[0048] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The preferred embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thoroughly understood.

[0049] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation method.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0051] Embodiment 1

[0052] See also Figure 1 The method of building a marketing vertical model based on a large language model includes the following processes:

[0053] Step 1: Establish the base model: Use an open source model as the base model for pre-training;

[0054] Step 2: Build a corpus: Use data crawling technology to crawl a large amount of high-quality marketing-related copywriting data from the Internet, and clean and pre-process these data to obtain a marketing copywriting corpus;

[0055] Step 3, incremental pre-training: Use the massive marketing copy collected in the corpus to perform incremental pre-training of all parameters of the base model to obtain the base model after incremental pre-training (the training task of incremental pre-training of the base model focuses on language modeling, that is, let the model predict and generate the next most likely word or word fragment in the sequence based on a given text sequence (i.e. context). This not only enhances the model's ability to process Chinese text, but also enables the model to inject rich marketing knowledge into the model by continuously learning the unique vocabulary, sentence patterns and logical structures in the marketing field);

[0056] Step 4. Build LoRA adapters for different industries: Based on the incremental pre-trained base model, use QLoRA (Quantization and Low-Rank Adaptation) technology to fine-tune the incremental pre-trained base model with the data of each segmented industry (during the fine-tuning process, the parameters of the base model are frozen, and only the parameters of the LoRA adapter are trainable), and each industry gets a LoRA adapter (LoRA adaptor); for example, through QLoRA technology, a LoRA adapter specifically for the home appliance industry is generated, so that the model can more accurately describe product characteristics when generating copywriting for the home appliance industry;

[0057] Step 5: Build a marketing vertical model: Connect the base model fine-tuned according to the data of each sub-industry with the LoRA adapters of different industries to obtain a marketing vertical model;

[0058] Step 6: Use of the marketing vertical model: Integrate LoRA adapters from multiple different industries to form a unified and flexible structure, so that it can realize task processing in different industries by loading each LoRA layer separately.

[0059] For open source large models, you can choose llama3, which can process text data in multiple languages ​​and is suitable for advanced applications such as synthetic data generation and model refinement.

[0060] The marketing copy data crawled by data crawling technology comes from websites or social media platforms (such as Weibo, WeChat public accounts, Zhihu, Xiaohongshu, etc.) and covers all walks of life.

[0061] Collection and processing of high-quality data from various industries: social media data crawling, using data crawling technology to extract copywriting with high reading volume, likes, collections, and comments from major social media platforms, and regard these copywriting as potential high-quality copywriting. AI technology is used to mark marketing needs, classify copywriting, and distinguish high-quality copywriting from low-quality copywriting;

[0062] By collecting and processing data from various industries, the high quality and industry relevance of the data are ensured, so that the constructed marketing vertical model can provide insightful and accurate copywriting services in the marketing field. Taking the home appliance industry as an example, the home appliance copywriting data accumulated by the collection platform and the high-interaction copywriting extracted from social media are stored in the database after manual or AI annotation for subsequent use.

[0063] The specific process of using QLoRA technology combined with data from various industry segments to fine-tune the incremental pre-trained base model and build a LoRA adapter is as follows:

[0064] Step 11: Load the base large language model that has been incrementally pre-trained;

[0065] Step 12: Decompose the weight matrix of the incremental pre-trained base large language model into two low-rank matrices to reduce the number of parameters that need to be adjusted;

[0066] Step 13: Use the processed high-quality data from various industries to fine-tune the incremental pre-trained base large language model, and obtain the LoRA adapter corresponding to each industry by adjusting the parameters of the low-rank matrix.

[0067] Each LoRA adapter in each industry is assigned a unique task identifier, which is used to select the corresponding LoRA adapter parameters. The input task identifier determines which LoRA adapter will be loaded into the base model to generate marketing copy for the corresponding industry. For example, when a user issues a copy generation request for the beauty industry, the system first identifies the task identifier and extracts the beauty industry LoRA adapter corresponding to the task identifier, and then dynamically loads the adapter parameters into the base model to generate beauty industry copy that meets user needs.

[0068] In the process of fine-tuning the incremental pre-trained base language model: For industries with sufficient training data (such as furniture, beauty, and food industries), SFT (supervised fine-tuning) technology can be directly used to fine-tune the model using a large amount of high-quality data to improve the model's generation capabilities;

[0069] For industries with relatively less training data (such as the film and television industry), in order to solve the problem of insufficient training data, a hybrid strategy of SFT (supervised fine-tuning) technology combined with RLHF (human feedback reinforcement learning) technology is adopted. Through model self-generated data, human feedback and reinforcement learning optimization, the amount of training data and model performance are gradually increased, thereby improving the large model's ability to generate high-quality copy for different industries.

[0070] See also Figure 2 For industries with less training data, the specific implementation process of combining SFT and RLHF to enhance the amount of training data and model performance is as follows:

[0071] Step 21: Preliminary fine-tuning (SFT): First, use high-scoring and high-quality marketing copy to perform preliminary supervised fine-tuning on the model, so that the model has basic marketing copy generation capabilities;

[0072] Step 22: Reward model training: The reward model is trained based on the human feedback data obtained in the above data collection phase, so that it can score the generated copy;

[0073] Step 23: Reinforcement learning optimization: Use reinforcement learning algorithms to optimize the copywriting generation model for a specific industry based on the score of the reward model;

[0074] Step 24: Use the reward model to score the generated new copy. The higher the score, the better the quality of the copy.

[0075] Step 25: Calculate the reward value of each generated copy based on the score of the reward model, and then use a reinforcement learning algorithm (such as PPO) to optimize the strategy, so that the generated copy can get a higher reward by optimizing the model parameters;

[0076] Step 26, Data Enhancement: By generating new high-quality copywriting and collecting human feedback, the amount of training data is gradually increased to further improve the model performance. Through cyclic iteration, the model is gradually optimized so that the copywriting it generates increasingly meets the expected standards, thereby obtaining the LoRA adapter for the corresponding industry.

[0077] Among them, the role of the reward model is to use the initially fine-tuned model to generate industry-specific marketing copy. The reward model will score the generated copy, and the score reflects the quality and relevance of the copy.

[0078] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: Through the method of constructing a marketing vertical model, the present invention significantly improves the performance of the large language model in the generation of marketing copy, especially the application effect in segmented industries. A large amount of marketing knowledge is injected into the incremental pre-training process of the large language model, so that the model performs better in Chinese processing and marketing copy generation. The large model is fine-tuned in combination with QLoRA technology and RLHF technology, which further improves the accuracy and practicality of the model in specific industries. The deployment and calling of the marketing vertical model realizes the specific needs of segmented industries and improves the efficiency and quality of marketing copy generation.

[0079] Embodiment 2

[0080] See also Figure 3 By inserting the fine-tuned base language model into the LoRA adapters of different industries, the marketing vertical model is obtained, and the deployment and calling of models of different industries are realized. For example, the marketing model of the home appliance industry is obtained by combining the base language model with the LoRA adapter of the home appliance industry, so that the model can generate more professional home appliance marketing copy;

[0081] By integrating LoRA adapters from multiple industries, the simultaneous deployment and calling of models from different industries can be achieved to meet the specific marketing needs of different industry segments;

[0082] The deployment and calling of LoRA adapters in multiple industries include the following processes:

[0083] Step 31: Marketing vertical model deployment: Deploy the combined marketing vertical model within the enterprise or in the cloud to ensure its efficient operation;

[0084] Step 32, API interface: Develop and deploy API interfaces to facilitate external systems to call specific models of different industries to generate marketing copy.

[0085] The adapter parameters of multiple LoRA models are combined to form a unified and flexible structure, which enables task processing in different industries by loading each LoRA layer separately. It has the following characteristics:

[0086] Integrate multiple fine-tuning tasks, and the LoRA adapters corresponding to each fine-tuning task are centrally managed during deployment. By integrating these adapters, the system can share the same basic model, thus saving resources;

[0087] Efficient use of resources, since each LoRA adapter is only loaded when needed, it avoids the situation where all models reside in memory at the same time, which can save GPU memory and computing resources;

[0088] Permission management: Through the separate loading mechanism of adapters, different access rights can be assigned to different tasks to ensure that sensitive or task-specific data and models are not leaked in a shared resource environment;

[0089] Dynamic loading: LoRA adapters for each industry can be dynamically loaded into the base model as needed. This process is simplified by using tools such as the Peft library, which allows the required adapters to be quickly loaded via the API interface.

[0090] Multi-task support, by introducing the above-mentioned dynamic loading mechanism, allows a single base model to quickly switch between different tasks.

[0091] After the marketing vertical model generates marketing copy for the corresponding industry and puts it on the market, the RLSF (Reinforcement Learning with Social Media Feedback) technology is used to continuously track the published copy, and the strategy is adjusted using reinforcement learning to improve the large model's copy generation ability;

[0092] The RLSF (Reinforcement Learning with Social Media Feedback) technology dynamically adjusts the marketing copy delivered, including the following process:

[0093] Step 41: Determine the mission objectives

[0094] Task goal: Optimize the effect of marketing copy generation so that its performance on social media (such as clicks, likes, collections, and comments) is closer to that of popular copy;

[0095] Reward function: The reward function is set based on key indicators such as click-through rate, user interaction (likes, shares, views), etc., and higher rewards will be given to copywriting with high interaction and high exposure.

[0096] Step 42: Data collection and processing

[0097] Collect data: Collect feedback data of marketing copy that has been delivered through the ContaAI platform from social media platforms (such as Xiaohongshu), including clicks, likes, collections, comments, etc.

[0098] Data processing: Clean and format the collected data for further analysis and model training.

[0099] Step 43: Feedback signal generation

[0100] Convert metrics to rewards: Convert social media metrics into reward signals for reinforcement learning; for example, a post with high click-through rates can receive positive rewards, and a post with many comments and active user interactions can receive higher rewards.

[0101] Step 44: Reinforcement Learning Model Training

[0102] Model selection: Choose a suitable RL (reinforcement learning) algorithm, such as the Deep Q Network (DQN) to handle high-dimensional copywriting problems;

[0103] Environmental interaction: After the model generates the copy, it is placed on social media and the strategy is adjusted based on feedback signals (such as clicks, likes, etc.);

[0104] Strategy update: Use the collected reward signals to update the strategy so that the model gradually learns to generate more popular copy.

[0105] Step 45: Real-time optimization

[0106] Continuous learning: Continuously collect new social media feedback and continuously update the model to respond to changes in market trends and user preferences.

[0107] Step 46: Evaluate and monitor

[0108] Evaluation: Use indicators such as user interaction rate, click-through rate, etc. to evaluate the effectiveness of the copy;

[0109] Monitoring: Monitor copywriting performance, especially pay attention to abnormal changes or shifts in trends, and adjust models or strategies when necessary.

[0110] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: by using RLSF (reinforcement learning with social media feedback) technology, the copywriting generation capability of the large model can be continuously learned and improved in actual operations, not only based on historical data, but also able to adjust the generation strategy according to the latest feedback, thereby ensuring that the generated copywriting can better attract the target audience, meet market demand, and increase user engagement on social media platforms.

[0111] Embodiment 3

[0112] See also Figure 4 , a device for building a marketing vertical model based on a large language model, comprising a processor 11, a memory 13 and a user terminal device 12;

[0113] The client device 12 is connected to the processor 11 via an API interface 14 . The processor 11 and the memory 13 are integrated on the bus 10 .

[0114] The user terminal device 12 transmits the processing information to the processor 11 through the API interface 14, selects the LoRA adapter of its own industry and generates the marketing copy of the industry; the copy can be stored and backed up through the memory 13.

[0115] Embodiment 4

[0116] A storage medium for building a marketing vertical model based on a large language model. The storage medium is a hardware device or virtual device used to store model data.

[0117] The hardware storage device can be HDD (hard disk drive) or SDD (solid state drive);

[0118] Memory storage devices can be RAM (random access memory) or ROM (read-only memory);

[0119] Cloud storage services include AWS S3 (Amazon Simple Storage Service), which is an object storage service provided by Amazon Cloud Service (AWS), Google Cloud Storage, which is an object storage service provided by Google Cloud, and Azure Blob Storage, which is an object storage service provided by Microsoft Azure Cloud Service.

[0120] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0121] The LoRA adapter in this application is LoRA (Low-Rank Adaptation), which is a method for efficiently adjusting pre-trained models through low-rank decomposition. The specific steps include first training a large language model using a large-scale general data set, then performing low-rank decomposition on the model's weight matrix, and inserting the low-rank matrix to fine-tune the model parameters. This method achieves fine-tuning by inserting a low-rank matrix. Compared with directly adjusting all parameters, fewer parameters are required, and it can quickly adapt to new tasks or fields, reducing training time and resources;

[0122] QLoRA (Quantized Low-Rank Adaptation) is a variant of LoRA that combines quantization technology to make low-rank decomposition and parameter adjustment more efficient. The specific steps include low-rank decomposition of the model and then quantization of the inserted low-rank matrix to reduce accuracy and save computing and storage resources. QLoRA is suitable for deployment on resource-limited devices, such as mobile devices or embedded systems. It can deploy efficient pre-trained models in low-resource environments and adapt to specific tasks such as real-time translation or speech recognition.

[0123] This application uses data crawling technology to crawl copywriting data in strict compliance with national laws and regulations, respect cyberspace sovereignty and network security, and maintain a good network order.

Claims

1. A method for constructing a marketing vertical model based on a large language model, characterized by: Including the following process, Establishment of the base model: Select an open source model as the base model for pre-training; Constructing a corpus: Using data crawling to crawl a large amount of high-quality copywriting data related to marketing from the Internet, and then cleaning and preprocessing these data to obtain a marketing copywriting corpus; the marketing copywriting data crawled by data crawling comes from websites or social media platforms and covers all walks of life; Incremental pre-training: Using the massive marketing copy collected in the corpus, perform incremental pre-training of all parameters of the base model to obtain the base model after incremental pre-training; Build LoRA adapters for different industries: Based on the incremental pre-trained base model, QLoRA technology is used to fine-tune the incremental pre-trained base model based on the data of each subdivided industry, and each industry gets a LoRA adapter; Build a marketing vertical model: Connect the base model fine-tuned according to the data of each sub-industry with the LoRA adapters of different industries to obtain a marketing vertical model; Use of the marketing vertical model: Integrate LoRA adapters from multiple different industries to form a unified and flexible structure, enabling it to implement task processing in different industries by loading each LoRA layer separately.

2. The method for constructing a marketing vertical model based on a large language model according to claim 1, characterized in that: The training task of incremental pre-training of the base large model focuses on language modeling, that is, allowing the model to predict and generate the next most likely word or word fragment in the sequence based on a given text sequence. This not only enhances the model's ability to process Chinese text, but also enables it to inject rich marketing knowledge into the model by continuously learning the unique vocabulary, sentence patterns and logical structures in the marketing field.

3. The method for constructing a marketing vertical model based on a large language model according to claim 1, characterized in that: The specific process of using QLoRA technology combined with data from various industry segments to fine-tune the incremental pre-trained base model and build a LoRA adapter is as follows: Load the incrementally pre-trained pedestal language model; Decompose the weight matrix of the incrementally pre-trained base large language model into two low-rank matrices to reduce the number of parameters that need to be adjusted; Use the processed high-quality data from various industries to fine-tune the incremental pre-trained base large language model, and adjust the parameters of the low-rank matrix to obtain the LoRA adapter corresponding to each industry.

4. The method for constructing a marketing vertical model based on a large language model according to claim 3, characterized in that: Each LoRA adapter in each industry is assigned a unique task identifier, which is used to select the corresponding LoRA adapter parameters. The input task identifier determines which LoRA adapter will be loaded by the base model, thereby generating marketing copy for the corresponding industry.

5. The method for constructing a marketing vertical model based on a large language model according to claim 1, characterized in that: In the process of fine-tuning the incremental pre-trained base language model, for industries with sufficient training data, SFT technology is directly used to fine-tune the model using a large amount of high-quality data to improve the model's generation capabilities; For industries with relatively less training data, a hybrid strategy of combining SFT technology with RLHF technology is adopted. Through model self-generated data, human feedback and reinforcement learning optimization, the amount of training data and model performance are gradually increased, thereby improving the large model's ability to generate high-quality copy for different industries.

6. The method for constructing a marketing vertical model based on a large language model according to claim 5, characterized in that: The specific implementation process of combining SFT and RLHF to enhance the amount of training data and model performance is as follows: Step 21: Preliminary fine-tuning (SFT): First, use high-scoring and high-quality marketing copy to perform preliminary supervised fine-tuning on the model, so that the model has basic marketing copy generation capabilities; Step 22: Reward model training: The reward model is trained based on the human feedback data obtained in the above data collection phase, so that it can score the generated copy; Step 23: Reinforcement learning optimization: Use reinforcement learning algorithms to optimize the copywriting generation model for a specific industry based on the score of the reward model; Step 24: Use the reward model to score the generated new copy. The higher the score, the better the quality of the copy. Step 25: Calculate the reward value of each generated copy based on the score of the reward model, and then use the reinforcement learning algorithm to optimize the strategy, so that the generated copy can get a higher reward by optimizing the model parameters; Step 26, Data Enhancement: By generating new high-quality copywriting and collecting human feedback, the amount of training data is gradually increased to further improve the model performance. Through cyclic iteration, the model is gradually optimized so that the copywriting it generates increasingly meets the expected standards, thereby obtaining the LoRA adapter for the corresponding industry.

7. The method for constructing a marketing vertical model based on a large language model according to claim 1, characterized in that: The deployment and calling of LoRA adapters in multiple industries include the following processes: Step 31: Marketing vertical model deployment: Deploy the combined marketing vertical model within the enterprise or in the cloud to ensure its efficient operation; Step 32, API interface: Develop and deploy API interfaces to facilitate external systems to call specific models of different industries to generate marketing copy.

8. The method for constructing a marketing vertical model based on a large language model according to claim 7, characterized in that: The specific process of using RLSF technology to dynamically adjust the marketing copy includes the following: Determine mission objectives Task goal: Optimize the effect of marketing copy generation so that its performance on social media (such as clicks, likes, collections, and comments) is closer to that of popular copy; Reward function: The reward function is set based on key indicators such as click-through rate, user interaction (likes, shares, views), etc., and higher rewards will be given to copywriting with high interaction and high exposure; Data Collection and Processing Collect data: Collect feedback data of marketing copy that has been delivered through the ContaAI platform from social media platforms (such as Xiaohongshu), including clicks, likes, collections, comments, etc. Data processing: Clean and format the collected data for further analysis and model training; Feedback signal generation Metrics to Rewards: Converting social media metrics into reward signals for reinforcement learning; For example, a copy with high click-through rates can receive positive rewards, and a copy with many comments and active user interactions can receive higher rewards; Reinforcement Learning Model Training Model selection: Choose a suitable RL algorithm, such as a deep Q-network, to handle high-dimensional copywriting problems; Environmental interaction: After the model generates the copy, it is placed on social media and the strategy is adjusted based on the feedback signals; Strategy update: Use the collected reward signals to update the strategy so that the model can gradually learn to generate more popular copywriting; Real-time optimization Continuous learning: constantly collect new social media feedback and continuously update the model to respond to changes in market trends and user preferences; Evaluation and Monitoring Evaluation: Use indicators such as user views, click-through rate, etc. to evaluate the effectiveness of the copy; Monitoring: Monitor copywriting performance, especially pay attention to abnormal changes or shifts in trends, and adjust models or strategies when necessary.

9. The method for constructing a marketing vertical model based on a large language model according to claim 1, characterized in that: Collection and processing of high-quality data from various industries: Using data crawling technology, we extract texts with high reading volume, likes, collections, and comments from major social media platforms. We regard these texts as potential high-quality texts, and use AI technology to mark marketing needs, classify the texts, and distinguish between high-quality and low-quality texts.

10. The method for constructing a marketing vertical model based on a large language model according to claim 1, characterized in that: During the fine-tuning process of the incremental pre-trained base model on the data of each industry segment, the parameters of the base model are frozen and only the parameters of the LoRA adapter are trainable.

Images