Large model comprehensive evaluation method and system for traditional Chinese painting, electronic equipment and readable storage medium

Abstract

The invention relates to a large-model comprehensive evaluation method and system for a traditional Chinese painting, electronic equipment and a readable storage medium. The method comprises the following steps: a test set construction step, a pure model evaluation step, a retrieval enhancement generation evaluation step, a visual understanding ability evaluation step, a comprehensive evaluation generation step, a comprehensive evaluation report generation step, and a calculation of the similarity between extracted professional elements of the traditional Chinese painting and corresponding elements in reference answers. The invention provides a large model comprehensive evaluation method and system for a traditional Chinese painting, electronic equipment and a readable storage medium, which can perform multi-dimensional and multi-level comprehensive performance evaluation on a model.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and large language model evaluation technology, specifically to a comprehensive evaluation method, system, electronic device, and readable storage medium for large models of traditional Chinese painting. Background Technology

[0002] In recent years, Large Language Models (LLMs) have performed exceptionally well on general natural language processing tasks. However, when these models are applied to vertical fields with profound cultural connotations and high levels of specialization, such as "traditional Chinese painting," existing evaluation methods reveal significant shortcomings.

[0003] First, there is a lack of professionalism and depth. Current mainstream assessments mostly focus on general benchmarks (such as MMLU, GSM8K) or general text metrics (such as BLEU, ROUGE). These metrics cannot effectively assess the accuracy and depth of the model's understanding of professional knowledge in the field of traditional Chinese painting (such as brush and ink techniques, styles and schools, and the creation of artistic conception). Second, the assessment granularity is coarse. Existing methods usually give a general overall score to the model's response, lacking fine-grained differentiation and evaluation of different professional knowledge elements in the response (such as basic information, visual understanding, and semantic reasoning).

[0004] Furthermore, existing technologies lack the ability to utilize external knowledge and fail to combine retrieval-enhanced generation (RAG) technology to assess the model's ability to utilize external professional knowledge bases. This makes it impossible to distinguish whether the model relies on rote memorization or possesses genuine knowledge integration capabilities. Finally, visual evaluation is disconnected from the domain. In visual grounding evaluation, the general object detection metric (mAP) only focuses on object categories (such as "birds" and "flowers"), without considering the unique "technique" elements of traditional Chinese painting (such as "cuncao" and "mogu"), resulting in evaluation results that are out of touch with the actual needs of traditional Chinese painting art analysis.

[0005] Therefore, there is an urgent need in this field for a comprehensive evaluation system and method that is thorough, in-depth, and tailored to the characteristics of traditional Chinese painting. Summary of the Invention

[0006] In view of the shortcomings and problems in the existing technology, the present invention aims to solve the above-mentioned technical problems and provide a comprehensive evaluation system and method for large-scale models of traditional Chinese painting, which can perform multi-dimensional and multi-level comprehensive performance evaluation of the model.

[0007] To achieve the above objectives, the core technical solution adopted by this invention is as follows: This invention provides a comprehensive evaluation method for large-scale models of traditional Chinese paintings, the method being executed by computer equipment and including the following steps: S1: The test set construction step involves receiving traditional Chinese painting images, professional question-and-answer pairs related to traditional Chinese painting images, and reference answers, and constructing ground truth values ​​for object detection that include object category labels and painting technique labels. S2: Pure model evaluation step, directly input the questions from the professional question-and-answer pair into the large model to be evaluated, obtain the first answer, and score the first answer based on the preset question-and-answer ability index system; S3: The Search Enhancement Generation (RAG) evaluation step uses Search Enhancement Generation (RAG) technology to retrieve enhanced context related to the question from the Chinese painting professional knowledge base. The question and enhanced context are concatenated and input into the large model to be evaluated to obtain a second answer. The second answer is scored based on the question-answering ability index system. S4: Visual understanding ability assessment step: Input the Chinese painting image into the large model to be evaluated, obtain the target detection result, compare the target detection result with the target detection ground truth, calculate the mean average precision (mAP), and obtain the visual localization ability score; the comparison step includes: the detection is judged to be correct only when both the object category label and the painting technique label in the target detection result match the target detection ground truth. S5: Comprehensive evaluation generation step. Based on preset weight coefficients, the scores obtained from the pure model evaluation step, the scores obtained from the retrieval enhancement generation evaluation step, and the visual positioning capability score are weighted and summed to generate a comprehensive evaluation report. The question-and-answer ability indicator system includes a first-level overall scoring unit and a second-level professional element scoring unit. The second-level professional element scoring unit performs the following steps: using a predefined hierarchical data structure to extract structured Chinese painting professional elements from the first or second answer, and calculating the similarity between the extracted Chinese painting professional elements and the corresponding elements in the reference answer.

[0008] Furthermore, the predefined hierarchical data structure is in JSON format, and it contains at least the following four levels of traditional Chinese painting professional dimensions: The basic information layer is used to extract elements such as artwork name, size specifications, or collection location; The basic visual understanding layer is used to extract elements such as material shape, composition layout, brushstroke characteristics, or color atmosphere. The semantic understanding layer is used to extract elements such as subject matter, expression, artistic style, atmosphere, or symbolic meaning. The contextual reasoning layer is used to extract information about the painter, the creation date, inscriptions and seals, artistic heritage, historical context, or artistic status.

[0009] Furthermore, the first-level overall scoring unit performs the following processing: Calculate the n-gram-based text similarity between the answer and the reference answer; Calculate the cosine similarity between the vector representation of the answer and the vector representation of the reference answer; The answers and reference answers are input into a large language model for the judge, and the score output by the judge model is obtained.

[0010] Furthermore, the second-level professional element scoring unit undergoes the following processing: The feature extraction big language model is invoked, and professional elements are extracted from the reference answer and the answer of the big model to be evaluated, respectively, based on the hierarchical data structure. For each extracted professional element, calculate its element text similarity value and element vector cosine similarity between the answer and the reference answer; The average element text score is obtained by taking the arithmetic mean of the element text similarity values ​​of all extracted elements. The average feature vector score is obtained by taking the arithmetic mean of the cosine similarity of the feature vectors of all extracted features.

[0011] Furthermore, in step S4, the painting technique label is selected from at least one of the following sets: rubbing technique, boneless technique, double-outline technique, dotting technique, line drawing technique, and shading technique; The calculation of mean precision (mAP) specifically includes: calculating the mean precision based on multiple different intersection-to-union (IoU) thresholds, and calculating the arithmetic mean of the multiple mean precisions.

[0012] Furthermore, step S3 specifically includes: Perform semantic analysis on the question to generate a retrieval query vector; Retrieve a preset number of document fragments with the highest similarity to the query vector in the pre-built vector index of the Chinese painting professional knowledge base; Concatenate the question and the retrieved document fragments into prompt words in the format of: question-enhanced context; Input the prompt words into the large model to be evaluated, and instruct the model to generate a second answer based on the context.

[0013] Furthermore, a system is also provided that includes: The data preprocessing module is used to receive images of traditional Chinese paintings, professional question-and-answer pairs related to the images of traditional Chinese paintings, and reference answers, and to construct ground truth values ​​for object detection that include object category labels and painting technique labels; The pure model evaluation module is used to directly input the questions from the professional question-and-answer pairs into the large model to be evaluated, obtain the first answer, and score the first answer based on the preset question-and-answer ability index system; The Search Enhancement Generation (RAG) evaluation module is used to retrieve enhanced context related to the question from the Chinese painting professional knowledge base using Search Enhancement Generation (RAG) technology. The question and enhanced context are concatenated and input into the large model to be evaluated to obtain a second answer, and the second answer is scored based on the question-answering ability index system. The visual understanding ability assessment module is used to input traditional Chinese painting images into the large model to be evaluated, obtain object detection results, compare the object detection results with the object detection ground truth, calculate the mean average precision (mAP), and obtain a visual localization ability score. Among them, the comparison includes: the detection is judged to be correct only when both the object category label and the painting technique label in the object detection result match the object detection ground truth. The comprehensive evaluation generation module is used to generate a comprehensive evaluation report by weighting and summing the scores obtained by the pure model evaluation module, the retrieval enhancement generation evaluation module, and the visual positioning capability score according to preset weight coefficients. The question-and-answer ability index system includes a first-level overall scoring unit and a second-level professional element scoring unit. The second-level professional element scoring unit is used to extract structured Chinese painting professional elements from the first or second answer using a predefined hierarchical data structure, and to calculate the similarity between the extracted Chinese painting professional elements and the corresponding elements in the reference answer.

[0014] Furthermore, an electronic device is also provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements any of the above-mentioned comprehensive evaluation methods for large-scale models of traditional Chinese paintings.

[0015] 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 aforementioned comprehensive evaluation methods for large-scale models of traditional Chinese paintings.

[0016] Beneficial effects: 1. Enhanced professionalism and depth: By introducing a "layered extraction of professional elements" mechanism, based on a specific four-layer JSON structure, it is possible to delve into the internals of professional knowledge for fine-grained and multi-level evaluation, thereby improving the depth and accuracy of the evaluation and overcoming the superficiality of traditional indicators.

[0017] 2. Enhanced domain relevance: The visual positioning (Grounding) evaluation module is specifically designed for traditional Chinese painting. The "technique" tag matching logic is forcibly added to the mAP evaluation, making the evaluation results of object detection capability more reflective of the model's practical value in the field of art analysis and closely aligned with the actual needs of the domain.

[0018] 3. Improved comprehensiveness and comparability: For the first time, a dual evaluation mode of "pure model evaluation" and "RAG-enhanced model evaluation" is proposed, which can evaluate both the intrinsic knowledge of the model and its ability to utilize external knowledge, making the evaluation dimensions more comprehensive.

[0019] 4. Improved objectivity: The evaluation comprehensively utilizes traditional indicators (BLEU), vector space indicators (EmbeddingPrecision), and LLM referee indicators, with multiple dimensions mutually corroborating each other, making the evaluation results more objective and reliable. Attached Figure Description

[0020] 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: Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the question-answering ability index system in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the hierarchical structure (JSON Schema) of professional elements in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the system architecture of Embodiment 2 of the present invention; Figure 5 This is a schematic diagram of the hardware structure of the electronic device according to Embodiment 3 of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings: 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.

[0022] Example 1: As Figure 1 As shown, this embodiment provides a comprehensive evaluation method for large-scale models of traditional Chinese paintings. This method is executed by computer equipment (such as a server, workstation, or high-performance personal computer). The method includes the following steps: S1: Steps to build the test set The data preprocessing module receives and manages images of traditional Chinese paintings, related questions (Q&A pairs), and reference answers to build the test set required for evaluation.

[0023] Specifically, the data structure of the test set consists of three core parts: 1. Image data: High-resolution digital images of traditional Chinese paintings (e.g., "Dwelling in the Fuchun Mountains").

[0024] 2. Text-based Q&A pairs: Professional questions for each image (e.g., "Please analyze the composition and artistic style of this painting") and reference answers written by experts in the field of traditional Chinese painting.

[0025] 3. Grounding ground value data: Annotated data used for object detection evaluation. Furthermore, the ground value data in this embodiment is specially designed to include not only the location (boundary box) and category (such as "bird", "plum blossom", "rock") of the object in the drawing, but also the drawing technique label associated with the object.

[0026] 1. Examples of technique labels: including but not limited to "texturing and rubbing method" (such as hemp fiber texture, axe-cut texture, raindrop texture), "boneless method", "double-outline method", "dotting and dyeing method", "plain drawing method", "shading method", etc.

[0027] 2. Data structure example: { "object_id": 101, "category": "rock", "bbox": [100, 200, 50,80], "technique": "hemp fiber texture"} S2: Pure Model Evaluation Steps This step is used to assess the level of intrinsic knowledge of the large model being evaluated.

[0028] The computer's processor directly inputs the questions from the test set into the large model to be evaluated, and obtains its generated raw answers (referred to as the first answer). Subsequently, a preset question-and-answer (QA) capability index system is used to score the first answer, obtaining the pure model QA score.

[0029] like Figure 2 As shown, the QA capability indicator system adopts a two-tiered structure of "overall + professional elements": A. First Level: Overall Score A coarse-grained assessment of the overall quality of the model's responses is conducted from three dimensions: 1. Rule-based scoring: Calculate the n-gram text similarity between the model's answer and the reference answer (e.g., using the BLEU algorithm).

[0030] 2. Vector-based scoring: Calculate the cosine similarity between the model's answer and the reference answer's embedding vectors. The processor calls a pre-trained text embedding model to convert both into vectors and calculates the cosine similarity. A preset threshold (e.g., 0.8) is set; if the similarity is greater than this threshold, it is recorded as a match (Precision=1); otherwise, it is recorded as a mismatch (Precision=0), or a continuous similarity value can be used directly as the score.

[0031] 3. Judge Model Scoring: A more powerful large language model (such as a general high-performance generative AI model) is used as the judge. The processor constructs prompt words such as "question", "reference answer" and "model response" as input to the judge model, which is required to score according to criteria such as accuracy and completeness (e.g., 1-10 points).

[0032] B. Second Level: Professional Element Scoring This level aims to provide a fine-grained assessment of the professionalism of the model's responses.

[0033] 1. Feature Extraction: The processor calls a large language model for feature extraction, following a predefined JSON Schema structure (e.g., ...). Figure 3 As shown, structured professional elements are extracted from the "reference answer" and "model answer" respectively.

[0034] In a preferred embodiment, a strict four-layer JSON structure is defined to guide LLM in feature extraction. During extraction, each feature contains two fields: {relevance, content}. The four-layer structure includes: 1. Basic Information Layer: Includes "Artwork Title", "Dimensions", and "Location". 2. Basic Visual Understanding Layer: This includes "material and form", "composition and layout", "brushwork characteristics", and "color atmosphere"; 3. Semantic understanding layer: including "subject content", "form and spirit expression", "artistic style", "atmosphere creation", and "symbolic meaning"; 4. Contextual reasoning layer: including "painter information", "creation date", "inscriptions and seals", "artistic heritage", "historical context", and "artistic status".

[0035] For example, the element of "composition and layout" extracted from the reference answer is: "using a long scroll scattered perspective composition method"; the corresponding element extracted from the model answer is: "long scroll composition, scattered perspective".

[0036] 2. Element Score Calculation: For each extracted professional element, calculate its text similarity (BLEU) and vector similarity (Embedding Precision) with the corresponding element in the reference answer. Finally, take the arithmetic mean to obtain the "average element score".

[0037] S3: Retrieval Enhancement Generation (RAG) Evaluation Steps This step is used to evaluate the model's ability to utilize external knowledge bases.

[0038] 1. Retrieval: The processor performs semantic analysis on the input question, generates a query vector, and retrieves a preset number (Top-K) of document fragments most relevant to the question from a pre-built Chinese painting professional knowledge base (including painting theory, historical materials, technique analysis, etc.) (enhanced context).

[0039] 2. Concatenation and Generation: Concatenate the "question" with the retrieved "enhanced context" to form a new prompt word, input it into the large model to be evaluated, and obtain its generated enhanced answer (referred to as the second answer).

[0040] 3. Scoring: The QA capability indicator system (including the overall score of layer A and the professional element score of layer B) from step S2 above is called again to score the second answer and obtain the RAG enhanced QA score.

[0041] By comparing the scores of S2 and S3, the extent to which RAG technology improves model performance can be intuitively quantified.

[0042] S4: Visual Comprehension Assessment Steps (Grounding) This step evaluates the model's ability to detect objects in traditional Chinese paintings, particularly its ability to identify "techniques".

[0043] 1. Input and Detection: Input the Chinese painting image into the large model to be evaluated (which must have visual capabilities) and obtain the target detection results. The detection results include bounding box coordinates, category labels, and technique labels.

[0044] 2. Comparison and calculation: The detection results are compared with the ground truth values ​​containing technique information constructed in S1.

[0045] Furthermore, in this embodiment, a strict matching logic is used to determine whether the detection is correct: it is only considered a True Positive if both the object category label (e.g., "rock") and the painting technique label (e.g., "hemp fiber texture") of the detection result are consistent with the true value, and the intersection-over-union (IoU) of the bounding box is greater than a threshold. If the technique label is incorrect, even if the positions overlap, it is considered a detection error.

[0046] In this embodiment, the mAP@[.5:.95] standard is adopted. That is, the Intersection over Union (IoU) threshold is calculated in the range of 0.5 to 0.95, and the value is taken in increments of 0.05 (0.5, 0.55, ..., 0.95). The average precision (AP) at each threshold is calculated, and finally the average of these APs is calculated.

[0047] S5: Comprehensive Evaluation Generation Steps The comprehensive evaluation module aggregates and weights the scoring results from the above modules.

[0048] Furthermore, in this embodiment, in order to generate the final comprehensive evaluation report, the system uses a weighted summation method to calculate the total score.

[0049] Let the pure model QA score be S_{pure}, the RAG-enhanced QA score be S_{rag}, and the grounding ability score be S_{grounding}.

[0050] The preset weighting coefficients are W_1, W_2, and W_3.

[0051] Total Score = W_1 \times S_{pure} + W_2 \times S_{rag} + W_3 \times S_{grounding}.

[0052] In one specific embodiment, to emphasize the model's basic visual recognition and knowledge enhancement capabilities, W_1=0.3, W_2=0.3, and W_3=0.4 can be set.

[0053] Finally, a comprehensive evaluation report is generated, which includes radar charts, score comparison tables, and detailed analysis text, clearly showing the model's strengths and weaknesses in pure knowledge, knowledge enhancement, and visual foundation.

[0054] Example 2: Figure 4 As shown, this embodiment provides a comprehensive evaluation system for large-scale models of traditional Chinese paintings. The system includes: Data preprocessing module: Used to receive and manage Chinese painting images, related questions (Q&A pairs) and reference answers, and to build the test set required for evaluation.

[0055] Pure model evaluation module: This module is used to directly input the questions from the test set into the large model to be evaluated, obtain the first answer, and call the QA indicator system calculation module to score it.

[0056] The retrieval enhancement generation evaluation module is used to retrieve the enhanced context related to the question from the Chinese painting professional knowledge base, and input the question and enhanced context into the large model to obtain a second answer, and call the QA indicator system calculation module to score it.

[0057] QA indicator system calculation module: core calculation unit, including overall scoring unit (calculating BLEU, EmbeddingPrecision, LLM judge score) and professional element scoring unit (extracting elements based on JSON Schema and calculating average BLEU / Precision).

[0058] Visual understanding ability assessment module: evaluates the model's target detection ability on Chinese painting images. The core is to compare the detection results containing "technique labels" with the ground truth and calculate mAP@[.5:.95].

[0059] Comprehensive evaluation generation module: This module summarizes and weights the scoring results from the above modules to generate the final comprehensive evaluation report.

[0060] Example 3: Figure 5 As shown, this embodiment provides an electronic device, including a memory and a processor.

[0061] Memory is used to store computer programs. The processor is connected to memory via a bus and is used to execute computer programs.

[0062] When the processor executes the computer program, it implements the various steps of the comprehensive evaluation method for large-scale traditional Chinese painting, as described in Example 1.

[0063] The processor can be a general-purpose processor (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices. The memory can be volatile or non-volatile memory, used to store instructions and data.

[0064] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A comprehensive evaluation method for large-scale models of traditional Chinese painting, characterized in that, The method is executed by a computer device and includes the following steps: The test set construction step involves receiving traditional Chinese painting images, professional question-and-answer pairs related to the traditional Chinese painting images, and reference answers, and constructing a ground truth for object detection that includes object category labels and painting technique labels; The pure model evaluation step involves directly inputting the questions from the professional question-and-answer pairs into the large model to be evaluated, obtaining the first answer, and scoring the first answer based on a preset question-and-answer ability index system. The retrieval enhancement generation evaluation step involves using retrieval enhancement generation technology to retrieve enhanced context related to the question from the Chinese painting professional knowledge base, concatenating the question with the enhanced context and inputting it into the large model to be evaluated to obtain a second answer, and scoring the second answer based on the question-answering ability index system. The visual understanding ability assessment steps involve inputting the traditional Chinese painting image into the large model to be evaluated, obtaining the target detection result, comparing the target detection result with the target detection ground truth, calculating the average accuracy, and obtaining a visual positioning ability score; wherein, the comparison includes: determining that the detection is correct only when both the object category label and the painting technique label in the target detection result match the target detection ground truth; The comprehensive evaluation generation step involves weighting and summing the scores obtained from the pure model evaluation step, the retrieval enhancement generation evaluation step, and the visual positioning capability score according to preset weight coefficients to generate a comprehensive evaluation report. The question-and-answer ability index system includes a first-level overall scoring unit and a second-level professional element scoring unit. The second-level professional element scoring unit performs the following steps: extracting structured Chinese painting professional elements from the first answer or the second answer using a predefined hierarchical data structure, and calculating the similarity between the extracted Chinese painting professional elements and the corresponding elements in the reference answer.

2. The comprehensive evaluation method for large-scale models of traditional Chinese painting according to claim 1, characterized in that, The predefined hierarchical data structure is in JSON format and contains at least the following four levels of professional dimensions related to traditional Chinese painting: The basic information layer is used to extract elements such as artwork name, size specifications, or collection location; The basic visual understanding layer is used to extract elements such as material shape, composition layout, brushstroke characteristics, or color atmosphere. The semantic understanding layer is used to extract elements such as subject matter, expression, artistic style, atmosphere, or symbolic meaning. The contextual reasoning layer is used to extract information about the painter, the creation date, inscriptions and seals, artistic heritage, historical context, or artistic status.

3. The comprehensive evaluation method for large-scale models of traditional Chinese painting according to claim 1, characterized in that, The first-level overall scoring unit performs the following processing: Calculate the n-gram-based text similarity between the answer and the reference answer; Calculate the cosine similarity between the vector representation of the answer and the vector representation of the reference answer; Input the answer and the reference answer into a large language model that acts as a judge, and obtain the score output by the large language model that acts as a judge.

4. The comprehensive evaluation method for large-scale models of traditional Chinese painting according to claim 1, characterized in that, The second-level professional element scoring unit performs the following processing: The feature extraction big language model is invoked, and professional elements are extracted from the reference answer and the answer of the big model to be evaluated, respectively, based on the hierarchical data structure. For each extracted professional element, calculate its element text similarity value and element vector cosine similarity between the answer and the reference answer; The average element text score is obtained by taking the arithmetic mean of the element text similarity values ​​of all extracted elements. The average feature vector score is obtained by taking the arithmetic mean of the cosine similarity of the feature vectors of all extracted features.

5. The comprehensive evaluation method for large-scale models of traditional Chinese painting according to claim 1, characterized in that, In the visual comprehension assessment step, the painting technique label is selected from at least one of the following: rubbing technique, boneless technique, double-outline technique, dotting technique, line drawing technique, and shading technique; The calculation of the average precision specifically includes: calculating the average precision based on multiple different cross-union ratio thresholds, and calculating the arithmetic mean of the multiple average precisions.

6. The comprehensive evaluation method for large-scale models of traditional Chinese painting according to claim 1, characterized in that, The retrieval enhancement generation evaluation step specifically includes: Perform semantic analysis on the problem to generate a retrieval query vector; Retrieve a predetermined number of document fragments with the highest similarity to the retrieval query vector from the vector index of a pre-constructed Chinese painting professional knowledge base; The question and the retrieved document fragments are concatenated into prompt words in the format of question-enhanced context; The prompt words are input into the large model to be evaluated, instructing the model to generate the second answer based on the context.

7. A comprehensive evaluation system for large-scale models of traditional Chinese painting, characterized in that, The system includes: The data preprocessing module is used to receive traditional Chinese painting images, professional question-and-answer pairs related to the traditional Chinese painting images, and reference answers, and to construct target detection ground truth values ​​containing object category labels and painting technique labels; The pure model evaluation module is used to directly input the questions in the professional question-and-answer pairs into the large model to be evaluated, obtain the first answer, and score the first answer based on the preset question-and-answer ability index system. The retrieval enhancement generation evaluation module is used to retrieve enhanced context related to the question from the Chinese painting professional knowledge base using retrieval enhancement generation technology, concatenate the question and the enhanced context and input them into the large model to be evaluated to obtain a second answer, and score the second answer based on the question answering ability index system; The visual understanding ability assessment module is used to input the traditional Chinese painting image into the large model to be evaluated, obtain the target detection result, compare the target detection result with the target detection ground truth, calculate the average accuracy, and obtain a visual positioning ability score; wherein, the comparison includes: the detection is judged to be correct only when both the object category label and the painting technique label in the target detection result match the target detection ground truth; The comprehensive evaluation generation module is used to perform a weighted summation of the scores obtained by the pure model evaluation module, the scores obtained by the retrieval enhancement generation evaluation module, and the visual positioning capability scores according to preset weight coefficients, and generate a comprehensive evaluation report. The question-and-answer ability index system includes a first-level overall scoring unit and a second-level professional element scoring unit. The second-level professional element scoring unit is used to extract structured Chinese painting professional elements from the first answer or the second answer using a predefined hierarchical data structure, and to calculate the similarity between the extracted Chinese painting professional elements and the corresponding elements in the reference answer.

8. An electronic 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 comprehensive evaluation method for large-scale models of traditional Chinese paintings as described in any one of claims 1 to 6.

9. 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 comprehensive evaluation method for large-scale models of traditional Chinese paintings as described in any one of claims 1 to 6.

Images