Fine-tuning corpus filtering method and device based on large model, equipment and medium
By using a large model scoring and correction method, the problems of low efficiency and insufficient security in fine-tuning corpus filtering are solved, achieving efficient and reliable corpus security filtering and ensuring the security and accuracy of model output.
Patent Information
- Application Number
- CN202511456597.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-13
- Publication Date
- 2026-01-09
AI Technical Summary
Existing fine-tuning corpus filtering methods are inefficient and prone to errors, failing to effectively guarantee corpus security and causing the model to output harmful or biased content.
By acquiring target fine-tuning corpora, performing feature extraction, and using multiple large models for scoring and correction, combined with security prompt words and domain classification models, a security score is generated to filter out unsafe corpora.
It improves the filtering efficiency of fine-tuning corpora, ensures corpus security, reduces manual intervention, and enhances the reliability of model output.
Smart Images
Figure CN121303341A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and more specifically, to a method, apparatus, device, and medium for fine-tuning corpus filtering based on a large model. Background Technology
[0002] With the widespread application of Large Language Models (LLMs) across various industries, the limitations of general-purpose LLMs in specific domains have become increasingly apparent. To improve their performance in these domains, fine-tuning training with corpora specific to that domain is necessary. However, due to the wide variety and sheer volume of fine-tuning corpora, the problem of attackers poisoning these corpora is becoming increasingly prominent. Fine-tuning corpus poisoning refers to attackers carefully constructing insecure text and mixing it into normal fine-tuning corpora. When such corpora are used for fine-tuning large models, it can cause the model to produce harmful, biased, or inappropriate output.
[0003] Existing methods for filtering the content of fine-tuned corpora mainly include two types: manual filtering and filtering based on regular expression rules. Manual filtering is extremely inefficient and its effect is unstable. Its filtering quality depends on the professional level of the screeners and is prone to errors and omissions. Filtering based on regular expression rules requires pre-defining filtering rules and is prone to missed or false judgments, which can lead to poor performance of the fine-tuned model or the generation of harmful, biased or inappropriate output content. Summary of the Invention
[0004] The purpose of this application is to provide a method, apparatus, device, and medium for filtering fine-tuned corpora based on large models, so as to solve the above-mentioned problems existing in the prior art, improve the filtering efficiency of fine-tuned corpora, and ensure the security of fine-tuned corpora.
[0005] Firstly, a method for fine-tuning corpus filtering based on a large model is provided, which may include: Obtain the target fine-tuning corpus; Feature extraction is performed on the target fine-tuning corpus to obtain corpus features; Based on the corpus features, match the security prompt words corresponding to the configured corpus features with multiple first target large models and multiple second target large models; The target fine-tuning corpus is scored based on the security prompt words using multiple matched first target large models to obtain multiple first scores; Multiple first scores are corrected based on the security prompt words using multiple matched second target large models to obtain multiple second scores; Based on multiple second scores, a safety score for the target fine-tuning corpus is obtained; If the security score is less than the configured security threshold, then the target fine-tuning corpus is filtered.
[0006] In an optional implementation, feature extraction is performed on the target fine-tuning corpus to obtain corpus features, including: Keyword extraction is performed on the target fine-tuning corpus to obtain the first feature; Feature extraction is performed on the target fine-tuning corpus to obtain the second feature; The first feature and the second feature are input into a pre-trained domain classification model to obtain the domain classification result corresponding to the target fine-tuning corpus; The first feature, the second feature, and the domain classification result are used as the corpus features of the target fine-tuning corpus. In an optional implementation, multiple first scores are corrected based on the security prompt words using multiple matched second-target large models, including: For any second target big model, the second target big model is used to correct multiple first scores based on the security prompt words to obtain the corrected score corresponding to the second target big model; The first scores are corrected using the corrected scores corresponding to the second target large model, resulting in multiple corrected first scores; Multiple second scores are obtained based on multiple first scores after correction of different second objective large models.
[0007] In one optional implementation, multiple second scores are obtained based on multiple first scores corrected from different second objective large models, including: Obtain the weights of different second-objective large models on the domain classification results corresponding to the target fine-tuning corpus; Based on the weights of different secondary objective models in the classification results of the corresponding domains, multiple first scores after correction of different secondary objective models are weighted and summed to obtain multiple second scores.
[0008] In an optional implementation, after obtaining the security score of the target fine-tuning corpus, the method further includes: The corpus features of the target fine-tuning corpus and the security score are stored in the configured corpus scoring library.
[0009] In an optional implementation, after obtaining the corpus features, the method further includes: Calculate the similarity between the features of each historical corpus in the corpus scoring database and the features of the target fine-tuning corpus; The historical security score corresponding to the historical corpus features with a similarity greater than the configured similarity threshold is used as the security score of the target fine-tuning corpus.
[0010] Secondly, a fine-tuning corpus filtering device based on a large model is provided, which may include: The acquisition unit is used to acquire the target fine-tuning corpus. The extraction unit is used to extract features from the target fine-tuning corpus to obtain corpus features; The matching unit is used to match the security prompt words corresponding to the configured corpus features with multiple first target large models and multiple second target large models according to the corpus features. The scoring unit is used to score the target fine-tuning corpus based on the security prompt words using multiple matched first target large models to obtain a first score; The verification unit is used to modify the first score based on the security prompt words using multiple matched second target large models to obtain a second score; A filtering unit is used to obtain a security score for the target fine-tuning corpus based on multiple second scores. If the security score is less than a configured security threshold, the target fine-tuning corpus is filtered.
[0011] Thirdly, an electronic device is provided, which includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements any of the steps described in the first aspect above.
[0012] Fourthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the steps of any of the methods described in the first aspect above.
[0013] This application enables multiple large models to simulate the role of domain experts, score and cross-validate target fine-tuned corpora, accurately identify the safety of corpus content, and improve filtering effect; by leveraging the deep semantic understanding and context awareness capabilities of modern large language models, there is no need for manual intervention to write complex filtering rules, which greatly reduces the complexity of system implementation. Attached Figure Description
[0014] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application 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.
[0015] Figure 1 A system architecture diagram of a fine-tuning corpus filtering method based on a large model provided for embodiments of this application; Figure 2 A flowchart illustrating a fine-tuning corpus filtering method based on a large model, provided for embodiments of this application; Figure 3 A schematic diagram of a fine-tuning corpus filtering device based on a large model provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0016] 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 a part of the embodiments of this application, and not all of the 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. Unless otherwise defined, the technical or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art. The words "first," "second," and similar terms used in this application do not indicate any order, quantity, or importance, but are only used to distinguish different components. The words "comprising" or "including," etc., mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but do not exclude other elements or objects. The words "connected," "coupled," or "connected," etc., are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0017] The fine-tuning corpus filtering method based on a large model provided in this application can be applied to... Figure 1 In the system architecture shown, such as Figure 1As shown, the system may include: a server in the enterprise's backend and terminals for enterprise employees. The server can be a physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal may be a user equipment (UE) such as a mobile phone, smartphone, laptop, digital radio receiver, personal digital assistant (PDA), tablet computer (PAD), handheld device, in-vehicle device, wearable device, computing device, or other processing device connected to a wireless modem, mobile station (MS), mobile terminal, etc. The terminal and server can be directly or indirectly connected via wired or wireless communication methods; this application does not limit the connection.
[0018] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.
[0019] Figure 2 This is a flowchart illustrating a fine-tuning corpus filtering method based on a large model, provided as an embodiment of this application. Figure 2 As shown, the method may include: Step S210: Obtain the target fine-tuning corpus.
[0020] The target fine-tuning corpus is obtained by reading from the data source and performing preprocessing and standardization operations.
[0021] Step S220: Extract features from the target fine-tuning corpus to obtain corpus features.
[0022] Specifically, feature extraction from the target fine-tuning corpus includes: Keyword extraction is performed on the target fine-tuning corpus to obtain the first feature; specifically, mature NLP libraries or tools are used to extract keywords, noun phrases, and named entities from the target fine-tuning corpus to obtain the first feature. Feature extraction is performed on the target fine-tuning corpus to obtain the second feature; specifically, the target fine-tuning corpus is input into a pre-trained sentence embedding model to obtain a semantic vector; text style features of the target fine-tuning corpus are extracted to obtain a text vector; the second feature is obtained based on the semantic vector and the text vector; the text style features include content structure, sentence length, lexical complexity, sentiment polarity, and whether special symbols are included; The first and second features are input into a pre-trained domain classification model to obtain the domain classification result corresponding to the target fine-tuning corpus; the first feature, the second feature, and the domain classification result are used as the corpus features of the target fine-tuning corpus. Step S230: Based on the corpus features, match the security prompt words corresponding to the configured corpus features with multiple first target large models and multiple second target large models.
[0023] Before matching the configured security prompt words and multiple first-target large models and multiple second-target large models according to the corpus features, the method may further include: Configure security warning words corresponding to different corpus features, specifically including: For any given domain, acquire the domain security knowledge document for that domain. This document includes: domain regulations, industry standards and ethical guidelines, a sensitive information dictionary, and a typical risk case library. The sensitive information dictionary contains sensitive words, phrases, and entities specific to that domain (such as disease names, drug formulations, and customer account information). The typical risk case library contains past cases of inappropriate content and their harmfulness analysis. Extract keywords and entities from the domain security knowledge document to obtain an initial sensitive feature dictionary for that domain. This initial sensitive feature dictionary includes: entities, risk levels, and association rules. Obtain a sensitive feature dictionary enhanced from the initial sensitive feature dictionary for this domain; the sensitive feature dictionary includes feature tags added by domain experts to the initial sensitive feature dictionary; the feature tags include: risk type, content style, and content structure; generate multiple feature clusters for this domain based on the sensitive feature dictionary for this domain; obtain a pre-built security assessment template for this domain and sample corpora of different feature clusters for this domain; the security assessment template includes: a prompt word engineering framework and security assessment dimensions; the sample corpora include secure text, insecure text, and standard scores; For any feature cluster, add the feature cluster to the security assessment template of the corresponding domain to obtain the initial security prompt words for the feature cluster; use the initial prompt words of the feature cluster to call the corresponding large model to score the sample corpus of the feature cluster; optimize the initial security prompt words based on the scoring results and the standard scores in the sample corpus to obtain the security prompt words for the feature cluster; based on the security prompt words of different feature clusters, obtain the security prompt words for the domain.
[0024] In practice, based on the domain classification results, security prompt words for the corresponding domain are matched; based on the first feature and the second feature, security prompt words for the corresponding feature clusters of the corresponding domain are matched to obtain security prompt words corresponding to the corresponding corpus features; using sample corpora corresponding to different feature clusters of different domains, different large models are trained to obtain multiple first-objective large models and multiple second-objective large models corresponding to different corpus features.
[0025] Step S240: Use the matched first target large model to score the target fine-tuning corpus based on the safety prompt words to obtain multiple first scores.
[0026] For each piece of fine-tuned corpus content to be tested, multiple large language models are concurrently invoked for scoring using specially designed security prompts. Specifically, 3 to 5 large models are selected and invoked using the designed content security scoring prompts to obtain the scoring results. The content security scoring prompts explicitly require the large models to score the fine-tuned corpus content as security experts in the fine-tuning domain, with scores ranging from 0 to 10 (higher scores indicate stronger security), and the scores are directly output.
[0027] Step S250: Use the matched second target big model to correct the multiple first scores based on the security prompt words to obtain multiple second scores.
[0028] Specifically, for each score result of the fine-tuned corpus to be verified, multiple large language models are called concurrently for verification using specially designed content security verification prompts.
[0029] Specifically, multiple first scores are corrected based on security prompts using multiple matched second-target large models, including: For any second objective model, multiple first scores are corrected based on safety prompts using the second objective model to obtain corrected scores corresponding to the second objective model; each first score is corrected using the corrected scores corresponding to the second objective model to obtain multiple corrected first scores; multiple second scores are obtained based on the multiple first scores corrected by different second objective models; wherein, the weights of different second objective models in the domain classification results corresponding to the target fine-tuning corpus are obtained; based on the weights of different second objective models in the corresponding domain classification results, the multiple first scores corrected by different second objective models are weighted and summed to obtain multiple second scores.
[0030] In practical applications, 3 to 5 large models are selected and invoked using pre-designed content security verification prompts to obtain verification results. The content security verification prompts explicitly require the large models to act as fine-tuning domain security experts, checking for misjudgments or omissions in the large model scoring results of step three, and directly outputting the corrected scores.
[0031] Step S260: Based on multiple second scores, obtain the security score of the target fine-tuning corpus; if the security score is less than the configured security threshold, filter the target fine-tuning corpus; if the security score is not less than the configured security threshold, retain the target fine-tuning corpus.
[0032] In practice, multiple second scores are weighted and summed to obtain the safety score of the target fine-tuning corpus.
[0033] Specifically, by comparing the safety scores and safety thresholds of each target fine-tuning corpus, it is determined whether the target fine-tuning corpus is safe, filtering out unsafe target fine-tuning corpora, and retaining safe target fine-tuning corpora.
[0034] In some embodiments of this application, after obtaining the security score of the target fine-tuning corpus, the method further includes: Store the corpus features and security scores of the target fine-tuned corpus into the configured corpus scoring library.
[0035] In some embodiments of this application, after obtaining the corpus features, the method further includes: Calculate the similarity between the features of each historical corpus in the corpus scoring database and the features of the target fine-tuning corpus; use the historical security score corresponding to the historical corpus features with a similarity greater than the configured similarity threshold as the security score of the target fine-tuning corpus.
[0036] In one embodiment of this application, the method further includes: The analysis results are obtained by analyzing whether each entry in the target fine-tuning corpus contains sensitive words from the sensitive feature dictionary corresponding to the domain classification results; the analysis results represent the frequency of sensitive words contained in the target fine-tuning corpus. Based on the analysis results, the suspiciousness of each entry in the target fine-tuning corpus is calculated; Generate inference instructions for target corpus entries whose suspicion level exceeds the configured suspicion level threshold; The sensitive feature dictionary, target corpus entries, their suspiciousness, and inference instructions are input into each first target large model to obtain the first score and inference chain output by each first target large model; wherein, the inference chain is used to characterize the determination process of the first score; Obtain the rationality analysis results of the reasoning chain of the first major model of each objective to the first major model of other objectives; Based on the rationality analysis results of the primary models for different objectives, the weights of the primary models for each objective in the domain classification results corresponding to the target fine-tuning corpus are determined.
[0037] Corresponding to the above method, embodiments of this application also provide a fine-tuning corpus filtering device based on a large model, such as... Figure 3 As shown, the device includes: Acquisition unit 310 is used to acquire the target fine-tuning corpus; The extraction unit 320 is used to extract features from the target fine-tuning corpus to obtain corpus features; Matching unit 330 is used to match the security prompt words corresponding to the configured corpus features and multiple first target large models and multiple second target large models according to the corpus features; Scoring unit 340 is used to score the target fine-tuning corpus based on security prompt words using multiple matched first target large models to obtain the first score; The verification unit 350 is used to correct the first score based on the security prompt words using multiple matched second target large models to obtain the second score; The filtering unit 360 is used to obtain a security score for the target fine-tuning corpus based on multiple second scores. If the security score is less than the configured security threshold, the target fine-tuning corpus is filtered.
[0038] The functions of each functional unit in the large-model-based fine-tuning corpus filtering device provided in the above embodiments of this application can be implemented through the above-described method steps. Therefore, the specific working process and beneficial effects of each unit in the large-model-based fine-tuning corpus filtering device provided in the embodiments of this application will not be repeated here.
[0039] This application also provides an electronic device, such as... Figure 4 As shown, it includes a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440.
[0040] Memory 430 is used to store computer programs; When the processor 410 executes the program stored in the memory 430, it performs the following steps: Obtain the target fine-tuning corpus; Feature extraction is performed on the target fine-tuning corpus to obtain corpus features; Based on the corpus features, match the security prompt words corresponding to the configured corpus features with multiple primary target large models and multiple secondary target large models; The target fine-tuning corpus was scored based on security prompt words using multiple matched first-target large models, resulting in multiple first scores; Multiple first scores are corrected based on security prompt words using multiple matched second target large models to obtain multiple second scores; Based on multiple secondary scores, a safety score for the target fine-tuning corpus is obtained; If the security score is less than the configured security threshold, the target fine-tuning corpus will be filtered.
[0041] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.
[0042] The communication interface is used for communication between the aforementioned electronic devices and other devices.
[0043] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0044] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0045] The implementation methods and beneficial effects of the various components of the electronic device in the above embodiments for solving the problem can be found in [reference needed]. Figure 2 The steps in the illustrated embodiments are used to implement the electronic device. Therefore, the specific working process and beneficial effects of the electronic device provided in this application will not be repeated here.
[0046] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform any of the large-model-based fine-tuning corpus filtering methods described in the above embodiments.
[0047] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the large-model-based fine-tuning corpus filtering methods described in the above embodiments.
[0048] Those skilled in the art will understand that the embodiments in this application can be provided as methods, systems, or computer program products. Therefore, the embodiments in this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments in this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0049] This application describes embodiments of methods, apparatus (systems), and computer program products according to embodiments of this application with reference to flowchart illustrations and / or block diagrams. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0050] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0051] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0052] Although preferred embodiments have been described in this application, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of this application.
[0053] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of the embodiments of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of this application and its equivalents, then these modifications and variations are also intended to be included in the embodiments of this application.
Claims
1. A method for fine-tuning corpus filtering based on a large model, characterized in that, The method includes: Obtain the target fine-tuning corpus; Feature extraction is performed on the target fine-tuning corpus to obtain corpus features; Based on the corpus features, match the security prompt words corresponding to the configured corpus features with multiple first target large models and multiple second target large models; The target fine-tuning corpus is scored based on the security prompt words using multiple matched first target large models to obtain multiple first scores; Multiple first scores are corrected based on the security prompt words using multiple matched second target large models to obtain multiple second scores; Based on multiple second scores, a safety score for the target fine-tuning corpus is obtained; If the security score is less than the configured security threshold, then the target fine-tuning corpus is filtered.
2. The method as described in claim 1, characterized in that, Feature extraction is performed on the target fine-tuning corpus to obtain corpus features, including: Keyword extraction is performed on the target fine-tuning corpus to obtain the first feature; Feature extraction is performed on the target fine-tuning corpus to obtain the second feature; The first feature and the second feature are input into a pre-trained domain classification model to obtain the domain classification result corresponding to the target fine-tuning corpus; The first feature, the second feature, and the domain classification result are used as the corpus features of the target fine-tuning corpus.
3. The method as described in claim 2, characterized in that, Multiple first scores are corrected based on the security prompt words using multiple matched second-target large models, including: For any second target big model, the second target big model is used to correct multiple first scores based on the security prompt words to obtain the corrected score corresponding to the second target big model; The first scores are corrected using the corrected scores corresponding to the second target large model, resulting in multiple corrected first scores; Multiple second scores are obtained based on multiple first scores after correction of different second objective large models.
4. The method as described in claim 3, characterized in that, Based on multiple first scores corrected by different second-objective large models, multiple second scores are obtained, including: Obtain the weights of different second-objective large models on the domain classification results corresponding to the target fine-tuning corpus; Based on the weights of different secondary objective models in the classification results of the corresponding domains, multiple first scores after correction of different secondary objective models are weighted and summed to obtain multiple second scores.
5. The method as described in claim 1, characterized in that, After obtaining the security score of the target fine-tuning corpus, the method further includes: The corpus features of the target fine-tuning corpus and the security score are stored in the configured corpus scoring library.
6. The method as described in claim 5, characterized in that, After obtaining the corpus features, the method further includes: Calculate the similarity between the features of each historical corpus in the corpus scoring database and the features of the target fine-tuning corpus; The historical security score corresponding to the historical corpus features with a similarity greater than the configured similarity threshold is used as the security score of the target fine-tuning corpus.
7. A fine-tuning corpus filtering device based on a large model, characterized in that, The device includes: The acquisition unit is used to acquire the target fine-tuning corpus. The extraction unit is used to extract features from the target fine-tuning corpus to obtain corpus features; The matching unit is used to match the security prompt words corresponding to the configured corpus features with multiple first target large models and multiple second target large models according to the corpus features. The scoring unit is used to score the target fine-tuning corpus based on the security prompt words using multiple matched first target large models to obtain a first score; The verification unit is used to modify the first score based on the security prompt words using multiple matched second target large models to obtain a second score; A filtering unit is used to obtain a security score for the target fine-tuning corpus based on multiple second scores. If the security score is less than a configured security threshold, the target fine-tuning corpus is filtered.
8. An electronic device, characterized in that, The electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.