A method, system, device, and storage medium for meta-information-driven synthesis training of large-scale traditional Chinese painting models.

Abstract

This invention relates to the field of artificial intelligence technology, specifically a method, system, device, and storage medium for meta-information-driven synthesis training of large-scale Chinese painting models. The method is executed by a computer device and includes the following steps: Step S1: Data acquisition and parsing; Step S2: Meta-information-driven structured data synthesis; Step S3: Multi-level quality filtering; Step S4: Progressive supervised fine-tuning; Step S5: Reinforcement learning based on a generative reward model. The main objective of this invention is to provide an integrated method and system for meta-information-driven synthesis training of large-scale Chinese painting models, capable of generating high-quality professional data at low cost, effectively suppressing hallucinations, and possessing honest expression capabilities.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system, device, and storage medium for meta-information-driven synthesis training of large-scale traditional Chinese painting models. Background Technology

[0002] Traditional Chinese painting (Guohua) carries a profound historical and cultural heritage, and its appreciation involves complex art historical knowledge, a unique understanding of techniques, and a high level of aesthetic cultivation. With the development of multimodal large model (VLM) technology, it has become possible to use AI to assist in the appreciation of Guohua. However, existing technologies face serious challenges in this vertical field:

[0003] First, high-quality professional data is extremely scarce and expensive to acquire; professional annotation in the field of traditional Chinese painting requires a very high level of expertise, and traditional manual annotation methods are inefficient and cannot meet the data requirements for large-scale model training.

[0004] Secondly, the existing general-purpose VLM has a particularly serious "illusion" problem when processing Chinese painting content; the model often fabricates incorrect dynasty attribution, author information or technique descriptions out of thin air, and even describes elements that do not exist in the picture out of thin air (such as fabricating modern buildings in landscape paintings).

[0005] Furthermore, existing models generally lack the ability to express uncertainty (overconfidence); when faced with issues such as unclear seals, disputed ownership of paintings, or questions outside the boundaries of knowledge, existing models tend to give a seemingly certain but actually incorrect answer, and such misinformation is very likely to mislead users in professional fields.

[0006] Finally, existing training processes typically separate "data synthesis" from "model training," making it impossible to dynamically adjust data strategies based on the model's real-time performance. Furthermore, existing RLHF (Reinforcement Learning Based on Human Feedback) lacks evaluation design tailored to the professional dimensions of traditional Chinese painting. Summary of the Invention

[0007] In response to the problems of the extreme scarcity and high cost of acquiring high-quality professional data and the serious "illusion" problem in existing technologies, the main purpose of this invention is to provide a meta-information-driven synthesis training integrated method and system for large-scale Chinese painting models, which can generate high-quality professional data at low cost, effectively suppress illusions and have the ability to express truthfully.

[0008] To achieve the above objectives, the core technical solution adopted by this invention is as follows: To achieve the above objectives, this invention provides a meta-information-driven synthesis training method for large-scale models of traditional Chinese paintings. The method is executed by a computer device and includes the following steps:

[0009] Step S1: Data Acquisition and Analysis

[0010] The original dataset of Chinese paintings is obtained through a network interface. The original dataset includes high-resolution digital images and corresponding structured metadata. The structured metadata includes at least basic information, artistic information, physical information and cultural information.

[0011] Step S2: Meta-information-driven structured data synthesis

[0012] Based on structured meta-information, image-text pair training data is generated using a pre-defined synthesis strategy; the image-text pair training data includes deterministic data and uncertain expression data.

[0013] Deterministic data includes factual questions and answers generated based on explicit meta-information, descriptive content combined with image features, and professional analysis based on known techniques;

[0014] Uncertainty-based expression data includes information-deficient expressions generated for fields with missing meta-information, controversial and questionable expressions generated for points of academic debate, and knowledge boundary expressions generated for areas not covered by the knowledge graph.

[0015] Step S3: Multi-stage quality filtration

[0016] The synthesized image-text pair training data is input into the multi-layer filtering module, which sequentially performs language quality filtering, factual accuracy filtering, professionalism filtering, and balance filtering to output a high-quality training set.

[0017] Step S4: Progressive Supervised Fine-Tuning (SFT)

[0018] The pre-trained multimodal large model is progressively fine-tuned in three stages using a high-quality training set, successively completing basic fact learning, professional expression training, and uncertainty expression training to obtain the SFT base model.

[0019] Step S5: Reinforcement Learning Based on Generative Reward Model

[0020] We construct a dual-model architecture that includes a question generator model and an answer generator model, and use a generative reward model (GenRM) as an evaluator for joint optimization training.

[0021] GenRM scores the model output based on four dimensions: illusion detection, relevance assessment, usefulness assessment, and honesty assessment, and feeds the scores back to the dual-model architecture to update the model parameters.

[0022] Furthermore, in step S2, the process of generating the uncertainty expression data specifically includes:

[0023] Detect null fields in structured metadata and generate a text description indicating "existing data cannot be determined" by calling the missing information template;

[0024] If multiple conflicting attributions are found for the same work when searching related art history knowledge bases, the controversial template is used to generate a text description indicating that "scholars have different opinions";

[0025] Identify non-core entities or undefined concepts in structured meta-information, and use knowledge boundary templates to generate text descriptions indicating that "further in-depth professional research is needed".

[0026] Furthermore, in step S3, the specific execution process of the multi-layer filtering module is as follows:

[0027] Data containing grammatical errors is removed using a syntax checker and a fluency evaluation model.

[0028] The entity information in the data is compared with authoritative art history knowledge bases to eliminate data with factual errors;

[0029] The accuracy of terminology usage in the data is detected using a professionalism assessment model, and data with a professionalism level below a preset threshold is removed.

[0030] The ratio of statistically deterministic data to uncertain data is determined by sampling or weighting the data according to a preset balance factor to ensure a balance in the types of training data.

[0031] Furthermore, in step S4, the three-stage progressive fine-tuning specifically includes:

[0032] First sub-phase: Freeze the model's visual encoder parameters and update the language model parameters only using deterministic data containing basic historical facts and common sense;

[0033] The second sub-stage involves training the model using deterministic data containing art terminology and technique analysis to enhance its understanding of professional concepts in traditional Chinese painting.

[0034] The third sub-stage involves training the model using data representing uncertainty and adjusting its output probability distribution so that it can output text sequences representing uncertainty when faced with unknown or controversial information.

[0035] Furthermore, in step S5, the input to the generative reward model (GenRM) includes the current image, the generated question, the generated answer, and the corresponding original structured meta-information;

[0036] GenRM uses the original structured meta-information as a ground truth benchmark to calculate the illusion detection score of the generated answer in terms of factual accuracy;

[0037] GenRM evaluates the information entropy density contained in the generated answers and calculates a usefulness assessment score.

[0038] GenRM detects whether the generated answers contain uncertain words that are consistent with the facts and calculates an honesty assessment score.

[0039] Furthermore, the method also includes self-guidance mechanisms:

[0040] Based on the current capability assessment results of the SFT base model, guidance instructions are automatically generated.

[0041] The guidance includes adaptive tasks to maintain current capabilities, targeted tasks generated to address identified knowledge weaknesses, and progressively more complex tasks that enhance current capabilities.

[0042] Furthermore, a meta-information-driven synthesis training integrated system for large-scale traditional Chinese painting models is provided, the system comprising:

[0043] The data acquisition unit is used to acquire high-definition digital images and structured metadata of traditional Chinese paintings.

[0044] The data synthesis unit is used to generate training data for image-text pairs containing deterministic data and uncertain expression data based on structured meta-information;

[0045] The quality filtering unit is used to filter text-image pairs training data through a four-layer filtering mechanism: language quality, factual accuracy, professionalism, and balance.

[0046] The supervised fine-tuning unit is used to perform progressive training on the model using filtered data, which includes three sub-stages: basic facts, professional expressions, and expressions of uncertainty.

[0047] The reinforcement learning unit includes a question generator model, an answer generator model, and a generative reward model; the generative reward model is used to optimize the output of the answer generator model based on multidimensional evaluation metrics.

[0048] Furthermore, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a meta-information-driven synthesis training method for a large-scale Chinese painting model that achieves any of the above.

[0049] Furthermore, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements any of the above-mentioned meta-information driven synthesis training methods for a large-scale Chinese painting model.

[0050] Compared with the prior art, the beneficial effects of this invention are significant and multifaceted:

[0051] 1. Reduced costs and improved efficiency: The use of meta-information-driven synthesis technology replaces expensive manual annotation, significantly reducing data acquisition costs and improving training efficiency.

[0052] 2. Reduce illusions and improve accuracy: By introducing GenRM's four-dimensional evaluation mechanism, especially illusion detection and honesty assessment, the illusion rate of the model is reduced, ensuring the accuracy of the art historical facts in the output content.

[0053] 3. Enhance the ability to express uncertainty: “Uncertainty expression data” was synthesized and specially trained, enabling the model to learn to honestly express “don’t know” or “doubt” when facing knowledge boundaries, avoiding misleading results caused by overconfidence.

[0054] 4. Improved the depth of integration: It realized a closed-loop iteration of data generation and model training, and achieved a gradual improvement in model capabilities through the entry of a self-guided mechanism. Attached Figure Description

[0055] To enable those skilled in the art to more clearly and comprehensively understand the technical solutions of the present invention, preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the accompanying drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the accompanying drawings:

[0056] Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention;

[0057] Figure 2 This is a schematic diagram of the data synthesis and filtering process in Embodiment 1 of the present invention;

[0058] Figure 3 This is a schematic diagram of the progressive supervised fine-tuning strategy of Embodiment 1 of the present invention;

[0059] Figure 4 This is a schematic diagram of the reinforcement learning dual-model architecture of Embodiment 1 of the present invention;

[0060] Figure 5 This is a system hardware structure block diagram of Embodiment 1 of the present invention. Detailed Implementation

[0061] The present invention will now be described in detail with reference to the accompanying drawings:

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, several preferred embodiments of the invention will be described in detail below with reference to the accompanying drawings. It should be understood that the embodiments described herein are merely for explaining the invention and do not constitute any limitation on its scope of protection. Any modifications, equivalent substitutions, or improvements made based on the spirit and principles of this invention should be included within the scope of protection of this invention. Example

[0063] like Figure 1As shown, this embodiment provides a meta-information-driven synthesis and training integrated method for large-scale models of traditional Chinese paintings. This method is executed by a computer device (such as a high-performance server cluster) containing a processor and memory. The specific steps are as follows:

[0064] Step S101: Obtain the dataset of traditional Chinese paintings

[0065] The processor obtains raw data from authoritative data sources (such as the Palace Museum's digital platform and the database of Chinese paintings throughout history) via a network interface.

[0066] The data consists of:

[0067] 1. High-definition digital images: Scanned copies of traditional Chinese paintings with a resolution of no less than 1080P.

[0068] 2. Structured metadata data: contains the following explicitly defined fields:

[0069] 1. Basic information: title of the work, author's name, and dynasty of creation (e.g., "Song" or "Yuan").

[0070] 2. Art Information: Painting School (e.g., "Wu School"), Subject Matter (e.g., "Landscape" and "Flower and Bird"), Techniques (e.g., "Boneless" and "Axe-Cut Texture").

[0071] 3. Physical information: size of the artwork, materials used (e.g., "silk", "paper"), and current holding institution.

[0072] 4. Cultural Information: Inscription content (OCR-recognized text), seal information (seal text and attribution), and historical background description.

[0073] Step S102: Perform structured data synthesis based on metadata data

[0074] The processor uses a limited amount of seed data and employs a strategy that combines template generation and variant expansion to synthesize training data.

[0075] The template generation employs an automated data synthesis method driven by structured metadata. The system maintains a standardized JSON database containing basic information about the paintings. Each record includes structured fields such as artist, creation date, painting type, dimensions, collecting institution, technical characteristics, and subject matter. These metadata fields serve as atomic filling units for template instantiation, and are automatically assembled using predefined syntax templates.

[0076] The variant extension employs a data generation strategy based on multi-dimensional reconstruction of source data and instruction optimization. Technically, it leverages a structured Prompt (prompt word) to guide a large language model in generating question-answer pairs. The system features a four-layer progressive Prompt architecture: the task overview layer defines the role as a "Chinese painting appreciation question generation expert"; the input specification layer includes core parameters such as the original question, content nature, question style, painting information, and appreciation text; the task requirements layer constructs detailed constraints to ensure the naturalness of the questions and the quality of the answers; and the output format layer defines a standardized structural template. Precise control and data diversity are achieved through the cross-product combination of answer content nature and question style, with each combination configuring specific question type examples and generation points.

[0077] The synthesized data is divided into two categories:

[0078] Type A: Deterministic Data

[0079] 1. Factual Question and Answer: Generated based on "basic information". For example, input the metadata "Author: Wang Ximeng", and generate the question and answer pair "Q: Who is the author of this painting? A: This painting was created by Wang Ximeng, a painter from the Northern Song Dynasty".

[0080] 2. Descriptive content: Combining image features with "artistic information". For example, combining the "green landscape" tag, a description can be generated such as "The image uses green colors, and the texture of the rocks is delicate..."

[0081] 3. Professional Analysis: Based on "known techniques". For example, for the tag "axe-cut texture strokes", generate analytical text about the specific application of this technique in the painting.

[0082] Type B: Data expressing uncertainty

[0083] This is one of the core innovations of this invention, which aims to train the honesty of the model.

[0084] 1. Information Missing Expression: When a field in the metadata (such as "year of creation") is empty or marked as "unknown", the synthesis system generates the following text: "The specific year of creation of this painting cannot be determined by existing data."

[0085] 2. Controversial and questionable expression: When there are different records in the related knowledge base regarding the ownership of the same work, the following is generated: "Regarding the ownership of this painting, there is still controversy in the academic community. One theory is that..., and another theory is that...."

[0086] 3. Knowledge Boundary-Type Expression: For details not covered by meta-information (such as the identity of a blurry background figure in a painting), generate: "The specific identity of this person requires more in-depth professional verification and is currently inconclusive."

[0087] Step S103: Perform multi-level quality filtration

[0088] The processor executes a four-layer filtering mechanism to ensure data quality:

[0089] 1. Language quality filtering: Use grammar checkers (such as rule-based parsers) and fluency assessment models (such as models with perplexity PPL below a threshold) to filter data with grammatical errors or incoherent sentences.

[0090] 2. Fact accuracy filtering: Extract entities (person names, place names, dates) from the generated text, compare them with authoritative art history knowledge bases using triplets, and eliminate data with factual conflicts.

[0091] 3. Professionalism Filtering: The LLM-as-a-Judge method is used to score the professionalism of the samples.

[0092] A professional filtering approach using LLM-as-a-Judge is employed, utilizing paid multimodal APIs such as GPT-4V and Claude-3 to comprehensively assess the quality of the Chinese painting dataset. This method employs a carefully designed prompt-guided model acting as a "Chinese painting data cleaning expert," comprehensively evaluating the data across four dimensions: painting content analysis, image quality detection, information annotation verification, and illusion identification in the appreciation text. The model first analyzes the main elements, compositional style, and technical characteristics of the paintings. Then, it detects image quality issues such as incomplete paintings, blurriness, and atypical Chinese painting styles. It verifies the accuracy of annotation information such as the artist's dynasty and, most importantly, identifies illusionary content in the appreciation text that does not match the actual painting. The API returns a structured JSON evaluation report, including the problem type, severity, and suggested corrections. Subsequent processing is based on the severity of the detection results: data with low-severity problems are directly retained; medium-severity problems are transferred to a manual correction process for targeted improvement; and data with high-severity problems are directly filtered and deleted, ensuring the professionalism and accuracy of the final corpus.

[0093] 4. Balance Filtering: The ratio of statistical deterministic data to uncertain data is maintained within a preset range (e.g., 8:2) through oversampling or undersampling techniques to prevent the model from forgetting this ability due to insufficient data of one type.

[0094] Step S104: Perform Supervised Fine-Tuning (SFT) Training

[0095] A progressive strategy is adopted, and the model is trained in three sub-stages:

[0096] 1. First sub-stage (basic fact learning): The focus is on training the model using deterministic data containing basic historical facts, with the goal of reducing basic illusion errors (such as getting the dynasty wrong).

[0097] 2. Second sub-stage (professional expression training): Using data containing complex artistic terms and technique analysis, strengthen the model's correct use of professional concepts such as "cunfa" and "qiyun".

[0098] 3. Third Sub-stage (Uncertainty Statement Training): This stage focuses on training using Type B data. The goal of this stage is to adjust the model's confidence output so that it can generate a high-probability "reject answer" or "uncertainty statement" token when encountering out-of-distribution (OOD) or missing information, rather than fabricating answers.

[0099] Step S105: Self-guided reinforcement learning training

[0100] Construct a dual-model architecture for joint optimization:

[0101] Model components:

[0102] 1. Question Generator Model 51: Generates diverse questions based on images and meta-information.

[0103] 2. Solver Model 52: This is the large-scale Chinese painting model to be optimized, which generates answers to the questions.

[0104] The Questioner-Solver framework employs a dual-model interaction mechanism. The objective function of the question generator (Questioner) is E_{x~π_i, q_pred~π_q(·|x)} [R_q(x, q_pred)], while the objective function of the solver (Solver) is E_{x~π_i, y_pred~π_s(·|x,q)} [R_s(x, q, y_pred)]. The Solver's reward, Rs(x, q, y) = GenRM(x, q, y), is evaluated by the generation reward model based on the quality of the answer. The Questioner's reward, Rq(x, q) = { 1 if 0 < |{yi : yi = ymaj}| < N, 0 otherwise}, is designed based on moderate question difficulty. A reward is given when the number of mode answers is neither all nor zero, ensuring that the generated questions have a reasonable level of challenge.

[0105] The training process employs an alternating optimization strategy. In the first phase, the Questioner is fixed, and the GRPO algorithm is used to optimize the Solver, maximizing the GenRM reward through multi-sample generation and group advantage calculation. In the second phase, the Solver is fixed again, and the Policy Gradient Algorithm is used to optimize the Questioner, aiming to generate questions of moderate difficulty for the Solver. This adversarial training mechanism allows the Questioner to adaptively generate questions that match the Solver's current capabilities, while the Solver continuously improves its answer quality. The two models achieve spiral collaborative optimization in a game-like process, avoiding overfitting issues that may occur with static datasets and ensuring that the difficulty of the questions always matches the model's capabilities during training.

[0106] Generative Reward Model 53 (GenRM) and Four-Dimensional Evaluation:

[0107] GenRM is specifically trained for scoring Chinese painting appreciation texts, focusing on evaluating four core dimensions: illusion detection (checking whether entities in the answer conflict with meta-information), relevance assessment (checking whether the answer directly answers the question and whether it is irrelevant), usefulness assessment (calculating the information density of the answer), and honesty assessment (for questions involving uncertainty, checking whether the model correctly uses uncertainty expressions rather than forcing answers). Training data is obtained by sampling Chinese painting appreciation texts from large models such as GPT and Claude. A multi-dimensional evaluation principle encompassing illusion, relevance, usefulness, and honesty is established. Annotators with art history backgrounds are recruited to score the combination of (principle description, appreciation question, and multiple candidate answers) across the four dimensions, ultimately culminating in an overall score of 1-5. The model is based on the RewardAnything GenRM architecture and undergoes a warm start, receiving input (evaluation principle, appreciation question, candidate answer group, and painting meta-information) and outputting a scalar reward score for each candidate answer. During training, the annotators' ratings of 1-5 points are used as supervision signals, and the rating accuracy is used as the reward function. The GRPO (Group Relative Policy Optimization) algorithm with a thinking mechanism is used for training, enabling the model to accurately evaluate the professional quality scores of different candidate answers relative to the given principles and questions through the reasoning process.

[0108] The evaluation dimensions include:

[0109] 1. Hallucination detection: Check whether the entities in the answer conflict with the meta-information.

[0110] 2. Relevance assessment: Check whether the answer directly addresses the question and whether it is irrelevant.

[0111] 3. Usefulness assessment: Calculate the information density of the answer.

[0112] 4. Honesty assessment: For questions involving uncertainty, check whether the model correctly uses expressions of uncertainty, rather than forcing answers. Example

[0113] This embodiment describes in detail a meta-information-driven synthesis and training integrated system for large-scale traditional Chinese painting models.

[0114] like Figure 5 As shown, the system includes:

[0115] Data Acquisition Unit 1: Responsible for connecting to API interfaces such as those of the Palace Museum, and downloading image streams and metadata in JSON format.

[0116] Data synthesis unit 2: Internally integrates a deterministic data generator, an uncertain data generator, and a synthesis controller. The synthesis controller is responsible for scheduling the template library and calling the NLP model to perform variant extensions.

[0117] Quality Filtering Unit 3: It integrates a syntax analysis module, a knowledge base comparison module, a professionalism scoring module, and a data balancing module.

[0118] Training module: includes supervised fine-tuning unit 4 and reinforcement learning unit 5.

[0119] The supervised fine-tuning unit 4 is equipped with a staged training controller for sequentially loading different types of training data.

[0120] Reinforcement learning unit 5 integrates the Questioner model, the Solver model, and the GenRM evaluator. The GenRM evaluator uses the weighted sum of the four dimensions mentioned above as the final reward signal to guide the parameter updates of the Solver model. Finally, it should be emphasized that the above embodiments are merely illustrative of the technical concept and preferred implementation of the present invention, and are not intended to exhaustively describe or limit the scope of protection of the present invention. Any person skilled in the art, after understanding the spirit and core technical solutions of the present invention, may make various modifications, equivalent substitutions, or improvements based on the content disclosed in the present invention without departing from the basic principles of the present invention. These obvious variations or substitutions should all be considered to be included within the scope of protection claimed by the present invention.

Claims

1. A meta-information-driven synthesis training method for large-scale models of traditional Chinese paintings, characterized in that, The method is executed by a computer device and includes the following steps: Step S1: Data Acquisition and Analysis The original dataset of Chinese paintings is obtained through a network interface. The original dataset includes high-resolution digital images and corresponding structured metadata. The structured metadata includes at least basic information, artistic information, physical information, and cultural information. Step S2: Meta-information-driven structured data synthesis Based on the structured meta-information, image-text pair training data is generated using a preset synthesis strategy; the image-text pair training data includes deterministic data and uncertain expression data. The deterministic data includes factual questions and answers generated based on explicit meta-information, descriptive content combined with image features, and professional analysis based on known techniques; The uncertainty expression data includes information-deficient expressions generated for fields with missing meta-information, controversial and questionable expressions generated for points of academic controversy, and knowledge boundary expressions generated for areas not covered by the knowledge graph. Step S3: Multi-stage quality filtration The synthesized image-text pair training data is input into a multi-layer filtering module, which sequentially performs language quality filtering, factual accuracy filtering, professionalism filtering, and balance filtering to output a high-quality training set. Step S4: Progressive Supervision and Fine-tuning The pre-trained multimodal large model is progressively fine-tuned in three stages using the high-quality training set, successively completing basic fact learning, professional expression training, and uncertainty expression training to obtain the SFT base model. Step S5: Reinforcement Learning Based on Generative Reward Model We construct a dual-model architecture that includes a question generator model and an answer generator model, and use a generative reward model as an evaluator for joint optimization training. The generative reward model is used to score the model output based on four dimensions: illusion detection, relevance assessment, usefulness assessment, and honesty assessment, and the scores are fed back to the dual-model architecture to update the model parameters. In step S2, the process of generating the uncertainty expression data specifically includes: The null value field in the structured metadata is detected, and the missing information template is called to generate a text description indicating "existing data cannot be determined"; If multiple conflicting attributions are found for the same work when searching related art history knowledge bases, the controversial template is used to generate a text description indicating that "scholars have different opinions"; Identify non-core entities or undefined concepts in the structured meta-information, and call the knowledge boundary template to generate a text description indicating "requires further in-depth professional research"; In step S3, the specific execution process of the multi-layer filtering module is as follows: Data containing grammatical errors is removed using a syntax checker and a fluency evaluation model. The entity information in the data is compared with authoritative art history knowledge bases to eliminate data with factual errors; The accuracy of terminology usage in the data is detected using a professionalism assessment model, and data with a professionalism level below a preset threshold is removed. The ratio of statistically deterministic data to uncertain data is determined by sampling or weighting the data according to a preset balance factor to ensure a balance in the types of training data. In step S4, the three-stage progressive fine-tuning specifically includes: First sub-phase: Freeze the model's visual encoder parameters and update the language model parameters only using deterministic data containing basic historical facts and common sense; The second sub-stage involves training the model using deterministic data containing art terminology and technique analysis to enhance its understanding of professional concepts in traditional Chinese painting. The third sub-stage involves training the model using the aforementioned uncertainty expression data, adjusting the model's output probability distribution so that it outputs a text sequence representing uncertainty when faced with unknown or controversial information.

2. The method for meta-information-driven synthesis training of a large-scale traditional Chinese painting model according to claim 1, characterized in that, In step S5, the input to the generative reward model includes the current image, the generated question, the generated answer, and the corresponding original structured meta-information; The generative reward model uses the original structured meta-information as a truth benchmark to calculate the illusion detection score of the generated answer in terms of factual accuracy. The generative reward model evaluates the information entropy density contained in the generated answer and calculates a usefulness assessment score. The generative reward model detects whether the generated answers contain uncertain words that are consistent with the facts and calculates an honesty assessment score.

3. The method for meta-information-driven synthesis training of a large-scale traditional Chinese painting model according to claim 1, characterized in that, The method also includes a self-guiding mechanism: Based on the current capability assessment results of the SFT base model, guidance instructions are automatically generated; The guidance instructions include adaptive tasks for maintaining current capabilities, targeted tasks generated for identified knowledge weaknesses, and progressively more complex tasks that enhance current capabilities.

4. A meta-information-driven synthesis and training integrated system for large-scale traditional Chinese painting models, characterized in that: The system includes: The data acquisition unit (1) is used to acquire high-definition digital images and structured metadata of traditional Chinese paintings; The data synthesis unit (2) is used to generate image-text pair training data containing deterministic data and uncertain expression data based on the structured meta-information; The quality filtering unit (3) is used to filter the text-image pair training data through a four-layer filtering mechanism of language quality, factual accuracy, professionalism and balance; The supervised fine-tuning unit (4) is used to perform progressive training on the model using the filtered data, which includes three sub-stages: basic facts, professional expressions, and uncertainty expressions. The reinforcement learning unit (5) includes a question generator model (51), an answer generator model (52), and a generative reward model (53); the generative reward model (53) is used to perform feedback optimization on the output of the answer generator model (52) based on multidimensional evaluation metrics; The above system is used to execute the meta-information driven synthesis training method for a large-scale Chinese painting model according to any one of claims 1 to 3.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements a meta-information-driven synthesis training method for a large-scale Chinese painting model according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a meta-information-driven synthesis training method for a large-scale Chinese painting model according to any one of claims 1 to 3.

Images