A model evaluation method, device, storage medium and program product

Abstract

The application discloses a model evaluation method, device, storage medium and program product. The method comprises the following steps: acquiring multi-modal evaluation data and constructing a test set containing multiple evaluation tasks; distributing the test set to a model to be evaluated in an isolated execution environment; collecting output results of the model to be evaluated corresponding to each evaluation task; mapping the output results to task scores; performing a high-risk detection mechanism for a preset governance security task; performing a rating fuse operation in response to the governance security task triggering the high-risk detection mechanism; performing weighted calculation on the task scores according to preset weights of the evaluation tasks to obtain a comprehensive score in response to the high-risk detection mechanism not being triggered; generating a capability level report based on the comprehensive score; and performing consistency verification analysis on the output results and preset control standards and outputting a verification result. The application can realize objective, rigorous and safe model capability grading certification.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and large model evaluation technology, and more specifically, to a model evaluation method, device, storage medium, and program product. Background Technology

[0002] Large-scale language models are increasingly used in vertical domains, often for tasks involving the assessment and reasoning of emotional and cognitive states. Existing benchmarks generally suffer from insufficient task coverage, failing to effectively evaluate the performance of large models on core tasks such as feature extraction and quantification. Furthermore, current technologies have severe unimodal limitations, unable to realistically simulate the fusion requirements of multimodal data. Existing evaluation systems generally lack safety red-line mechanisms for high-risk scenarios; missed detection of high-risk signals or the generation of inappropriate intervention suggestions could lead to serious harm. Moreover, existing open evaluation platforms lack blind review and verification loops involving domain experts, making them vulnerable to manipulation by participants using detection methods to inflate rankings.

[0003] How to construct a closed large model evaluation system that covers multimodal inputs, has a mandatory circuit breaker mechanism, and can introduce expert consensus for consistency verification has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] This application provides a model evaluation method, device, storage medium, and program product, which realizes objective, rigorous, and secure model capability classification certification.

[0005] One of the purposes of this application is to provide a model evaluation method. This method includes:

[0006] Acquire multimodal evaluation data and construct a test set containing multiple evaluation tasks;

[0007] In an isolated execution environment, the test set is distributed to the model to be evaluated, and the output results of the model to be evaluated corresponding to each of the evaluation tasks are collected;

[0008] The output results are mapped to task scores, and a high-risk detection mechanism is executed for a preset governance security task.

[0009] In response to the high-risk detection mechanism being triggered by the aforementioned security governance task, a rating-based circuit breaker operation is executed;

[0010] In response to the failure to trigger the high-risk detection mechanism, a comprehensive score is obtained by weighting the scores of the tasks according to the preset weights of each of the evaluation tasks.

[0011] A capability level report is generated based on the comprehensive score, and the output results are compared with a preset reference standard for consistency verification analysis, and the verification results are output.

[0012] In one optional implementation, acquiring multimodal assessment data and constructing a test set containing multiple assessment tasks includes: receiving input data containing text data, scale data, numerical phenotypic data, and biometric data; standardizing the input data; extracting multiple related features based on the standardized input data; and using the related features to construct an emotion state representation task, a state judgment task, a multimodal integration task, a reasoning and decision-making task, and the governance and security task, to combine them to form the test set.

[0013] In one optional implementation, the step of weighting the task scores according to the preset weights of each of the evaluation tasks to obtain a comprehensive score includes: normalizing the task scores of each of the evaluation tasks to obtain a normalized task score; aggregating the normalized task scores of each of the evaluation tasks into multiple ability scores corresponding to each task category; and performing summation quantization based on the weight coefficients assigned to each ability score to obtain the comprehensive score, wherein the ability score corresponding to the state judgment domain is configured with the largest weight coefficient.

[0014] In one optional implementation, the high-risk detection mechanism for a preset governance security task includes detecting whether the output of the model to be evaluated contains a missed detection behavior label for high-risk input signals or a sensitive word combination that has the characteristics of generating inappropriate suggestions.

[0015] In one optional implementation, the rating circuit breaker operation includes: discarding the output results corresponding to other evaluation tasks besides the governance security tasks; severing the regular score calculation link and forcibly resetting the comprehensive score to zero; and downgrading the assessment rating in the capability level report to the pre-registered lowest capability level marker.

[0016] In one optional implementation, the step of distributing the test set to the model to be evaluated in the isolated execution environment and collecting the output results of the model to be evaluated corresponding to each of the evaluation tasks includes: constructing a closed container sandbox environment as the isolated execution environment and deploying an initial running image; setting network access isolation rules in the closed container sandbox environment and trunculating communication access requests to public external networks; asynchronously pushing data fragments in the test set to the data input nodes of the model to be evaluated through a local inter-process communication interface and mounting a private data volume; intercepting the standard output stream of the model to be evaluated using a monitoring daemon to collect the output results; and forcibly taking the model offline and destroying the current running instance of the closed container sandbox environment and reclaiming the corresponding computing resources after a single evaluation task is completed.

[0017] In one alternative implementation, the preset benchmark is a consensus reference result independently and in parallel generated by domain experts based on the same multimodal evaluation data under blind review isolation conditions without external reference information.

[0018] In one optional implementation, the step of performing consistency verification analysis on the output result and a preset comparison standard and outputting the verification result includes: extracting the first text content of the output result and the second text content of the consensus reference result; calculating the semantic similarity score between the first text content and the second text content; and outputting the verification result report with a high consistency conclusion in response to the semantic similarity score being greater than a first similarity threshold parameter.

[0019] In one optional implementation, the calculation of the semantic similarity score between the first text content and the second text content is configured to be implemented by the following steps: calling a pre-trained word vector extraction operator to map the first text content and the second text content into a first semantic vector V1 and a second semantic vector V2 of fixed dimensions, respectively; calculating the cosine distance parameter D between the first semantic vector V1 and the second semantic vector V2 in the vector space; in response to the cosine distance parameter D being less than a preset first distance threshold parameter T1, determining that the logical overlap between the two sets of content meets the standard and setting the semantic similarity score to the maximum constant value; in response to the cosine distance parameter D being between the first distance threshold parameter T1 and the second distance threshold parameter T2, applying a penalty decay operator to the cosine distance parameter D for inverse linear scaling to generate the final calibrated semantic similarity score.

[0020] The second objective of this application is to provide an electronic device, comprising: a memory, a processor, and a computer program stored in the memory; wherein the processor executes the computer program to implement the method described in any of the above optional embodiments.

[0021] The third objective of this application is to provide a computer-readable storage medium having a computer program stored thereon; when the computer program is executed by a processor, it implements the method described in any of the above optional embodiments.

[0022] The fourth objective of this application is to provide a computer program product, including computer instructions; when the computer instructions are executed by a processor, they implement the method described in any of the above optional embodiments. Attached Figure Description

[0023] Figure 1 This is a block diagram of the logical structure of the large model evaluation system provided in the embodiments of this application.

[0024] Figure 2 This is a flowchart of the model evaluation service execution provided in the embodiments of this application.

[0025] Figure 3 This is a flowchart of the model evaluation method provided in the embodiments of this application.

[0026] Explanation of reference numerals in the attached figures:

[0027] In the diagram: 101 - Evaluation Task Library, 102 - Multimodal Input Processing Module, 103 - Evaluation Execution Engine, 104 - Expert Blind Evaluation Verification Module, 105 - Capability Grading Output Module. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] The inventors discovered that general-domain evaluation paradigms cannot be directly transferred to the assessment of emotional and cognitive states. Emotional and cognitive problems typically exhibit high subjectivity, multimodal entanglement, and high-risk concealment. A large model with good text generation capabilities might provide seemingly fluent but unprofessional comforting remarks based on the patient's textual complaints, while missing deeply hidden intervention signals concealed in subtle speech pauses or eye movement delays. This phenomenon, lacking multimodal input support and safety redline constraints, not only fails to be identified in existing benchmarks but may even achieve high scores. Based on this, this application proposes a novel large-scale model evaluation framework by constructing a ten-task system covering specific core capabilities, introducing high-risk mandatory circuit breaker logic, and employing an expert-controlled blind review mechanism.

[0030] Figure 1 This is a logical structure block diagram of the large model evaluation system provided in this application embodiment. This system architecture, serving as the physical and logical carrier of the core method of this application, is composed of multiple highly decoupled functional components connected in series. The system includes an evaluation task library 101, a multimodal input processing module 102, an evaluation execution engine 103, an expert blind evaluation verification module 104, and a capability grading output module 105.

[0031] The evaluation task library 101 serves as the underlying rule data storage center. Internally, it persistently stores datasets covering ten specific evaluation tasks and corresponding scoring rule support files through a relational database and object storage cluster. By storing task configurations and scoring rules independently, it can support dynamic hot updates of the evaluation logic.

[0032] The multimodal input processing module 102 is located in the front-end access layer of the system and is configured to receive text, audio, video and physiological signal data generated from various sensors or analog terminals, and perform normalization, alignment and noise reduction and other standardization processing on them.

[0033] The evaluation execution engine 103 is the core scheduling hub of the system. By establishing a communication connection with the multimodal input processing module 102 and the evaluation task library 101, it distributes the standardized input data along with the corresponding task instructions to the large model under test in a controlled state, and asynchronously collects the inference results output by the model.

[0034] The expert blind review verification module 104 is connected in parallel to the system workflow. It is used to compare the output results of the large model with the pre-entered expert evaluation results in multiple dimensions and perform consistency analysis based on natural language processing.

[0035] The capability grading output module 105 is located at the end of the data link. It is responsible for calculating the comprehensive score based on the scoring results of the front-end engine and generating a standard capability level report from level zero to level four. Through the closed-loop topology of the above five functional modules, full lifecycle evaluation and management can be achieved from data access, model scheduling, dual-track verification to result output.

[0036] To support the background operation and human-computer interaction of the aforementioned system, this embodiment also supplements the necessary underlying infrastructure. The system also includes a configured data storage center and log audit database to support large-capacity, high-concurrency storage of container image files of the large models under test, consensus reference indicators for expert reviews, and massive evaluation audit logs. A human-computer interaction console gateway is configured at the outermost layer of the system. This gateway provides participating vendors with a front-end webpage for registration verification and isolated image uploads, and provides the evaluation operator with a back-end management dashboard for handling expert blind review input, generating capability level certificates, and maintaining dynamic ranking lists. By supplementing the above-mentioned configuration of storage and interaction gateways, the logical closed loop and engineerable operation of the overall evaluation system are ensured.

[0037] Figure 3 This is a flowchart of the model evaluation method provided in this application embodiment. The flowchart clearly demonstrates the core logical nodes for performing a complete large model capability certification. The method includes the following steps:

[0038] Step S301: Obtain multimodal evaluation data and construct a test set containing multiple evaluation tasks.

[0039] In the context of emotion and cognitive state assessment, the input data is not limited to single-dimensional text. This step receives raw input data, including text data, scale data, numerical phenotypic data, and biometric data, by calling the aforementioned multimodal input processing module 102. Data cleaning and feature alignment algorithms are then used to standardize the input data, eliminating dimensional differences caused by different acquisition devices.

[0040] To comprehensively explore the capability boundaries of large models, the system extracts multiple related features from standardized input data and uses these features to construct a test set encompassing five capability domains. These five capability domains are further divided into ten specific tasks: emotional state representation, state judgment, multimodal integration, reasoning and decision-making, and governance and security. To achieve fine-grained evaluation, the five capability domains are further refined into ten specific evaluation tasks.

[0041] Specifically, the construction logic and evaluation content configuration of the ten specific evaluation tasks are as follows: 1) The emotional state representation domain includes: explicit emotion recognition task (extracting basic emotion labels based on input text or facial expressions) and implicit emotion feature extraction task (mining potential psychological stress indicators based on speech intonation pauses or numerical phenotypic data). 2) The state judgment domain includes: a cognitive state quantitative assessment task (scoring the degree of cognitive impairment of the participants by combining scale data) and a mental health risk screening task (predicting the probability of high depression or high anxiety tendency). 3) The multimodal integration domain includes: the audiovisual spatiotemporal alignment task (evaluating the model's ability to align asynchronous biometric data with text streams) and the cross-modal conflict resolution task (evaluating the model's ability to identify the true intent when the text presents a positive signal but the physiological signal presents a negative signal). 4) The reasoning and decision-making domain includes: attribution logic chain generation task (requiring the model to output a structured reasoning path that leads to a specific cognitive state) and dynamic intervention strategy recommendation task (generating personalized guidance or medical advice text based on the determined state). 5) The governance security domain (i.e. governance security tasks) includes: high-risk intent miss detection test task (inputting the model with cryptic inducement data containing self-harm or aggressive tendencies to test the interception rate) and ethics and sensitive word interception task (detecting whether the suggested features generated by the model contain sensitive word combinations that violate medical ethics or legal norms).

[0042] To support more granular ablation evaluation experiments, the test set in this embodiment can be divided into three progressive modal levels: mode 1, mode 2, and mode 3. This progressively structured test set effectively identifies the performance bottlenecks of the large model under test at different information richness levels.

[0043] Step S302: Distribute the test set to the model to be evaluated in the isolated execution environment and collect the output results.

[0044] Large model evaluations often allow participants to directly obtain test data by calling external application programming interfaces (APIs), which creates vulnerabilities for probing the test set. To mitigate this risk, this application introduces a mandatory sandbox isolation mechanism.

[0045] The system's front-end gateway restricts direct calls to external application interfaces, requiring participants to upload container images containing the inference code of the model to be evaluated. The system utilizes virtualization technology to construct a closed container sandbox environment as the isolated execution environment, within which the initial running image is deployed. To achieve physical-level data leakage prevention, strict network access isolation rules are set within the closed container sandbox environment, cutting off communication access requests to public external networks at the kernel level.

[0046] During operation, the evaluation execution engine asynchronously pushes data fragments from the test set to the data input nodes of the model under evaluation via a local inter-process communication interface or by mounting a read-only private data volume, ensuring that the private test set data does not leave the database. A system-level monitoring daemon intercepts the standard output stream of the model under evaluation to securely collect output results. After a single evaluation task is completed, the system will forcibly shut down and use orchestration tools to completely destroy the currently running instance of the closed container sandbox environment, and reclaim the corresponding graphics processing unit (GPU) or central processing unit (CPU) computing resources.

[0047] Through a strict isolation sandbox operation and resource control mechanism, the cheating path of the model using external networks to access information or send test questions back to the outside can be effectively blocked, ensuring the authenticity of the evaluation results from the bottom layer of the project.

[0048] Step S303: Map the output results to task scores and execute a high-risk detection mechanism. The high-risk detection mechanism is executed for pre-defined governance security tasks.

[0049] After collecting the original text or structured output of the model for the evaluation task, the system uses an index conversion operator to map the original indexes into normalized task scores, thereby providing a unified dimension for subsequent mathematical aggregation.

[0050] This application incorporates a security threshold detection node into the core workflow control. For pre-defined governance security tasks, a specific high-risk detection mechanism is executed. The underlying logic of this mechanism is to detect whether the output of the model under evaluation contains undetected behavior labels for high-risk input signals, or whether the intervention text it generates contains sensitive word combinations that suggest inappropriate recommendations. This high-risk detection mechanism establishes the security threshold that the model may cross in practical applications.

[0051] Step S304: In response to triggering the high-risk detection mechanism, execute the rating circuit breaker operation.

[0052] When the security threshold routing node determines that a violation has occurred in the aforementioned tasks, the system will initiate a hard rating circuit breaker operation. The system will discard the output results of the model under evaluation on all evaluation tasks other than governance security tasks, ignoring its performance in knowledge question answering or information extraction; it will sever the regular weighted score calculation link on the underlying computing bus, forcibly resetting the model's global comprehensive score to zero using an interrupt command; and in the generated report, it will downgrade the model's capability rating to the lowest capability level flag pre-registered in the database. This circuit breaker and zeroing operation effectively prevents high-risk incidents that may arise from large models with serious logical flaws in healthcare scenarios, ensuring that the evaluation and certification system possesses access defense value.

[0053] Step S305: In response to the absence of a high-risk detection mechanism, a weighted calculation is performed to obtain a comprehensive score. The comprehensive score is obtained by weighting the task scores according to the preset weights of each evaluation task.

[0054] Once the model successfully passes the security check, the system execution flow transitions to the normal quantitative evaluation branch. The normalized task scores of each evaluation task are clustered by mean according to their respective capability domains, aggregating them into multiple capability scores corresponding to each task category.

[0055] The algorithm performs a summation quantification process based on the weighted coefficients assigned to each ability score. To highlight the specific needs of the emotion recognition scenario, the weighted algorithm specifies that the ability score corresponding to the state judgment domain is configured with the largest weighted coefficient.

[0056] To achieve a scientific overall assessment of various ability scores, as a preferred embodiment, a model using vector inner product and linear bias is employed for quantitative calculation. Based on a preset combination of floating-point constants and extracting benchmark scores, the calculation logic for the comprehensive score is implemented through the following formula:

[0057]

[0058] in, This represents the final overall score output by the system. This represents the state characterization capability score obtained from aggregation; This represents the state judgment capability score obtained from aggregation; This represents the multimodal integration capability score obtained from aggregation; This represents the reasoning and decision-making ability score obtained from aggregation; This represents the governance security capability score obtained from aggregation.

[0059] To meet the engineering requirements of accelerated calculations using array matrices at the system's underlying level, the aforementioned weighted calculation logic is configured as follows: Multiple sets of preset floating-point constants from the system configuration items are read as weight coefficients; the weight coefficients corresponding to the aforementioned state judgment domains are set to a constant value Wmax, with a baseline value of 0.30 for Wmax; the system obtains a first numerical array A1 synthesized from various capability groups, and a second numerical array A2 corresponding to each weight coefficient; the underlying arithmetic logic unit is called to calculate the inner product of the first numerical array A1 and the second numerical array A2, thereby generating a baseline score; this baseline score is extracted, and the system administrator is allowed to superimpose preset bias adjustment constants under specific calibration conditions to generate a global comprehensive score, which is then persistently written to the storage unit via a transaction mechanism.

[0060] The above-described weight allocation is merely an optimization example for this specific sentiment benchmark. In other embodiments, if the evaluation focuses more on biological feature analysis, those skilled in the art can dynamically adjust the weight coefficients of the multimodal integration capability score, as long as the circuit breaker logic takes precedence over the weighting logic, thus achieving the purpose of this application. Through differentiated weighting design, the true level of the model on core business indicators can be accurately characterized.

[0061] Step S306: Generate a capability level report based on the comprehensive score and perform a consistency verification analysis. Compare the output results with a preset comparison standard and output the verification results.

[0062] After obtaining the comprehensive score, the system will map the query to the preset segmentation function and generate the corresponding ability level assessment report from level zero to level four.

[0063] This application introduces a parallel workflow on the expert side. The pre-defined benchmark is a consensus reference result generated by a domain expert review committee through independent and parallel deduction based on the same multimodal evaluation data, under blind review isolation conditions without external reference information.

[0064] During the consistency verification analysis, the system extracts the first text content of the large model's output and the second text content of the expert consensus reference results. The semantic similarity score between the two is calculated using the natural language processing module. When this similarity score is greater than the system's preset first similarity threshold parameter, it indicates that the large model's reasoning logic highly aligns with the clinical intuition of human experts, thus outputting a verification result report recording a high degree of consistency and an acceptable threshold analysis.

[0065] To achieve high-resolution semantic consistency accuracy, the semantic similarity score calculation process is configured as follows: call the pre-trained word vector extraction operator to map the first text content and the second text content into a first semantic vector V1 and a second semantic vector V2 of fixed dimensions, respectively.

[0066] Calculate the cosine distance parameter D between the first semantic vector V1 and the second semantic vector V2 in the vector space. In a preferred embodiment, the cosine distance parameter D is calculated using the following mathematical formula:

[0067]

[0068] Where D represents the cosine distance parameter used to measure the degree of semantic disagreement between two texts; V1 represents the semantic feature vector of the output of the model to be evaluated; and V2 represents the semantic feature vector of the expert consensus reference result. The value of parameter D ranges from 0 to 2, with the value closer to 0 indicating that the semantics are closer.

[0069] After obtaining the distance parameter D, the system performs a two-level threshold judgment. If the cosine distance parameter D is less than the preset first distance threshold parameter T1, the system determines that the overlap between the model output and the expert logic is fully met, and sets the semantic similarity score to the maximum constant value. If the cosine distance parameter D is between the first distance threshold parameter T1 and the second distance threshold parameter T2, the system does not directly score the model zero. Instead, it introduces a penalty decay operator to parameter D and performs inverse linear scaling calculation to generate the final calibrated smooth semantic similarity score. By introducing word vector mapping and dual-threshold decay logic, the system can effectively accommodate differences in word choice and sentence construction within the model, truly reflecting the expert consistency level of its inherent logical reasoning ability.

[0070] Figure 2 This document illustrates a flowchart of the model evaluation process provided in an embodiment of this application. The flowchart describes how the participating parties and expert verification teams collaborate.

[0071] Participants register on the terminal and submit a container image containing the model. After the image is received by the closed evaluation platform, the system uses a frequency limiting module to intercept abnormally high-frequency submission requests and generates anti-brute-force submission logs. After successful verification, the image is pushed into a sandbox isolation environment for silent testing with a private test set not released from the database. During the execution of the evaluation task, the original indicators of each task are continuously converted into task scores. The system determines whether a security threshold has been triggered. If triggered, the red line path of resetting the comprehensive score to zero and forcibly downgrading is executed; if not triggered, the process proceeds to the quantitative evaluation path, clustering the task scores into five capability scores, executing a weighted summation algorithm, and finally outputting a capability level assessment and evaluation report. On the parallel expert verification side, a consensus standard is generated through a domain expert committee, and a blind evaluation comparing the model with experts is performed to ensure that the output results have interpretable evidence.

[0072] To curb the practice of using publicly available question banks for pre-tuning of models, this application introduces dynamic probes and hierarchical queue features into the management mechanism of the evaluation task bank.

[0073] When maintaining its internal private test set, the system constructs a two-tiered, hierarchical isolation structure consisting of a discovery queue and a validation queue. The discovery queue contains the latest non-public data streamed in real-time from real-world medical or psychological interaction scenarios. This data, after multimodal cleaning, is periodically injected into a sandbox environment. The validation queue retains classic, fixed-paradigm test questions to calibrate baselines. When distributing the test set, the evaluation execution engine introduces a pseudo-random seed, dynamically assembling the question sequence for a single evaluation according to a specific ratio. By constructing a two-tiered private queue mechanism based on a real-time streaming data pool, the system can leverage the time freshness and unpredictable structure of the data to prevent ranking manipulation based on historical data, thereby improving the objectivity of large model capability certification in industry entry requirements.

[0074] This application overcomes the limitations of traditional single-modal text evaluation by organically combining five categories and ten items of multimodal test domains, a high-risk red line mandatory circuit breaker mechanism, and a closed-loop expert-compared blind review. It utilizes an isolation sandbox to ensure the physical security of the test set, employs a weighted algorithm to highlight the importance of core judgment capabilities, and leverages a high-risk missed detection veto strategy to build a security defense for large-scale model applications. Through the synergistic effect of hardware and software logic, it systematically constructs a comprehensive evaluation benchmark for large-scale models with commercial certification value and authoritative significance.

[0075] Those skilled in the art will understand that various aspects of this application can be implemented as systems, methods, or program products. Various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation, or a combination of hardware and software aspects, collectively referred to herein as circuits, modules, or systems.

[0076] In another embodiment of this application, an electronic device is provided. This electronic device, serving as a large model evaluation server, includes a physical architecture layer of memory and a processor. The memory is configured as a non-volatile storage medium for persistently storing multimodal input data, sandbox image configurations, and computer programs. As a preferred embodiment, the memory can be random access memory, flash memory, read-only memory, etc., including but not limited to non-transient storage media such as optical disc storage, magnetic storage devices, or solid-state drives; this application does not specifically limit this. The processor can be a central processing unit or a graphics processing unit, configured to read and execute the computer program stored in the memory via an internal data bus. When the processor executes the program instructions, by calling its internal arithmetic logic units and registers, it can fully implement all the method steps regarding data acquisition, isolation distribution, red-line circuit breaking, weighted scoring, and expert blind review verification as described in the above embodiments. Through the collaboration of this hardware structure, high-throughput model concurrent evaluation computing power support can be provided.

[0077] In another embodiment of this application, a computer-readable storage medium is provided. Computer program code is physically burned onto it and stored in the form of a binary instruction stream. When the storage medium is configured and connected to a specific processor system, the program code is read, interpreted, and executed, thereby implementing the model evaluation method provided in any of the above-described method embodiments of this application. The computer-readable storage medium may take the form of any combination of one or more computer-readable media.

[0078] In another embodiment of this application, a computer program product is provided. This product includes computer instructions or a software development kit that exist in a specific compiled form on a physical carrier. When these computer instructions are invoked and executed by the processor of an external computing terminal or cloud server, they can instantiate functional units such as the evaluation task library, input processing module, execution engine, and verification module described in this application, thereby achieving closed-loop control of the entire evaluation process as described in any of the above method embodiments.

[0079] The terms "first," "second," etc., used in this specification are merely for distinguishing one entity from another and do not necessarily require or imply any such actual relationship or order between the entities. The terms "include," "comprise," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed.

[0080] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. The scope of protection of this application shall be determined by the scope of the claims.

Claims

1. A model evaluation method, characterized in that, include: Acquire multimodal evaluation data and construct a test set containing multiple evaluation tasks; In an isolated execution environment, the test set is distributed to the model to be evaluated, and the output results of the model to be evaluated corresponding to each of the evaluation tasks are collected. This process includes: constructing a closed container sandbox environment as the isolated execution environment and deploying an initial running image; setting network access isolation rules in the closed container sandbox environment and trunculating communication access requests to public external networks; asynchronously pushing data fragments from the test set to the data input nodes of the model to be evaluated through a local inter-process communication interface and mounted private data volumes; intercepting the standard output stream of the model to be evaluated using a monitoring daemon to collect the output results; and forcibly taking the model offline and destroying the currently running instance of the closed container sandbox environment and reclaiming the corresponding computing resources after each evaluation task is completed. The output results are mapped to task scores, and a high-risk detection mechanism is executed for a preset governance security task. In response to the high-risk detection mechanism being triggered by the aforementioned security governance task, a rating-based circuit breaker operation is executed; In response to the failure to trigger the high-risk detection mechanism, a comprehensive score is obtained by weighting the scores of the tasks according to the preset weights of each of the evaluation tasks. A capability level report is generated based on the comprehensive score, and the output results are compared with a preset comparison standard for consistency verification analysis, and the verification results are output. The process of comparing the output results with the preset comparison standard for consistency verification analysis and outputting the verification results includes: extracting the first text content of the output results and the second text content of the consensus reference results; calculating the semantic similarity score between the first text content and the second text content; and outputting the verification result report with a high consistency conclusion in response to the semantic similarity score being greater than a first similarity threshold parameter. The calculation of the semantic similarity score between the first text content and the second text content is configured to be implemented by the following steps: calling the pre-trained word vector extraction operator to map the first text content and the second text content into a first semantic vector V1 and a second semantic vector V2 of fixed dimensions respectively; Calculate the cosine distance parameter D between the first semantic vector V1 and the second semantic vector V2 in the vector space; in response to the cosine distance parameter D being less than a preset first distance threshold parameter T1, determine that the logical overlap of the two sets of content meets the standard and set the semantic similarity score to the maximum constant value; in response to the cosine distance parameter D being between the first distance threshold parameter T1 and the second distance threshold parameter T2, apply a penalty decay operator to the cosine distance parameter D for inverse linear scaling to generate the final calibrated semantic similarity score.

2. The method as described in claim 1, characterized in that, The process of acquiring multimodal evaluation data and constructing a test set containing multiple evaluation tasks includes: It receives input data that includes text data, scale data, numerical phenotypic data, and biometric data; The input data is standardized. Based on the standardized input data, multiple related features are extracted, and the related features are used to construct ten specific evaluation tasks covering five capability domains: emotional state representation, state judgment, multimodal integration, reasoning and decision-making, and governance and security, to form the test set.

3. The method as described in claim 1, characterized in that, The step of weighting the task scores according to the preset weights of each of the evaluation tasks to obtain the comprehensive score includes: The task scores of each of the aforementioned evaluation tasks are normalized to obtain normalized task scores; The normalized task scores of each of the aforementioned evaluation tasks are aggregated into multiple capability scores corresponding to each task category. The comprehensive score is obtained by summing and quantifying the weight coefficients assigned to each of the ability scores, wherein the ability score corresponding to the state judgment is assigned the largest weight coefficient.

4. The method as described in claim 1, characterized in that, The high-risk detection mechanism for a pre-defined governance security task includes detecting whether the output of the model under evaluation contains a missed detection behavior label for high-risk input signals or a sensitive word combination that generates inappropriate suggestions.

5. The method as described in claim 4, characterized in that, The execution of the rating-based circuit breaker operation includes: Discard the output results corresponding to the evaluation tasks other than the governance security tasks. The normal score calculation process is interrupted, and the comprehensive score is forcibly reset to zero. The assessment rating in the capability level report will be downgraded to the lowest pre-registered capability level marker.

6. The method as described in claim 1, characterized in that, The pre-set comparison standard is a consensus reference result independently and in parallel generated by domain experts based on the same multimodal evaluation data under blind review isolation conditions without external reference information.

7. An electronic device, characterized in that, include: Memory, processor, and computer program stored in said memory; When the processor executes the computer program, it implements the method of claim 1.

8. A computer-readable storage medium, characterized in that, It stores a computer program; when the computer program is executed by a processor, it implements the method of claim 1.

9. A computer program product, characterized in that, Includes computer instructions; when executed by a processor, the computer instructions implement the method of claim 1.

Images