A system and method for evaluating fine-grained bias of a text-to-image model

Abstract

The application belongs to the technical field of artificial intelligence, and aims at the technical problems of existing text-to-image model bias evaluation, such as static limitation of dimension coverage, insufficient fine-grained description, limitation of large-scale capability caused by artificial dependence, lack of structured association between evaluation samples and dimensions and subgroups, difficulty in quantifying and tracing bias results, and provides a text-to-image model fine-grained bias evaluation system and method, a parsing module parses the evaluation task input information into a structured task object, an extension module determines a fine-grained bias dimension set and a sub-group grouping rule, a construction module generates a corresponding evaluation prompt set, a generation module calls a to-be-tested text-to-image model to sample and generate image samples, and forms an evaluation sample library indexed by dimensions and subgroups, a quantification module performs multi-modal analysis and attribute feature extraction on the image samples, calculates bias quantification indicators, and obtains bias evaluation results. The application can improve the consistency, fine-grained coverage capability and result interpretability of the evaluation process.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a fine-grained bias assessment system and method for text-based graph models. Background Technology

[0002] With the development of generative artificial intelligence technology, text-based graph models are widely used in content creation, design assistance, advertising generation, and multimedia production. Their generated results may exhibit biases related to social attributes such as gender, race, age, occupation, and region. Due to the rapid iteration of text-based graph models and their plugins, the variety of configuration combinations, and the open application contexts, the triggering conditions and manifestations of bias are more complex. Traditional evaluation methods focusing on a few macro-level dimensions are insufficient to promptly cover bias manifestations in subdivided subgroups, long-tail dimensions, and cross-dimensional dimensions.

[0003] Existing bias assessment methods typically rely on manual design of prompts, sample selection, and annotation and review. Some methods also incorporate natural language models to construct static assessment sets or fixed prompt templates around pre-defined macro-social dimensions. However, existing bias assessment methods still suffer from problems such as high manual costs, long update cycles, insufficient capacity for expanding assessment dimensions, inconsistent subgroup grouping criteria, lack of structured indexing between image samples and assessment dimensions, and difficulty in fine-grained quantification and backtracking of bias assessment results. These issues make it difficult to support large-scale and continuous bias assessment of text-based image models in open scenarios. Furthermore, in further governance scenarios, the assessment results obtained by existing bias assessment methods often remain at the reporting level, lacking a direct connection with the text-based image model alignment mechanism, making it difficult to form a reusable and iterative closed-loop governance system. Summary of the Invention

[0004] This application addresses the technical problems of existing bias assessment methods for Wensheng graph models, which suffer from limitations in scalability due to static dimensional coverage, insufficient fine-grained characterization, and reliance on manual intervention. These problems include a lack of structured relationships between assessment samples, dimensions, and subgroups, and difficulties in quantifying and backtracking bias assessment results. The application provides a fine-grained bias assessment system and method for Wensheng graph models.

[0005] To achieve the above objectives, this application adopts the following technical solution: In the first aspect, this application proposes a fine-grained bias assessment system for text-based graph models, including a parsing module, an extension module, a construction module, a generation module, and a quantification module; The parsing module is used to receive the assessment task input information and parse it to obtain a structured task object used to drive the fine-grained bias assessment process. The extended module is used to determine a set of fine-grained bias dimensions based on the structured task object, and to determine the subgroup grouping rules corresponding to the set of fine-grained bias dimensions. The construction module is used to construct a set of evaluation prompts for calling the text-to-image model under test, based on the set of fine-grained bias dimensions and the subgroup grouping rules. The generation module is used to call the text-to-image model to be tested to sample according to the evaluation prompt set, generate image samples, and form an evaluation sample library indexed by fine-grained bias dimension and subgroup; The quantization module is used to obtain image samples from the evaluation sample library, as well as the fine-grained bias dimension index and subgroup index corresponding to the image samples, perform multimodal analysis on the obtained image samples, extract attribute features related to the fine-grained bias dimension, and calculate the bias quantification index according to the subgroup grouping rules and the structured task object to obtain the bias evaluation results of each fine-grained bias dimension.

[0006] Secondly, this application proposes a fine-grained bias assessment method for text-based graph models, including: Receive the assessment task input information and parse it to obtain the structured task object used to drive the fine-grained bias assessment process; Based on the structured task object, determine the set of fine-grained bias dimensions, and determine the subgroup grouping rules corresponding to the set of fine-grained bias dimensions; Based on the fine-grained bias dimension set and the subgroup grouping rules, construct a set of evaluation prompts for calling the text-to-image model to be tested; The text-to-image model to be tested is invoked to sample according to the set of evaluation prompts, generating image samples and forming an evaluation sample library indexed by fine-grained bias dimensions and subgroups; Image samples are obtained from the evaluation sample library, along with the fine-grained bias dimension index and subgroup index corresponding to the image samples. Multimodal analysis is performed on the obtained image samples to extract attribute features related to the fine-grained bias dimension. Bias quantification index is calculated according to the subgroup grouping rules and the structured task object to obtain the bias evaluation results for each fine-grained bias dimension.

[0007] Thirdly, this application provides a computer device, including: a processor and a computer-readable storage medium; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the above-described fine-grained bias assessment method for text-based graph models.

[0008] Fourthly, this application proposes a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed the above-described fine-grained bias assessment method for text-based graph models.

[0009] Compared with the prior art, this application has the following beneficial effects: This application proposes a fine-grained bias assessment system for text graph models, comprising a parsing module, an extension module, a construction module, a generation module, and a quantification module. This system establishes a continuous, automated processing chain for bias assessment of text graph models, from task input, dimension determination, prompt construction, sample generation to bias quantification. The parsing module parses the assessment task input information into structured task objects, enabling subsequent modules to operate based on a unified task object. This reduces inconsistencies in understanding the assessment objectives, model objects, and constraints across different processing stages, improving the executability and consistency of the assessment process. The extension module determines the set of fine-grained bias dimensions based on the structured task objects and establishes corresponding subgroup grouping rules. This allows the system to move beyond a few pre-defined macro-bias dimensions and develop more granular and context-specific assessment dimensions and grouping criteria tailored to the specific assessment task. This enhances the system's ability to cover biases in subgroups, implicit biases, and differential biases between different subgroups. The construction module constructs an assessment prompt set based on the fine-grained bias dimension set and subgroup grouping rules, establishing a correspondence between the assessment prompts and the fine-grained bias dimensions and subgroup groups, thereby improving the relevance, comparability, and traceability of the assessment samples. The generation module calls the text-based image model to be tested to sample according to the assessment prompt set, forming an assessment sample library indexed by the fine-grained bias dimensions and subgroups. This creates a structured index relationship between image samples, assessment prompts, fine-grained bias dimensions, and subgroups, supporting batch and large-scale sample generation and management, reducing the cost of manually constructing samples and manually organizing assessment data. The quantification module obtains image samples and their corresponding fine-grained bias dimension indices and subgroup indices from the assessment sample library. It performs multimodal analysis on the obtained image samples and extracts attribute features related to the fine-grained bias dimensions. Then, according to the subgroup grouping rules and combined with the structured task objects, it calculates the bias quantification index, enabling the visual attributes, semantic representation, and subgroup differences in the text-based image model output to be transformed into calculable bias assessment results. This application enables the automation, structuring, and quantification of the bias assessment process for Wensheng graph models. While reducing reliance on manual prompt design, manual sample preparation, and manual annotation, it improves the fine-grained bias dimension coverage, the consistency between sample generation and statistics, and the interpretability and verifiability of bias assessment results. This alleviates the problems of statically limited dimension coverage, insufficient fine-grained characterization, and limited scalability of existing Wensheng graph model bias assessment methods.

[0010] This application also proposes a method for evaluating fine-grained bias in text-based graph models, an electronic device, and a computer-readable storage medium, which possess all the advantages of the aforementioned fine-grained bias evaluation system for text-based graph models. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of a fine-grained bias assessment system for the graph model in this application. Figure 2 This is a flowchart illustrating a method for evaluating fine-grained bias in the graph model presented in this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] Generative artificial intelligence has developed rapidly in the field of digital content production in recent years. Among them, text-based image models, as generative models that can automatically generate image content based on natural language text prompts, have been widely used in various scenarios. Users only need to input descriptions such as people, scenes, styles, and compositions, and the model can generate corresponding images, thereby significantly reducing the threshold for image creation and improving content production efficiency. With the development of large-scale pre-trained models, diffusion models, and multimodal learning techniques, the generation quality, semantic understanding ability, and style transfer ability of text-based image models have been continuously enhanced, and more and more models and their plugins are being integrated into design platforms, content management systems, and enterprise-level production processes. However, text-based image models usually rely on massive amounts of text and image data for training. These training data come from a wide range of sources, vary in quality, and may contain factors such as uneven distribution of groups, stereotypes, historical biases, and contextual ambiguities that already exist in real society. This may cause the model to unintentionally replicate or amplify biases related to social attributes such as gender, race, age, occupation, region, and cultural identity in the generated results. Therefore, with the large-scale application of text-based graph models, identifying, assessing, and managing the risks of bias in the generated results has become an important technical direction for building model security, content compliance, and trustworthy artificial intelligence.

[0015] In practical applications, the bias problem of text-based image models does not always appear in an explicit and easily identifiable way. Instead, it may be hidden in the generation details such as character appearance, occupational association, skin color presentation, age distribution, clothing style, scene setting, or cultural symbols. Due to the randomness and diversity of text-based image models, the same prompt word may generate significantly different images under different random seeds, different plugins, different style parameters, or different model versions, and the bias performance may also change accordingly. This makes bias assessment unable to rely solely on a small number of samples or a single generation result, but requires statistical analysis of a sufficient number of prompts, samples, and generated images. At the same time, user prompts in real business scenarios are usually more open and complex, and may contain multiple social attributes, aesthetic styles, cultural contexts, and task goals, resulting in bias triggering conditions exhibiting combined, implicit, and long-tail characteristics. Although the traditional method of relying on manually constructing prompts, manually selecting samples, and manually annotating images has a certain degree of accuracy, it is prone to problems such as insufficient evaluation sample coverage, long annotation cycle, high manual cost, difficulty in reproducing results, and insufficient consistency across scenarios when facing large-scale models, multiple version iterations, and multi-scenario deployments. It is difficult to meet the needs of continuous and large-scale bias assessment.

[0016] To address the aforementioned issues, several bias assessment and governance solutions for text-based image models have been proposed in existing technologies. A common approach involves manually pre-setting several macro-level bias dimensions, such as gender, race, age, and occupation. Fixed assessment prompts are then designed around these macro-level bias dimensions. The generated images are then batch-generated, and the results are statistically analyzed using either manual review or automated classification models to determine the degree of bias on specific macro-level bias dimensions. Some solutions further incorporate natural language processing techniques or large language models to assist in prompt generation, expand test samples, or automatically analyze image descriptions to reduce manual workload. Other solutions utilize visual language models for image content recognition, such as determining the gender, age, skin color, occupational symbols, or scene attributes of people in an image, and calculating the distribution differences of different groups in the generated results. Furthermore, at the governance level, existing technologies typically reduce the probability of biased output by filtering training data, optimizing prompts, adding safety rules, fine-tuning the model, and employing reinforcement learning human feedback or safe alignment strategies. These methods can, to some extent, identify and mitigate some known bias problems and provide a reference for pre-deployment evaluation or compliance review of models.

[0017] However, existing solutions still have significant limitations. First, many bias assessment solutions rely on preset dimensions and fixed prompt templates, essentially covering only a limited number of known bias types, making it difficult to promptly identify new biases that emerge with model updates, changes in plugin combinations, or shifts in application scenarios. Second, the granularity of assessments under macro-level bias dimensions is usually coarse, making it difficult to further characterize the differences between subgroups within the macro-level bias dimension, and also difficult to systematically analyze the cross-bias generated by the intersection of multiple social attributes. Third, while some automated bias assessment solutions can reduce manual annotation, they still require continuous manual maintenance of the assessment set and dimension system. Once the business scenario changes, the assessment data and judgment rules need to be replenished, resulting in insufficient scalability and reusability. In addition, existing bias assessment results often remain at the level of risk reports or statistical indicators, failing to form a direct linkage with subsequent model training, data correction, security alignment, or policy updates. The governance process still relies heavily on human experience and judgment, making it difficult to form a closed loop of automatic discovery, quantitative interpretation, policy generation, effect verification, and continuous iteration. Therefore, under the conditions of open world and multi-scenario deployment, how to improve the systematicness and scalability of bias dimension coverage, reduce the dependence on manual construction and annotation, enhance the ability to identify long-tail bias and cross bias, and effectively feed the evaluation results back to the model alignment and governance process remains a technical problem that needs to be solved in the field of bias governance of text-based graph models.

[0018] Based on the above, this application proposes a fine-grained bias assessment system and method for text-based graph models. The following is a detailed description of this application in conjunction with embodiments and accompanying drawings.

[0019] like Figure 1 As shown, this application proposes a fine-grained bias evaluation system for text-based graph models, used to automatically evaluate the generation results of the text-based graph model under different bias dimensions and different subgroup conditions. The system may include a parsing module, an extension module, a construction module, a generation module, and a quantization module. Each module can be deployed on the same computing device or on multiple computing nodes and communicate through a data interface. Each module can be implemented by program instructions executed by a processor or by a combination of hardware and software. Specifically: The parsing module is used to receive the assessment task input information and parse it to obtain the structured task object used to drive the fine-grained bias assessment process.

[0020] It should be noted that the input information for the assessment task can originate from user-input natural language assessment requests, task configurations issued by the assessment platform, or pre-configured batch assessment tasks. The parsing module parses this input information, transforming the originally unstructured or semi-structured assessment requirements into structured task objects that can drive the subsequent fine-grained bias assessment process. This structured task object serves as the unified task carrier for this fine-grained bias assessment process, ensuring consistent assessment criteria across the extension module, construction module, generation module, and quantification module. This avoids inconsistencies in dimensions, groupings, or statistical conditions that might arise from different modules parsing the task separately.

[0021] The extension module is used to determine the set of fine-grained bias dimensions based on structured task objects, and to determine the subgroup grouping rules corresponding to the set of fine-grained bias dimensions.

[0022] The fine-grained bias dimension set is used to limit the bias observation directions that need to be covered in this assessment, while the subgroup grouping rules are used to limit the grouping method of different samples under the same fine-grained bias dimension. Through this setting, the system can form an executable and statistical assessment space starting from the assessment task input information, so that subsequent prompt construction and quantitative analysis are all carried out around the same fine-grained bias dimension and grouping rules.

[0023] The construction module is used to construct a set of assessment prompts for calling the text image model under test, based on the set of fine-grained bias dimensions and the subgroup grouping rules.

[0024] Each assessment prompt in the set is associated with its corresponding fine-grained bias dimension, subgroup, and assessment prompt index, ensuring that each prompt can be traced back to its corresponding assessment dimension and grouping conditions during subsequent generation and quantification. The construction module ensures comparability of assessment prompts corresponding to different subgroups under the same fine-grained bias dimension during prompt construction, thereby reducing the interference of non-target semantic differences on bias assessment results.

[0025] The generation module is used to call the text-to-image model to be tested to sample according to the set of assessment prompts, generate image samples, and form an assessment sample library indexed by fine-grained bias dimensions and subgroups.

[0026] When generating image samples, the generation module establishes an index relationship between the image samples and their corresponding assessment prompts, fine-grained bias dimensions, and subgroups, forming an assessment sample library indexed by fine-grained bias dimensions and subgroups. Therefore, this assessment sample library not only preserves the image samples themselves but also their generation sources, enabling the quantification module to perform statistical analysis on the image samples according to fine-grained bias dimensions and groupings.

[0027] The quantization module is used to obtain image samples from the evaluation sample library, as well as the fine-grained bias dimension index and subgroup index corresponding to the image samples. It performs multimodal analysis on the obtained image samples, extracts attribute features related to the fine-grained bias dimension, and calculates the bias quantification index according to the subgroup grouping rules and the structured task object to obtain the bias evaluation results of each fine-grained bias dimension.

[0028] Therefore, the system can convert the output differences of the Wensheng graph model under different fine-grained bias dimensions into comparable and verifiable quantitative results, thereby alleviating the problems of static limitation of bias dimension coverage, insufficient fine-grained characterization, and strong reliance on manual intervention in existing bias assessment methods.

[0029] The present application will be further described in detail below through some more detailed embodiments.

[0030] In some embodiments of this application, a fine-grained bias assessment system for text-based image models is deployed in an assessment environment and connected to the text-based image model under test. The text-based image model under test can be a generative model capable of generating images based on text prompts, including diffusion models and their variants, a model version obtained through secondary training, or a combined model with plugins. Plugins or lightweight parameter update components can be plugged into the text-based image model to change or enhance its specific generative capabilities or style. After the system connects to the text-based image model under test and its plugins, a continuous workflow is formed around assessment task parsing, fine-grained dimension expansion, prompt construction, image sampling, multimodal quantization, result feedback, and lightweight updates.

[0031] In some embodiments, the system can be implemented using a multi-agent collaborative framework. Multiple agents, under the control of a scheduler, divide tasks and exchange information through shared memory or messaging mechanisms. Different agents can respectively undertake responsibilities such as dimension generation, cross-expansion, prompt generation, compliance filtering, sampling execution, result parsing, and quantitative analysis. The aforementioned agents do not limit the system to a fixed number of subroutines, but rather coordinate the functions of the parsing module, expansion module, construction module, generation module, and quantization module during system operation, enabling bias dimension expansion, sample generation, and result analysis to be performed continuously under the same task caliber.

[0032] During the initialization phase, the system establishes an evaluation dataset and a toxicity dataset, setting the bias threshold τ, sampling parameters, and result storage structure. The evaluation dataset stores evaluation prompts, image samples, evaluation prompt indexes, fine-grained bias dimension indexes, subgroup indexes, and generative metadata. The toxicity dataset stores high-biased samples subsequently identified as highly biased and their corresponding evaluation prompts. Sampling parameters may include parameters used to call the text-based graph model, such as the number of samples, random seed strategy, number of generation steps, and scale. The result storage structure stores structured task objects, bias dimension maps, evaluation prompt sets, evaluation sample libraries, bias evaluation results, alignment training samples, re-evaluation sample libraries, and alignment effect reports. By unifying the data support layer before the evaluation begins, the system ensures that the data written and read by subsequent modules has a consistent index, thereby improving the reproducibility and auditability of the evaluation results. Specific implementation methods for each module may include: The parsing module receives users' natural language assessment requests and converts them into structured task objects that can be directly executed by the subsequent system. The parsing module does not directly perform fine-grained bias assessment; instead, it standardizes user input into a unified task carrier, ensuring that the extension, construction, generation, quantization, and alignment modules all rely on the same input format. This eliminates the need for individual modules to interpret user input, reducing process interruptions caused by missing fields, inconsistent interpretations, or parameter conflicts.

[0033] Specifically, the parsing module extracts at least three types of information from the assessment task input. The first type is the information of the text-to-image model to be tested, including the model name, model version, plugin configuration, or style parameters. The second type is the assessment target information, including macro-level bias dimensions and subgroup cues explicitly mentioned by the user. Macro-level bias dimensions can include high-level social attributes such as gender, race, age, occupation, and region. Subgroup cues can include objects of explicit user interest, such as women or certain professions, and serve as seed conditions for subsequent fine-grained dimension expansion. The third type is assessment constraints and parameter information, including contextual constraints, sampling parameters, assessment budget, stopping conditions, and threshold configurations. Contextual constraints can include the generation context such as scene, style, and language. Sampling parameters can include the number of samples and a random seed strategy. The assessment budget can include the maximum number of dimensions, the maximum number of prompts, or the maximum number of rounds.

[0034] The parsing module maps the extracted text-to-image (TPE) model name and plugin configuration to a pre-defined model registry, forming the TPE model identifier. It maps the evaluation target information to a pre-defined dimension ontology, forming the bias dimension identifier. Then, combining contextual constraints, sampling parameters, evaluation budget, and threshold configuration, it generates a structured task object. This structured task object can include fields such as the TPE model identifier, a macro-level bias dimension set, subgroup seeds, contextual constraints, sampling parameters, evaluation budget, and threshold configuration. For fields not provided by the user but required for workflow operation, the parsing module can complete them using default strategies. For mutually exclusive configurations or abnormal parameters, the parsing module can perform consistency checks and conflict resolution. The parsing module can also generate a parsing report, recording which fields came from explicit user input, which fields were completed by the system, whether there were any ambiguities or conflicts, and the corresponding processing results.

[0035] The structured task object serves as the data entry point between modules in the system. The extension module reads the macro-level bias dimension set, subgroup seeds, contextual constraints, and assessment budget to generate a fine-grained bias dimension set. The construction module reads the contextual constraints and linguistic conditions to limit the semantic scope of the assessment prompts. The generation module reads the model identifier, plugin configuration, and sampling parameters to invoke the text-to-image model to be tested. The quantification module reads the grouping rules, threshold configuration, and sampling parameters to determine the statistical caliber and threshold determination method. Through these data transfer relationships, the structured task object output by the parsing module becomes the unified control object for the entire assessment process.

[0036] Based on the macro-level bias dimensions defined by the structured task objects, the extension module automatically generates and maintains a set of executable fine-grained bias dimensions. The core function of this module is to define the assessment space covered by subsequent evaluations and further refine the macro-level bias dimensions into dimension objects that can be directly used for prompt construction, model sampling, and quantitative statistics. Fine-grained bias dimensions can be represented as risk concerns such as subgroups, specific contexts, narrative patterns, and their cross-combinations. Bias hypotheses can be expressed as verifiable statements about the potential bias exhibited by the text-to-image model under a specific fine-grained bias dimension, used to drive subsequent assessment scenario construction and assessment prompt generation.

[0037] The input to the extension module primarily comes from the structured task objects output by the parsing module, including at least the set of macro-bias dimensions, explicitly mentioned subgroup cues or seed conditions, contextual constraints (such as scene / style / language), and assessment budget and stopping conditions (such as maximum number of dimensions, maximum number of prompts, maximum number of rounds). Furthermore, the extension module can receive anomalous signals and evidence indices from the quantification module, such as high-bias dimension markers, triggering contexts, and sample numbers, and further refine, cross-expand, or prioritize these dimensions in the next iteration.

[0038] The extension module, based on macro-level bias dimensions, subgroup cues, contextual constraints, and assessment budget, derives subgroup dimensions, long-tail dimensions, and cross-dimensional dimensions. Subgroup dimensions are used to segment different groups under the same macro-level bias dimension. Long-tail dimensions cover dimensions that are easily overlooked in traditional static assessment sets but may trigger risks in open contexts. Cross-dimensional dimensions characterize bias manifestations under the combined influence of multiple social attributes or contextual conditions. The extension module performs standardized naming, hierarchical organization, deduplication and merging, priority ranking, and budget trimming on the derived dimensions to obtain the set of fine-grained bias dimensions for this round. This action chain ensures that dimension expansion maintains openness while prioritizing coverage of high-risk or high-value dimensions within the assessment budget.

[0039] The output of the extension module includes two types of results. The first type is the set of fine-grained bias dimensions for this round, which is a list of fine-grained bias dimensions after normalization, deduplication, and budget trimming. Each fine-grained bias dimension includes at least a unified dimension identifier, a corresponding subgroup set, available context labels, and suggested comparison / grouping rules. The unified dimension identifier is used to establish a stable pointer between assessment prompts, assessment samples, assessment results, and alignment data. The corresponding subgroup set is used to limit the grouping statistics objects of the quantification module. Available context labels constrain the scenario, style, and language used by the construction module when generating assessment prompts. The suggested comparison / grouping rules are used to determine the comparison method between different subgroups within the same fine-grained bias dimension. The second type is the bias dimension map update results, which are records of updates to the fine-grained bias dimension tree or bias dimension map, including new nodes, hierarchical relationships, and priority information. These are used for interpretable result presentation, cross-round reuse, and providing a unified pointer for specifying fine-grained bias dimensions for subsequent alignment modules.

[0040] In some embodiments of this application, the extension module generates or updates a bias dimension map. The bias dimension map can be a tree structure or a graph structure, used to structurally represent macro-bias dimensions, fine-grained bias dimensions, and their relationships. The bias dimension map includes macro-bias dimension nodes, fine-grained bias dimension nodes, and hierarchical or relational relationships between macro-bias dimension nodes and fine-grained bias dimension nodes. When the quantification module outputs a high-bias result for a certain fine-grained bias dimension, the extension module updates the fine-grained bias dimension node based on the bias assessment result. For example, it may mark the fine-grained bias dimension node as a high-risk node, increase its priority, supplement the triggering context, or continue to derive finer subgroup dimensions and cross dimensions under the fine-grained bias dimension node. Through the above feedback updates, the system can dynamically expand the coverage of bias dimensions as the assessment progresses, rather than relying on repeated assessments of a fixed set of dimensions.

[0041] During the evaluation iteration phase, the dimension generation agent and the cross-expansion agent propose fine-grained bias hypotheses based on existing macro-bias dimensions, historical evaluation results, and anomalous signals, and write them into the bias dimension map. The prompt generation agent constructs contrast prompts or counterfactual prompts based on the fine-grained bias dimensions, ensuring that the prompt group only changes the target dimension conditions while maintaining semantic consistency in other aspects. The sampling execution agent performs multiple samplings for each evaluation prompt to reduce the perturbation caused by the random generation of the text image model. The parsing and quantization agent performs attribute extraction, narrative signal recognition, and inter-group difference calculation on the generated images, and outputs significance or confidence. The quantization results are fed back to the bias dimension map update process to guide the next round of dimension refinement, cross-expansion, and evaluation prompt optimization. This iterative process can be executed repeatedly until a preset stopping condition is met.

[0042] The construction module generates and maintains a dataset directly usable for evaluation and alignment, based on the set of fine-grained bias dimensions and subgroup grouping rules determined by the extension module. The construction module organizes cue samples, subgroup identifiers, and contextual constraints around the fine-grained bias dimensions, enabling the generation module to generate image samples according to a unified standard, and the quantification module to statistically analyze and backtrack evidence using a unified index. Upon discovering fine-grained bias, the construction module also provides the alignment module with a subset of toxicity data corresponding to the fine-grained bias dimension, supporting targeted lightweight updates and re-evaluation comparisons.

[0043] In some embodiments of this application, the construction module generates a set of executable assessment prompts around the assessment scenario and assessment prompts. The assessment scenario includes a generation context such as scenario type, character or object setting, style constraints or semantic constraints, used to trigger or observe bias performance. The assessment prompts are text instructions that invoke the text-generated graph model, and the prompt template can be a prompt structure with replaceable slots, used to generate contrast prompts while maintaining semantic consistency. Contrast prompts or counterfactual prompts can be a set of prompts that only change a certain sensitive attribute or dimension condition while keeping the rest of the semantics consistent, used to compare the generation differences between different groups and reduce semantic interference.

[0044] The construction module can employ three data generation methods. The first is public corpus templating, where the module extracts descriptions related to the target dimension from publicly available text corpora, restructures them to form standardized prompt templates, and constructs semantically consistent contrastive prompt groups for different subgroups within the same fine-grained bias dimension. This method leverages existing corpora to cover common contexts while improving comparability between different subgroups through template rewriting. The second is manual template supplementation, adding a small number of high-quality prompt templates to dimensions with insufficient coverage or requiring stable reproducibility to ensure comparability and reproducibility of key dimensions. The third is self-instruction expansion generation, where the construction module uses a large language model to expand prompt variants under contextual and compliance constraints, focusing on supplementing the long-tail semantic space and trigger samples under cross-conditions. When the quantification module returns an anomaly fine-grained bias dimension, the construction module can perform targeted expansion on that dimension to improve risk localization accuracy.

[0045] After the construction module generates assessment prompts, it performs deduplication and consistency checks on the prompts. Deduplication removes duplicate or highly similar assessment prompts to prevent low-gain samples from consuming the assessment budget. Consistency checks confirm that assessment prompts within the same control group maintain consistency in scenario, behavior, style, language, and semantic constraints, except for the target dimension conditions. This ensures that the inter-group differences obtained by the quantification module are closer to the differences caused by the target dimension. The processed assessment prompt set can include assessment prompt text, assessment prompt template identifier, fine-grained bias dimension identifier, subgroup identifier, contextual label, control group identifier, and assessment prompt index.

[0046] The construction module employs a combined dynamic and static data management approach. After the evaluation prompts are actually used by the generation module and generate images, the evaluation prompts, images, and their metadata are written into the evaluation prompt set for long-term storage, used for reproduction, auditing, and comparison before and after lightweight updates. Unless the agent maintaining the construction module executes a cleanup strategy, the construction module does not actively delete data already stored in the database. Meanwhile, the construction module is continuously updated as the system runs. When the extension module adds a fine-grained bias dimension, or the quantification module reports a high bias trigger condition, the construction module supplements the relevant fine-grained bias dimension with evaluation prompts, refines subgroup coverage or expands cross conditions, and deduplicates and reduces the weight of duplicate or low-gain evaluation prompts, which can improve coverage efficiency with a limited budget.

[0047] The construction module can be invoked in two phases. The first phase occurs before the evaluation of the specified fine-grained bias dimension. In this phase, the construction module builds or completes the evaluation dataset for the selected fine-grained bias dimension and outputs a set of executable evaluation prompts. The second phase is the alignment preparation phase before re-evaluation. After the quantification module identifies high-biased samples, the construction module extracts the high-biased samples and evaluation prompts corresponding to the fine-grained bias dimension from the stored samples and organizes them into a subset of toxicity data for that fine-grained bias dimension according to alignment requirements. If necessary, the construction module can also perform targeted expansion around that fine-grained bias dimension, providing a data foundation for subsequent lightweight updates and re-evaluations.

[0048] The generation module calls the interface of the raw text image model to be tested. It batch-generates the evaluation prompt set provided by the construction module according to the fine-grained bias dimension selected by the extension module, forming an evaluation sample library for use by the quantification module. The generation module uses the raw text image model identifier and plugin configuration determined by the parsing module as the calling basis, requests and generates images one by one according to the evaluation prompt set, and samples each evaluation prompt multiple times to reduce the impact of the randomness of the raw text image model output on the statistical conclusions.

[0049] During the generation process, the generation module uniformly records the Wensheng image model version, plugin parameters, sampling parameters, random seed, hint index, fine-grained bias dimension index, and subgroup index. This metadata, along with the generated images, is stored in a database, enabling subsequent bias quantification, evidence backtracking, model version comparison, and re-evaluation before and after lightweight updates to trace the sample source. When significant semantic mismatches or missing subjects render the sample unusable, the generation module can perform local regeneration without altering the fine-grained bias dimension and control structure, storing the regeneration reason and configuration in the database. Local regeneration does not change the target fine-grained bias dimension and control relationship, thus preserving the comparability between different subgroups within the same control group.

[0050] After generation, the generation module outputs an assessment sample library organized by fine-grained bias dimension and subgroup indexes. The assessment sample library includes image samples, corresponding assessment prompts, parameters for calling the raw text image model to be tested, and generation metadata. The quantization module can read and statistically analyze image samples according to the fine-grained bias dimension index and subgroup index. The construction module can extract high-bias samples according to the fine-grained bias dimension index in the subsequent alignment stage. The alignment module can construct alignment training samples based on the assessment prompts and image samples corresponding to the high-bias samples. Therefore, the generation module not only handles the calling function of the raw text image model to be tested but also establishes a traceable relationship between assessment prompts, image samples, and fine-grained bias dimensions in the assessment process.

[0051] The quantification module performs bias statistics and interpretable summarization on the evaluation samples generated by the generation module, and compares and analyzes the evaluation results of the same dimension before and after alignment, outputting a quantitative report that can be used for decision-making and auditing. The quantification module uses the fine-grained bias dimensions and their subgroup grouping rules defined by the extension module as the statistical caliber, performing multimodal analysis and correlation discrimination on the generated results corresponding to each evaluation prompt in the evaluation sample library, obtaining sample-level scores and fine-grained bias dimension-level statistics.

[0052] When performing multimodal analysis on image samples, the quantization module can perform processes such as image description generation, attribute recognition, text-image consistency calculation, and narrative label matching to form structured features. Attribute features related to fine-grained bias dimensions can include character representation, occupational symbols, scene elements, emotional expressions, posture, and other observable attributes. For fine-grained bias dimensions involving narrative patterns or implicit stereotypes, the quantization module can also extract narrative signals from image descriptions or labeling results and associate them with the corresponding fine-grained bias dimensions.

[0053] At the sample level, the quantification module outputs traceable evaluation results for each image sample, such as a target group relevance score, whether a bias judgment was triggered, key evidence annotations, corresponding assessment prompts, and an image sample index. The target group relevance score indicates the degree of matching between the image sample and the current subgroup or target dimension conditions; the bias judgment indicates whether the image sample exhibits bias related to the current fine-grained bias dimension. Key evidence annotations record the attribute features, scene elements, or narrative signals that led to the judgment. These sample-level results enable the system to trace back from dimensional statistics to specific images and assessment prompts.

[0054] At the dimensional level, the quantification module outputs bias metrics for each fine-grained bias dimension across subgroups, such as bias rate, mean score, and distribution statistics, and also outputs cross-subgroup differences, such as maximum difference, relative risk ratio, or significance test results. Bias quantification metrics can include between-group difference metrics, trigger rate difference metrics, distribution difference metrics, and their statistical significance or confidence level outputs. Significance or confidence levels can be obtained through resampling, permutation tests, confidence intervals, or threshold rules to characterize the reliability of the bias quantification results. The quantification module calculates bias quantification metrics according to subgroup grouping rules, combined with sampling parameters and threshold configurations, and determines whether the fine-grained bias dimension exceeds a preset threshold, thereby determining whether to proceed to the toxic data extraction and alignment process.

[0055] To support closed-loop governance, the quantification module processes both sets of evaluation results before and after the lightweight update, and outputs comparative conclusions under the same evaluation prompt set, the same sampling strategy, and the same grouping criteria. Specifically, the quantification module outputs the bias rate before the lightweight update for each fine-grained bias dimension. Bias rate after lightweight update and improvement amount , It can further provide the improvement amount of each subgroup and monitoring indicators for whether performance rebound / oversuppression has occurred, such as changes in overall generation quality or mean relevance. When a certain fine-grained bias dimension still exceeds the preset threshold after a lightweight update, the quantification module outputs a residual risk warning and sends back the trigger sample and corresponding evaluation warning index to support the reconstruction of toxicity data or further refinement of the fine-grained bias dimension.

[0056] The output of the quantification module can include structured result tables, visualization charts, and interpretable evidence packages. The structured result tables contain a fine-grained list of bias dimensions, a list of subgroups, sample size, bias rate, mean / quantiles, difference indicators, threshold judgments, corresponding evidence indexes, and comparison fields before and after lightweight updates. The visualization charts visually represent the bias distribution and the effects of lightweight updates, and can include bar charts comparing dimensional bias rates, subgroup bias rate distribution charts (box lines or histograms), and heatmaps or difference matrices of cross-dimensional biases. The interpretable evidence packages are used for audit backtracking and can include representative samples (image samples + assessment prompts) for each highly biased sample, a summary of triggering conditions, and key attributes or narrative signals leading to the judgment. Based on the above outputs, the quantification module generates a bias assessment report and an alignment effect report, providing a unified basis for toxic data extraction, alignment strategy selection, and lightweight update iterations.

[0057] In some embodiments of this application, the system further includes a shared memory module. The shared memory module connects to the extension module, construction module, generation module, quantization module, and alignment module, respectively, and is used to store the fine-grained bias dimension set, the assessment prompt set, the assessment sample library, bias assessment results, and update records of the fine-grained bias dimension nodes. The shared memory module can also store the bias dimension map, historical assessment results, the assessment prompt library, representative evidence, high-risk triggering patterns, toxic data subsets, alignment effect reports, and iteration status.

[0058] The shared memory module plays a crucial role in cross-round reuse and feedback coordination within the system. When the version of the text graph model under test is updated, the plugin configuration changes, or the application context shifts, the system can read the existing bias dimension maps, assessment prompt templates, and high-risk trigger patterns from the shared memory module, and perform incremental expansion and re-validation based on these. Since the bias patterns between adjacent text graph model versions or similar plugin combinations may be similar, the shared memory module can reduce the cost of repetitive manual design and maintenance, and accelerate assessment convergence and risk localization under new versions or plugin configurations. The shared memory module also enables anomalous signals from the quantization module to be fed back to the expansion and construction modules, thus forming a feedback chain of dimension refinement, assessment prompt optimization, and sample regeneration.

[0059] In some embodiments of this application, the system further includes an alignment module. The alignment module is used to extract high-bias samples and their assessment prompts related to the corresponding fine-grained bias dimension from the assessment sample library when the bias assessment result meets a preset bias condition, forming a toxicity data subset. Based on the toxicity data subset, alignment training samples are constructed, and then the alignment training samples are used to perform a lightweight update on the text-to-image model to be tested. The preset bias condition can be triggered by a threshold, for example, when the bias rate (BiasRate) of a certain fine-grained bias dimension d is... d When the value exceeds the preset threshold τ, the system enters the lightweight update phase. The lightweight update phase performs targeted parameter updates on the text image model under test and the fine-grained bias dimension d, and after the update, the same fine-grained bias dimension d is regenerated and re-evaluated to form a closed loop of detection, correction and verification.

[0060] For the detected fine-grained bias dimension d, the alignment module adjusts the toxicity data subset D. tox Two sets of aligned training samples are constructed. The first set is Group A (target group), where the target group is denoted as g. Group A contains samples related to the target group g and is used to suppress specific bias triggering directions under the fine-grained bias dimension d. The second set is Group B (dimensionality coverage group), which covers multiple group samples under the fine-grained bias dimension d. This is used to improve the generalization of lightweight updates and avoid overfitting to only a single group. The toxic data subset D... tox A single sample in the algorithm can be represented as (x, p), where p is the text prompt and x is the image sample corresponding to the text prompt or its latent variable representation. By distinguishing between the target group and the dimension coverage group, the alignment module can update the trigger direction for high-biased samples while maintaining the overall coverage of multiple groups under this fine-grained bias dimension.

[0061] During the lightweight update phase, the system invokes a secure alignment tool and employs a lightweight update strategy based on LoRA (Low-Rank Adaptation). Let the principal parameters of the original text-based graph model be θ0, where the... The original weights of a linear layer (or attention projection layer) that can be injected with LoRA are W. (0) LoRA reparameterizes the weights of linear layers or attention projection layers to W. : W = W (0) + s B A

[0062] Among them, s This is the scaling factor for the linear layer or the attention projection layer. , These are all newly added low-rank matrix parameters, where r represents the low-rank order and d... in With d out Let W represent the input and output dimensions of the weights in the linear layer or attention projection layer, respectively. The original weights W are frozen during fine-tuning. (0) Only train the newly added parameter set φ = { A B This allows for controlled updates without altering the main parameters θ0 of the original graph model.

[0063] The aforementioned parameter relationships enable the system to separate the main parameters of the original graph model from the low-rank incremental parameters oriented towards the fine-grained bias dimension d. (0) Maintain the original image generation capabilities of the original model. B A Provides a low-rank increment for the current fine-grained bias dimension, φ = { A B } represents the set of parameters actually updated during training. Since only the newly added low-rank matrix parameters are trained, the training cost required for lightweight updates is lower than that for a full update of the main model, and it is convenient to save, load, or roll back the corresponding newly added low-rank matrix parameters for different fine-grained bias dimensions.

[0064] As an example, within the diffusion model training framework, the system uses a noisy prediction target for lightweight updates. Let t denote the diffusion time step, and x... t Let represent the noisy sample at diffusion time step t, and ε represent the injected Gaussian noise. Let c(p) represent the text condition vector encoded by the text cue p. Let... , representing the predicted noise value of the textural model after freezing the main parameters θ0 and introducing φ. The optimization objective of lightweight updates is defined as minimizing the noise prediction error:

[0065] Among them, ||·||2 2 Let L2 represent the squared loss. Through the above training, the system obtains a low-rank incremental parameter φ related to the bias dimension d, which is used to suppress the bias response of the graph-text model on this fine-grained bias dimension. Represents the expectation operator. This indicates the data from the subset D of toxicity data. tox Extract (x, p) from the middle.

[0066] Optionally, the system can construct bias offsetting weights based on the learned low-rank terms for bias suppression during the generation stage. Specifically, for the ... A linear layer or attention projection layer that can be injected with LoRA is used to construct offset weights. This allows for the suppression of bias-related feature responses at the parameter level. The bias offsetting weights are a further optional approach; the system can also directly use the trained low-rank incremental parameters to perform lightweight updates after generation.

[0067] After the lightweight update is completed, the generation module, based on the same set of assessment prompts, sampling parameters, and subgroup grouping rules as before the lightweight update, calls the updated Wensheng image model to perform sampling, generating lightweight updated image samples. The lightweight updated image samples, along with their corresponding fine-grained bias dimension index, subgroup index, and assessment prompt index, are then written into the re-evaluation sample library. Maintaining the same set of assessment prompts, sampling parameters, and subgroup grouping rules ensures that the comparison before and after the lightweight update primarily reflects the bias changes brought about by the Wensheng image model parameter update, rather than differences introduced by changes in prompts, sampling, or grouping criteria.

[0068] The quantization module retrieves lightly updated image samples from the re-evaluation sample library, along with the corresponding fine-grained bias dimension index and subgroup index. It calculates the lightly updated bias rate and generates a bias improvement measure based on the bias rates before and after the update. To quantify the effect of the light update, let... This represents the bias rate of the Wensheng graph model before the lightweight update on the fine-grained bias dimension d, let BR d {post(A / B)} This represents the bias rate of the lightweight updated Wensheng graph model on the fine-grained bias dimension d, where post(A / B) indicates that it can be calculated under Group A, Group B, or a combination of both. (Bias improvement amount) Defined as BR d {post(A / B)} .when When the value is greater than 0, it indicates that the bias rate of the fine-grained bias dimension d decreases after the lightweight update, achieving a positive bias reduction effect.

[0069] The quantization module outputs an alignment effect report. This report includes at least the bias rate before the lightweight update, the bias rate after the lightweight update, the amount of bias improvement, the alignment training sample identifiers, and the retest sample index. Optionally, the alignment effect report may also include the identifiers of the toxic data subset used for alignment, the visual statistical results of the retest, the representative sample index, and residual risk warnings. When a certain fine-grained bias dimension after the lightweight update still exceeds a preset threshold τ, the quantization module will trigger the index of the image sample and the corresponding assessment warning to be sent back to the shared memory module. The extension module can further refine this fine-grained bias dimension, the construction module can supplement the assessment warnings or toxic data for this fine-grained bias dimension, and the alignment module can perform another lightweight update. Thus, the system forms a closed-loop bias reduction process of assessment, alignment, and retesting.

[0070] Through the aforementioned system structure and operation, this application enhances the systematicity and scalability of bias assessment dimension coverage in open-world environments. The parsing module transforms natural language assessment requirements into a unified task object, resolving inconsistencies in input criteria across different modules. The extension module and bias dimension map refine macro-level bias dimensions into subgroup dimensions, long-tail dimensions, and cross-dimensional dimensions, alleviating the problem of insufficient coverage of static dimension sets. The construction module integrates publicly available corpus templates, manual template supplementation, and self-instruction extension generation, reducing the cost of manually designing prompts and maintaining assessment sets. The generation module improves the reproducibility of assessment samples through multiple sampling and metadata recording. The quantification module enhances the interpretability and stability of bias assessment results through multimodal parsing, inter-group difference calculation, and significance or confidence output. The alignment module, based on toxicity data extraction and LoRA lightweight updates, enables bias assessment results to directly drive text-based graph model governance and verify effectiveness through re-evaluation.

[0071] In summary, in the pre-launch audit scenario of the text-based image model under test, evaluators can input the model name, version information, plugin configuration, macro-level bias dimensions of interest, and evaluation context constraints of the text-based image model under test. The parsing module converts the above input into structured task objects. The extension module derives subgroup dimensions, long-tail dimensions, and cross-dimensional dimensions based on macro-level bias dimensions such as gender, race, age, occupation, and region, as well as subgroup cues explicitly provided by the user. The construction module generates corresponding contrast prompts or counterfactual prompts. The generation module calls the text-based image model under test to generate image samples. The quantification module outputs the bias rate, between-group differences, significance or confidence level, and representative evidence. The system can provide verifiable evaluation results for the online review, compliance assessment, and risk management of text-based image models.

[0072] In the plugin combination testing scenario, the text graph model under test can be a combined model with one or more plugins loaded. The parsing module maps plugin configurations and model versions to the identifier of the text graph model under test. The generation module records metadata such as plugin parameters, model version, and random seed when calling the text graph model. The quantification module outputs bias evaluation results based on the same fine-grained bias dimension and subgroup grouping rules. Through this scenario, the system can compare bias performance under different plugin configurations and provide a risk basis for plugin deployment or combination deployment.

[0073] In model version update scenarios, the system can read existing bias dimension maps, prompt templates, historical assessment results, and high-risk trigger patterns from the shared memory module. After the parsing module identifies the new text-based graph model version, the extension module performs incremental expansion and re-validation based on the existing bias dimension map, the construction module prioritizes reusing existing assessment prompt sets and supplements them with prompt variations adapted to the new context, the generation module performs sampling according to the same or comparable sampling strategy, and the quantification module compares the bias assessment results under the new version with historical results. Through this scenario, the system can reduce the cost of repeatedly maintaining assessment data across versions and improve the consistency of continuous monitoring.

[0074] In the scenario of fine-grained bias closed-loop governance, when the quantification module determines a certain fine-grained bias dimension dBiasRate d When the value exceeds a preset threshold τ, the alignment module extracts high-biased samples and their evaluation prompts from the evaluation sample library to form a toxicity data subset D. tox And construct the target group Group A and the dimension coverage group Group B. The alignment module freezes the main parameters θ0 of the text image model to be tested, and only trains the newly added low-rank matrix parameter φ = {A B After a lightweight update, the generation module and the quantization module will be re-evaluated using the same set of evaluation prompts, sampling parameters, and subgroup grouping rules. The quantization module will then... BR d {post(A / B)} and Output an alignment effect report. Through this scenario, bias assessment results can be directly transformed into alignment training samples and reassessment basis, thus forming a closed-loop process of detection, correction, and verification.

[0075] In continuous monitoring scenarios, the system can receive multiple assessment tasks in batches at different times and periodically retest the Wensheng graph model using the bias dimension map, assessment prompt library, bias assessment report, and evidence package recorded by the shared memory module. If the quantification module detects an increase in the bias rate of a certain fine-grained bias dimension, the extension module can increase the priority of that fine-grained bias dimension in the next round of assessment, the construction module can supplement the comparison prompts for that fine-grained bias dimension, the generation module can increase the sampling quantity under that fine-grained bias dimension, and the quantification module can recalculate the significance or confidence level. Through this scenario, the system can support the large-scale and continuous assessment of the Wensheng graph model and its plugin ecosystem.

[0076] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of each block is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple blocks may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0077] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0078] Accordingly, such as Figure 2 As shown, this application also proposes a fine-grained bias assessment method for text-based graph models, which may include: S1 receives the assessment task input information and parses it to obtain the structured task object used to drive the fine-grained bias assessment process. S2, determine the set of fine-grained bias dimensions based on the structured task object, and determine the subgroup grouping rules corresponding to the set of fine-grained bias dimensions; S3, construct a set of evaluation prompts for calling the text-to-image model to be tested, based on the fine-grained bias dimension set and the subgroup grouping rules; S4, the text-to-image model to be tested is called to sample according to the set of evaluation prompts, and image samples are generated to form an evaluation sample library indexed by fine-grained bias dimension and subgroup; S5. Obtain image samples from the evaluation sample library, as well as the fine-grained bias dimension index and subgroup index corresponding to the image samples. Perform multimodal analysis on the obtained image samples, extract the attribute features related to the fine-grained bias dimension, and calculate the bias quantification index according to the subgroup grouping rules and the structured task object to obtain the bias evaluation results of each fine-grained bias dimension.

[0079] This application also provides a computer device, which may include a processor and a computer-readable storage medium.

[0080] The processor can be a processor or controller, such as a Central Processing Unit (CPU), a general-purpose processor, a Digital Signal Processor (DSP), an Application-Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The processor can be used to support an electronic device in performing the method steps provided in the above embodiments.

[0081] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the above-described fine-grained bias evaluation method for the Wensheng graph model.

[0082] The computer-readable storage media involved in this application include random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage media known in the art.

[0083] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A fine-grained bias assessment system for text-based graph models, characterized in that, It includes a parsing module, an extension module, a construction module, a generation module, and a quantization module; The parsing module is used to receive the assessment task input information and parse it to obtain a structured task object used to drive the fine-grained bias assessment process. The parsing yields structured task objects used to drive the fine-grained bias assessment process, including: Extract the text image model information to be tested, the evaluation target information, and the evaluation constraints and parameter information from the input information of the evaluation task; The text-to-image model information to be tested is mapped to the text-to-image model identifier, and the evaluation target information is mapped to the bias dimension identifier. Combined with the evaluation constraints and parameter information, a structured task object is obtained. The evaluation target information includes macro-bias dimensions and subgroup cues, and the evaluation constraints and parameter information include contextual constraints, sampling parameters, evaluation budget, and threshold configuration. The extended module is used to determine a set of fine-grained bias dimensions based on the structured task object, determine the subgroup grouping rules corresponding to the set of fine-grained bias dimensions, and generate or update the bias dimension map. The determination of the set of fine-grained bias dimensions based on the structured task object includes: Based on the macro-bias dimension, the subgroup cues, the contextual constraints, and the assessment budget, the subgroup dimension, long-tail dimension, and cross-dimensional dimension are derived. The derived subgroup dimensions, long-tail dimensions, and cross dimensions are normalized, hierarchically organized, deduplicated and merged, prioritized and pruned by budget to obtain the fine-grained bias dimension set. The bias quantification index is calculated based on the subgroup grouping rules and the structured task objects, including: The bias quantification index is calculated based on the subgroup grouping rules, combined with sampling parameters and threshold configuration. The bias dimension map includes macro-bias dimension nodes, fine-grained bias dimension nodes, and hierarchical or association relationships between the macro-bias dimension nodes and the fine-grained bias dimension nodes; Updating the bias dimension map includes updating the fine-grained bias dimension nodes based on the bias assessment results obtained from the quantification module; The construction module is used to construct a set of evaluation prompts for calling the text-to-image model under test, based on the set of fine-grained bias dimensions and the subgroup grouping rules. The generation module is used to call the text-to-image model to be tested to sample according to the evaluation prompt set, generate image samples, and form an evaluation sample library indexed by fine-grained bias dimension and subgroup; The quantization module is used to obtain image samples from the evaluation sample library, as well as the fine-grained bias dimension index and subgroup index corresponding to the image samples, perform multimodal analysis on the obtained image samples, extract attribute features related to the fine-grained bias dimension, and calculate the bias quantification index according to the subgroup grouping rules and the structured task object to obtain the bias evaluation results of each fine-grained bias dimension.

2. The fine-grained bias assessment system for text-based graph models according to claim 1, characterized in that, It also includes a shared memory module; The shared memory module is connected to the extension module, construction module, generation module and quantization module respectively, and is used to store the fine-grained bias dimension set, the assessment prompt set, the assessment sample library, the bias assessment results, and the update records of the fine-grained bias dimension nodes.

3. The fine-grained bias assessment system for text-based graph models according to claim 2, characterized in that, It also includes an alignment module; The alignment module is used to extract high-bias samples and their assessment prompts related to the corresponding fine-grained bias dimension from the assessment sample library when the bias assessment result meets the preset bias conditions, forming a toxic data subset, and constructing alignment training samples based on the toxic data subset. The image model to be tested is then updated using aligned training samples; the high-biased samples are image samples whose bias quantification index value is greater than a preset value. During lightweight updates, the main parameters of the text image model to be tested are frozen, and the newly added low-rank matrix parameters are trained.

4. The fine-grained bias assessment system for text-based graph models according to claim 3, characterized in that: The generation module is also used to, after the text-to-image model under test is lightly updated, call the lightly updated text-to-image model under test for sampling based on the evaluation prompt set, sampling parameters and subgroup grouping rules before the lightly updated update, generate lightly updated image samples, and write the lightly updated image samples and their corresponding fine-grained bias dimension index, subgroup index and evaluation prompt index into the re-evaluation sample library. The quantization module is further configured to obtain lightly updated image samples from the re-evaluation sample library, as well as fine-grained bias dimension indexes and subgroup indexes corresponding to the image samples, calculate the lightly updated bias rate, generate a bias improvement amount based on the bias rate before and after the lightly updated bias rate, and output an alignment effect report including the bias rate before the lightly updated bias rate, the bias rate after the lightly updated bias rate, the bias improvement amount, the alignment training sample identifier, and the re-evaluation sample index.

5. A fine-grained bias assessment method for text-based graph models based on the system described in claim 1, characterized in that, include: Receive assessment task input information and parse it to obtain a structured task object used to drive the fine-grained bias assessment process; Based on the structured task object, a set of fine-grained bias dimensions is determined, and the subgrouping rules corresponding to the set of fine-grained bias dimensions are determined. Based on the fine-grained bias dimension set and the subgroup grouping rules, construct a set of evaluation prompts for calling the text-to-image model to be tested; The text-to-image model to be tested is invoked to sample according to the set of evaluation prompts, generating image samples and forming an evaluation sample library indexed by fine-grained bias dimensions and subgroups; Image samples are obtained from the evaluation sample library, along with the fine-grained bias dimension index and subgroup index corresponding to the image samples. Multimodal analysis is performed on the obtained image samples to extract attribute features related to the fine-grained bias dimension. Bias quantification index is calculated according to the subgroup grouping rules and the structured task object to obtain the bias evaluation results for each fine-grained bias dimension.

6. A computer device, characterized in that, include: Processor and computer-readable storage media; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the fine-grained bias assessment method for the text-based graph model as described in claim 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in claim 5 for fine-grained bias assessment of the Wensheng graph model.

Images