Method for assisting AIGC creation in combination with large language model
By performing data annotation and reinforcement learning of large language models, the problems of poor interpretability and high training cost are solved, and more efficient and accurate AIGC creation is achieved, and the generated content is more interpretable and quality.
Patent Information
- Application Number
- CN202411478699.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-22
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Large language models have poor interpretability in AIGC creation, high training costs, and deviations and uncertainties in the generated content, making it difficult to understand the logical basis for generating content.
By annotating text data, the interpretability of the model is enhanced, and reinforcement learning is carried out in combination with human feedback, and the large language model is fine-tuned to improve the quality and efficiency of generated content. The steps such as data preprocessing, labeling, fine-tuning and model fusion are used to generate Multimedia works.
It improves the accuracy and interpretability of generated content, reduces the demand for computing resources, saves costs, and improves the quality and efficiency of AIGC creation.
Smart Images

Figure CN119004266A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of large language models, and in particular to a method for assisting AIGC creation by combining a large language model. Background Art
[0002] With the rapid development of artificial intelligence technology, the status of large language models in AIGC (artificial intelligence generated content) creation has become increasingly prominent. In recent years, large language models such as GPT and Wenxinyiyan have continued to emerge, bringing new changes to content creation. With its huge number of parameters and massive training data, large language models can better understand natural language and generate high-quality content in various forms such as text, images, and audio.
[0003] At present, the research status of AIGC creation shows diversified characteristics. On the one hand, many scientific research institutions and enterprises have invested in the research and development of large language models, and continuously improved the performance and generalization ability of the models. For example, OpenAI's GPT series of models perform well in natural language processing tasks and can generate articles with clear logic and fluent language. On the other hand, the application of AIGC in various industries is also constantly expanding. From news writing, advertising creativity to artistic creation, AIGC can be seen.
[0004] However, despite the remarkable achievements of large language models in AIGC creation, there are still some problems and challenges. The main problems are as follows: 1. The model has poor interpretability, and the generated content may have certain deviations and uncertainties, which makes it difficult for creators to accurately understand the logical basis of the generated content during use. For example, when using a large language model to generate an article, the source and reasoning process of some sentences may not be clear.
[0005] 2. The training of large language models requires a lot of computing resources and data. According to statistics, training a large language model may require thousands of high-performance servers and cost millions or even tens of millions of dollars, which is very costly.
[0006] Therefore, it is of great theoretical and practical significance to conduct in-depth research on the application of large language models in AIGC creation and explore more efficient and accurate creation methods. Summary of the invention
[0007] Based on the technical problems existing in the background technology, the present invention proposes a method for assisting AIGC creation by combining a large language model, which improves the interpretability of the model by annotating text data; and improves the quality and efficiency of AIGC creation by fine-tuning and reinforcing learning the large language model.
[0008] The present invention proposes a method for creating AIGC by combining a large language model, comprising the following steps: S1: Data preparation and preprocessing: Capture the characteristics and contextual information of natural language, generate fluent and coherent text, clean the text data, remove repeated, irrelevant information and noise, and then perform word segmentation to split the text into words with independent meanings to ensure the quality and availability of data, in preparation for the subsequent large language model training and use; S2: Data text annotation: By classifying and annotating data text, meaningful information is added to the data text, providing supervision signals for the large language model to help it perform natural language understanding and dialogue management. The annotated data set provides accurate label information for the model, helping the large language model learn the intrinsic connection between data, enhance the interpretability of the large language model, and understand the internal working mechanism of the large language model; S3: Reinforcement learning method based on human feedback: Use human feedback as a reward signal to guide the learning of the large language model, so that the large language model can adapt to different task objectives and continuously improve during the interaction process; S4: Large language model tuning: By adjusting the model's learning rate, batch size, and number of iterations, observe the changes in the model's performance indicators on the validation set, optimize the model's training process, and improve the model's performance and effect on specific tasks; S5: Model fusion and integration: First use a large language model to generate text content, then combine it with the image generation model and use partitioned prompt words to generate corresponding images. Finally, use the audio processing model to generate appropriate audio, integrate the content together to form a complete multimedia work.
[0009] Preferably, in step 2, Construct the dataset to be labeled D={(X i ,Y i )}|D| i=1 , Where X i Represents the entity pairs in the text, Y i represents the relationship between entity pairs, |D| is the total amount of labeled data in the dataset. In order to improve the execution efficiency of the data augmentation method, representative data is selected from the dataset D to be labeled.
[0010] Preferably, the representativeness includes uncertainty and diversity, wherein, Uncertainty uses the minimum confidence level to measure the uncertainty of the data U(X i ), by training an original relation extraction model to predict the relation probability P(X i |Y i ), U(Xi) = 1 - P(X i |Y i ) ; Diversity is measured using cosine similarity. After using sentence-bert to convert the data into vectors, the vector V of each data is calculated. xi The distance between the average vector V of all data, L(X i ) = cos(sim(V Xi , V)); The specific marking steps include: S21, select 8n data with the highest uncertainty from the dataset D to be labeled, calculate the uncertainty of all data, and store the first 4n data with the highest uncertainty into a temporary dataset temp; S22. Select the 2n data with the highest diversity from the 4n data as the final labeled data, calculate the diversity of all data in the temporary data set temp, and store the top n data with the highest diversity into the labeled data set D'.
[0011] Preferably, in S3, the reinforcement learning method specifically includes: S31: supervised fine-tuning phase, building a pre-trained language model containing directive input prompts and expected output results, and fine-tuning the large language model; S32: In the reward model training phase, sample prompts are input into the model to generate a series of output texts, which are evaluated by the annotators to train a reward model that can predict human preference outputs; S33: In the reinforcement learning fine-tuning stage, the output of the reward model is used as the reward signal in the reinforcement learning process to further optimize the model.
[0012] Preferably, in step S31, the specific steps are: Set a pre-trained language model f W :X i →Y i , where W represents the trainable parameters, and is trained on a small amount of labeled data set D' to obtain the teacher model f W tea ,The training process adopts the traditional fine-tuning method or the cue fine-tuning paradigm; Through the teacher model, each unlabeled sample X i ’ Perform pseudo-labeling to generate the probability distribution of each category: Y i ’ =argmax p(y=c|f W tea (X i’ )) Among them, p(.) is the probability distribution, Y i ’ Y i The corresponding pseudo label, f W The student model f can be obtained by training on pseudo-labeled data W stu .
[0013] Preferably, in step S32, the specific steps are: Design a soft reward function, use the mean method to pre-cluster the labeled data and pseudo-labeled data respectively, after obtaining the output text predicted by the reinforcement learning reasoning framework, compare the cluster labels of the output text and the target result respectively on the basis of mean pre-clustering, and take the intersection of the comparison results of the two. The calculation equation of the soft reward function is as follows: r C = Ⅱ{K(d p )}∩Ⅱ{ K(d q )} R C =1(if d p = d q ) or R C =0 Among them, r C represents the intersection of the comparison results, K() is a probability function, called Kolmogorov distribution, which is used to describe the distribution of a random variable, d p Indicates output text, d q represents the target result, R C Indicates the size of the soft reward.
[0014] Preferably, in step S33, the reward signal is graded into reinforcement levels, specifically including: No reinforcement level: When the reward model parameters cause the database to crash, no reinforcement is performed and the reward value is determined directly based on the empirical formula; Low reinforcement level: if the original database performs well, additional rewards are given; if the original data performs poorly, additional penalties are imposed; High reinforcement level: When the reward model parameters meet the optimization requirements, additional reward values are given based on the reward values of the original performance results.
[0015] Preferably, the additional reward value is designed as follows: For the performance indicator c of the reward model, calculate the initial value according to the change and the next time interval measurement value , By integrating the reinforcement function and the reinforcement level, the tuning knowledge is utilized, thereby achieving the purpose of optimizing the model tuning effect. The reinforcement function of each performance indicator c is designed as follows: when ; when .
[0016] The present invention also provides a computing device, comprising: one or more processors; A storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned creation method.
[0017] The present invention also proposes a computer-readable storage medium, in which a program is stored, and when the program is executed by a processor, the above-mentioned creation method is implemented.
[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention improves the interpretability of the model by adding annotations to the data text annotations, making the generated content more accurate; 2. The present invention reduces the amount of computation and saves costs while improving the quality and efficiency of AIGC creation by fine-tuning and reinforcing learning of large language models. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 This is a flow chart of a method for assisting AIGC creation by combining a large language model, as proposed by the present invention. DETAILED DESCRIPTION
[0020] The present invention will be further explained below in conjunction with specific embodiments.
[0021] Reference Figure 1 This embodiment proposes a method for creating AIGC using a large language model, including the following steps: S1: Data preparation and preprocessing: Capture the characteristics and contextual information of natural language, generate fluent and coherent text, clean the text data, remove repeated, irrelevant information and noise, and then perform word segmentation to split the text into words with independent meanings to ensure the quality and availability of data, in preparation for the subsequent large language model training and use; S2: Data text annotation: By classifying and annotating data text, meaningful information is added to the data text, providing supervision signals for the large language model to help it perform natural language understanding and dialogue management. The annotated data set provides accurate label information for the model, helping the large language model learn the intrinsic connection between data, enhance the interpretability of the large language model, and understand the internal working mechanism of the large language model; S3: Reinforcement learning method based on human feedback: Use human feedback as a reward signal to guide the learning of the large language model, so that the large language model can adapt to different task objectives and continuously improve during the interaction process; S4: Large language model tuning: By adjusting the model's learning rate, batch size, and number of iterations, observe the changes in the model's performance indicators on the validation set, optimize the model's training process, and improve the model's performance and effect on specific tasks; S5: Model fusion and integration: First use a large language model to generate text content, then combine it with the image generation model and use partitioned prompt words to generate corresponding images. Finally, use the audio processing model to generate appropriate audio, integrate the content together to form a complete multimedia work.
[0022] Among them, in step 2, Construct the dataset to be labeled D={(X i ,Y i )}|D| i=1 , Where X i Represents the entity pairs in the text, Y i represents the relationship between entity pairs, |D| is the total amount of labeled data in the dataset. In order to improve the execution efficiency of the data augmentation method, representative data is selected from the dataset D to be labeled.
[0023] The representativeness of this embodiment includes uncertainty and diversity, wherein, Uncertainty uses the minimum confidence level to measure the uncertainty of the data U(X i ), by training an original relation extraction model to predict the relation probability P(X i |Y i ), U(Xi) = 1 - P(X i |Y i ) ; Diversity is measured using cosine similarity. After using sentence-bert to convert the data into vectors, the vector V of each data is calculated. xi The distance between the average vector V of all data, L(X i ) = cos(sim(V Xi , V)); The specific marking steps include: S21, select 8n data with the highest uncertainty from the dataset D to be labeled, calculate the uncertainty of all data, and store the first 4n data with the highest uncertainty into a temporary dataset temp; S22. Select the 2n data with the highest diversity from the 4n data as the final labeled data, calculate the diversity of all data in the temporary data set temp, and store the top n data with the highest diversity into the labeled data set D'.
[0024] In step S3, the reinforcement learning method specifically includes: S31: supervised fine-tuning phase, building a pre-trained language model containing directive input prompts and expected output results, and fine-tuning the large language model; S32: In the reward model training phase, sample prompts are input into the model to generate a series of output texts, which are evaluated by the annotators to train a reward model that can predict human preference outputs; S33: In the reinforcement learning fine-tuning stage, the output of the reward model is used as the reward signal in the reinforcement learning process to further optimize the model.
[0025] In step S31, the specific steps are: Set a pre-trained language model f W :X i →Y i , where W represents the trainable parameters, and is trained on a small amount of labeled data set D' to obtain the teacher model f W tea ,The training process adopts the traditional fine-tuning method or the cue fine-tuning paradigm; Through the teacher model, each unlabeled sample X i ’ Perform pseudo-labeling to generate the probability distribution of each category: Y i ’ =argmax p(y=c|f W tea (X i ’ )) Among them, p(.) is the probability distribution, Y i’ Y i The corresponding pseudo label, f W The student model f can be obtained by training on pseudo-labeled data W stu .
[0026] In step S32, the specific steps are: Design a soft reward function, use the mean method to pre-cluster the labeled data and pseudo-labeled data respectively, after obtaining the output text predicted by the reinforcement learning reasoning framework, compare the cluster labels of the output text and the target result respectively on the basis of mean pre-clustering, and take the intersection of the comparison results of the two. The calculation equation of the soft reward function is as follows: r C = Ⅱ{K(d p )}∩Ⅱ{ K(d q )} R C =1(if d p = d q ) or R C =0 Among them, r C represents the intersection of the comparison results, K() is a probability function, called Kolmogorov distribution, which is used to describe the distribution of a random variable, d p Indicates output text, d q represents the target result, R C Indicates the size of the soft reward.
[0027] The reward function of this embodiment improves the traditional binary function. Due to the simple reward calculation logic, the reward matrix obtained by the agent is a sparse matrix containing a large number of 0s and a small number of 1s. This will lead to slow model convergence and poor model learning effect.
[0028] In step S33, the reward signal is graded according to the reinforcement level, which specifically includes: No reinforcement level: When the reward model parameters cause the database to crash, no reinforcement is performed and the reward value is determined directly based on the empirical formula; Low reinforcement level: if the original database performs well, additional rewards are given; if the original data performs poorly, additional penalties are imposed; High reinforcement level: When the reward model parameters meet the optimization requirements, additional reward values are given based on the reward values of the original performance results.
[0029] The design method of the additional reward value is: For the performance indicator c of the reward model, calculate the initial value according to the change and the next time interval measurement value , By integrating the reinforcement function and the reinforcement level, the tuning knowledge is utilized, thereby achieving the purpose of optimizing the model tuning effect. The reinforcement function of each performance indicator c is designed as follows: when ; when .
[0030] In step 5, during the image generation process, the image is divided into regions and controlled by specific text descriptions through partition prompt words. The prompt word content can be adjusted at any time according to different needs. If you are not satisfied with the generated image, you can change the content or style of a certain area of the image by modifying the prompt word. In addition, you can create more diverse image effects by combining different prompt words. The specific process includes: Parameter control, dividing the image area by setting different ratio and position parameters, for example, using parameters like "1:1, 1:2, 1:2" to represent the ratio relationship between the overall area and each sub-area, and position parameters like "0:0, 0:0, 0:1" to determine the starting coordinates of each area; Mask segmentation determines the partitions of the image by drawing masks, which allows for more intuitive and detailed partitioning based on specific image content.
[0031] For example, when drawing characters and building scenes, the latent variable pair plug-in is used to set the segmentation parameters "1:1, 1:2, 1:2" to divide the area, and then use prompt words such as "2girls, upper body, outdoors, blurry backgroundAND 2girls, upper body, outdoors, blurry background, upper body, outdoors, blurry background" to clarify the characteristics and styles of characters in different areas. At the same time, the self-assembled LoRA plug-in can process the LoRA model to make character drawing more diverse. In scene construction, using the partition diffusion plug-in and the partition prompt word plug-in, different areas such as background and foreground can be finely divided like drawing multiple character images. For example, in the partition drawing function of TiledDiffusion, by checking and enabling "Move ControlNet tensor to CPU (if applicable)", setting parameters such as latent space block overlap, the generation of specified content at a specified location can be precisely controlled. For example, you can create a background area that includes the sea, the beach, and the sun, a foreground area that includes a person wearing a white dress, a straw hat, with long black hair, a delicate face, and an elegant figure, and a foreground area that includes a perfect dog on the beach. By setting the partition prompt words reasonably, you can construct a rich and vivid scene.
[0032] Case 1 The method provided in this embodiment is used to generate text content. The following is a simple Python code example showing how to generate a fragment of an article about AIGC.
[0033] First, configure your API key.
[0034] Then, set up your API key: Generating text Next, use to generate a text about AIGC.
[0035] Output example Running the above code will generate a paragraph about the importance of large language models in AIGC, similar to the following: Large Language Models (LLMs) play a key role in AI-generated content (AIGC). Through deep learning of massive data, high-quality, natural and fluent text content can be generated. LLMs are widely used in AIGC, including automatic writing, translation, dialogue systems, and programming assistance. Its core advantage is the ability to understand and generate complex language structures to meet content needs in different scenarios. With the continuous advancement of technology, the position of LLMs in AIGC will become increasingly important, driving content creation into a new intelligent era.
[0036] Case 2 When this method is applied to building an intelligent private domain question-answering assistant, since it is difficult and expensive to train your own data set, the "zero-shot-prompt" method is adopted, that is, "0 samples" do not need to be trained and the model is used directly. However, "zero-shot" also has difficulties. It is necessary to put all the private domain knowledge into the input and hand it over to LLM for processing as context and questions. Since the length of the input text of the model is limited. The emergence of vector databases can solve this problem. Milvus vector database is used to store private domain knowledge, and the Embedding algorithm directly uses the algorithm provided by OpenAI. By converting text knowledge or other data formats into vector data through the Embedding algorithm, and then storing it in the vector database. In this way, the database can find the closest data based on the similarity algorithm and return it to the large language model, and use the open source framework LangChain to quickly build end-to-end applications. LangChain can easily manage interactions with language models, link multiple components together, and integrate additional resources such as APIs and databases. By using LangChain, models and databases can be easily integrated. The intelligent private domain Q&A assistant plays an important role in private domain knowledge Q&A. Within the enterprise, it can quickly answer employees' questions about company policies, product information, etc., improving work efficiency. In the field of customer service, it can provide customers with instant answers and improve customer satisfaction.
[0037] It can answer user questions quickly and accurately, saving users time. For example, when a user asks about the characteristics of a company's products, the intelligent private domain Q&A assistant can immediately give a detailed answer. Through semantic queries of the vector database, it can retrieve results similar to the input content to improve the accuracy of the answer. It can provide personalized answers based on the user's historical conversations and specific needs. As private domain knowledge continues to increase, the capabilities of the Q&A assistant can be easily expanded by simply vectorizing the new knowledge and storing it in the database.
[0038] The intelligent private domain question and answer assistant realized by the large language model assisted creation method provided in this embodiment provides an efficient, accurate and personalized solution for the management and application of private domain knowledge.
[0039] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A method of combining a large language model to assist AIGC creation, It is characterized in that The following steps are involved: S1: Data preparation and preprocessing: Capture the characteristics and contextual information of natural language, generate fluent and coherent text, clean the text data, remove repeated, irrelevant information and noise, and then perform word segmentation to split the text into words with independent meanings to ensure the quality and availability of data, in preparation for the subsequent large language model training and use; S2: Data text annotation: By classifying and annotating data text, meaningful information is added to the data text, providing supervision signals for the large language model to help it perform natural language understanding and dialogue management. The annotated data set provides accurate label information for the model, helping the large language model learn the intrinsic connection between data, enhance the interpretability of the large language model, and understand the internal working mechanism of the large language model; S3: Reinforcement learning method based on human feedback: Use human feedback as a reward signal to guide the learning of the large language model, so that the large language model can adapt to different task objectives and continuously improve during the interaction process; S4: Large language model tuning: By adjusting the model's learning rate, batch size, and number of iterations, observe the changes in the model's performance indicators on the validation set, optimize the model's training process, and improve the model's performance and effect on specific tasks; S5: Model fusion and integration: First use a large language model to generate text content, then combine it with the image generation model and use partitioned prompt words to generate corresponding images. Finally, use the audio processing model to generate appropriate audio, integrate the content together to form a complete multimedia work.
2. According to claim 1, a method for assisting AIGC creation by combining a large language model, It is characterized in that In the step 2, Construct the dataset to be labeled D={(X i ,Y i )}|D| i=1 , Where X i Represents the entity pairs in the text, Y i represents the relationship between entity pairs, |D| is the total amount of labeled data in the dataset. In order to improve the execution efficiency of the data augmentation method, representative data is selected from the dataset D to be labeled.
3. According to claim 2, a method for assisting AIGC creation by combining a large language model, It is characterized in that The representation includes uncertainty and diversity, where Uncertainty uses the minimum confidence level to measure the uncertainty of the data U(X i ), by training an original relation extraction model to predict the relation probability P(X i |Y i ), U(Xi) = 1 - P(X i |Y i ) ; Diversity is measured using cosine similarity. After using sentence-bert to convert the data into vectors, the vector V of each data is calculated. xi The distance between the average vector V of all data, L(X i ) = cos(sim(V Xi , V)); The specific marking steps include: S21, select 8n data with the highest uncertainty from the dataset D to be labeled, calculate the uncertainty of all data, and store the first 4n data with the highest uncertainty into a temporary dataset temp; S22. Select the 2n data with the highest diversity from the 4n data as the final labeled data, calculate the diversity of all data in the temporary data set temp, and store the top n data with the highest diversity into the labeled data set D'.
4. According to claim 1, a method for assisting AIGC creation by combining a large language model, It is characterized in that In S3, the reinforcement learning method specifically includes: S31: supervised fine-tuning phase, building a pre-trained language model containing directive input prompts and expected output results, and fine-tuning the large language model; S32: In the reward model training phase, sample prompts are input into the model to generate a series of output texts, which are evaluated by the annotators to train a reward model that can predict human preference outputs; S33: In the reinforcement learning fine-tuning stage, the output of the reward model is used as the reward signal in the reinforcement learning process to further optimize the model.
5. According to claim 4, a method for assisting AIGC creation by combining a large language model, It is characterized in that In step S31, the specific steps are: Set a pre-trained language model f W :X i →Y i , where W represents the trainable parameters, and is trained on a small amount of labeled data set D' to obtain the teacher model f W tea ,The training process adopts the traditional fine-tuning method or the cue fine-tuning paradigm; Through the teacher model, each unlabeled sample X i ’ Perform pseudo-annotation to generate the probability distribution of each category: Y i ’ =argmax p(y=c|f W tea (X i ’ )) Among them, p(.) is the probability distribution, Y i ’ Y i The corresponding pseudo label, f W The student model f can be obtained by training on pseudo-labeled data W stu .
6. According to claim 4, a method for assisting AIGC creation by combining a large language model, It is characterized in that In step S32, the specific steps are: Design a soft reward function, use the mean method to pre-cluster the labeled data and pseudo-labeled data respectively, after obtaining the output text predicted by the reinforcement learning reasoning framework, compare the cluster labels of the output text and the target result respectively on the basis of mean pre-clustering, and take the intersection of the comparison results of the two. The calculation equation of the soft reward function is as follows: r C =Ⅱ{K(d p )}∩Ⅱ{ K(d q )} R C = 1 (if d p = d q ) or R C = 0 Among them, r C represents the intersection of the comparison results, K() is a probability function, called Kolmogorov distribution, which is used to describe the distribution of a random variable, d p Indicates output text, d q represents the target result, R C Indicates the size of the soft reward.
7. According to claim 4, a method for assisting AIGC creation by combining a large language model, It is characterized in that In step S33, the reward signal is graded according to the reinforcement level, which specifically includes: No reinforcement level: When the reward model parameters cause the database to crash, no reinforcement is performed and the reward value is determined directly based on the empirical formula; Low reinforcement level: if the original database performs well, additional rewards are given; if the original data performs poorly, additional penalties are imposed; High reinforcement level: When the reward model parameters meet the optimization requirements, additional reward values are given based on the reward values of the original performance results.
8. According to claim 7, a method for assisting AIGC creation by combining a large language model, It is characterized in that The design method of the additional reward value is: For the performance indicator c of the reward model, calculate the initial value according to the change and the next time interval measurement value , By integrating the reinforcement function and the reinforcement level, the tuning knowledge is utilized, thereby achieving the purpose of optimizing the model tuning effect. The reinforcement function of each performance indicator c is designed as follows: when ; when 。 9. A computing device, It is characterized in that include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in any one of claims 1 to 8.
10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a program, which implements the method according to any one of claims 1 to 8 when executed by a processor.
Citation Information
Patent Citations
Image processing method and device, equipment and medium
CN117711001A
Programming method, device and equipment based on AIGC and storage medium
CN117992033A
Method and device for generating multimedia product of music and medium
CN118427372A
Text generation model training method and text generation method
CN118468821A
Architecture for explainable reinforcement learning
US20220147876A1
Cited By
Memory enhanced artist agent autonomous learning and creation system and method
CN119294470A
Financial risk assessment method and related equipment
CN119963311A
Training data synthesis method of voice assistant model, voice assistant system and computer equipment
CN120783729A
Zeolite multi-task synthesis auxiliary method and system based on large language model
CN120808923A
