Financial large model fine-tuning interpretability auditing method, device and equipment and medium

Abstract

The invention discloses a financial large model fine-tuning interpretability auditing method, device and equipment and a medium, and relates to the field of artificial intelligence and financial science and technology. The method comprises the steps that in the fine tuning process based on the LoRA adapter, the influence of a training sample on parameters of the LoRA adapter is recorded in real time, and a time sequence influence matrix snapshot is generated; establishing a multi-level index and retrieval mechanism from model output to a training sample based on the snapshot; after audit query is received, reverse influence tracing analysis from output to a sample is executed, and an analysis result is obtained; generating a visual responsibility affiliation report; and automatically generating a formatted audit report based on the report and the compliance list. According to the method, the whole flow of the fine tuning process of the large financial model can be traced and audited, the root of the problem can be accurately positioned, the auditing efficiency is remarkably improved, and the strict compliance supervision requirement of the financial industry is met.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and financial technology, and in particular to a method, apparatus, device and medium for fine-tuning the interpretability audit of a large financial model. Background Technology

[0002] With the deep integration and application of artificial intelligence technologies, represented by large language models, in the financial field, fine-tuning pre-trained large models based on specific scenario data has become a key technical path for building specialized financial models such as credit approval, risk prediction, and intelligent investment advisory. However, the financial industry has strict requirements for compliance, transparency, and accountability, and the model decision-making process must be explainable and auditable. Traditional large model fine-tuning methods often treat the fine-tuning process as a black box, focusing only on the final model performance indicators while ignoring the inherent causal relationship between training data, parameter updates, and the final output behavior during the fine-tuning process.

[0003] Current research on model interpretability largely focuses on feature attribution analysis during the inference phase, such as explaining why a model produces a specific output for a single input through attention weights or gradient methods. However, these methods have significant limitations: they cannot trace back to the model during the fine-tuning training phase, specifically which training samples and in what ways influenced the evolution of model parameters, leading to the fine-tuning black box problem; when the fine-tuned model exhibits bias or errors, existing technologies struggle to accurately pinpoint the root cause samples from massive amounts of training data, resulting in a broken chain of responsibility; and there is a lack of systematic audit support that meets financial regulatory requirements and spans the entire lifecycle of model fine-tuning, making it impossible to automatically generate structured compliance evidence and reports.

[0004] These technical shortcomings present financial institutions with numerous challenges when applying fine-tuned large models, including uncontrollable model behavior, difficulty in managing compliance risks, and high audit costs. Therefore, the industry urgently needs a systematic solution that can deeply record the fine-tuning process, support fine-grained impact tracing, and automatically generate compliance audit reports to achieve explainable, auditable, and controllable fine-tuning activities for large financial models. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and medium for auditing the interpretability of fine-tuning of large financial models, aiming to achieve full traceability and auditability of the fine-tuning process of large financial models, thereby breaking through their black-box state and meeting the rigid requirements of financial regulators for the interpretability of model decisions.

[0006] In a first aspect, embodiments of the present invention provide a method for auditing the interpretability of fine-tuning of a large financial model, comprising: During the fine-tuning of the large financial model based on LoRA adapter, the specific impact of each training sample on the LoRA adapter parameters is recorded in real time, and a snapshot of the time-series impact matrix containing the relationship between the training samples and the LoRA adapter is generated and saved. Based on the snapshot of the time-series influence matrix, a multi-level index and fast retrieval mechanism is established from the token that triggers audit requirements in the output of the financial big model to the fine-tuning training samples. Receive audit traceability query requests for the output results of the financial big model, and perform reverse impact traceability analysis from the output results of the financial big model to the training samples based on the multi-level index and fast retrieval mechanism to obtain the analysis results; Based on the analysis results, a visual responsibility attribution report is generated, showing the complete impact propagation path from the problem output to the root training sample; Based on the visualized responsibility attribution report and the pre-set financial compliance checklist, a formatted compliance audit report that meets financial regulatory requirements is generated.

[0007] Secondly, embodiments of the present invention also provide a financial large-scale model fine-tuning interpretability auditing device, comprising: The matrix snapshot generation module is used to record the specific impact of each training sample on the LoRA adapter parameters in real time during the fine-tuning of the large financial model based on the LoRA adapter, and to generate and save a snapshot of the time-series impact matrix containing the relationship between the training samples and the LoRA adapter. The retrieval mechanism generation module is used to establish a multi-level index and fast retrieval mechanism based on the time-series influence matrix snapshot, from the token that triggers audit requirements in the output results of the financial big model to the fine-tuning training samples. The traceability analysis module is used to receive audit traceability query requests for the output results of the financial big model, and based on the multi-level index and fast retrieval mechanism, to perform reverse impact traceability analysis from the output results of the financial big model to the training samples, and obtain the analysis results. The attribution report generation module is used to generate a visual attribution report based on the analysis results, showing the complete impact propagation path from the problem output to the root training sample; The audit report generation module is used to generate a formatted compliance audit report that meets financial regulatory requirements based on the visualized responsibility attribution report and the pre-set financial compliance checklist.

[0008] Thirdly, embodiments of the present invention also provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned financial large model fine-tuning interpretability audit method.

[0009] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described financial large-scale model fine-tuning interpretability auditing method.

[0010] This invention provides a precise data foundation for retrospective analysis by recording and saving sample-level impact matrix snapshots in real time. By establishing a multi-level indexing and retrieval mechanism from output to training samples, it achieves rapid, targeted reverse tracing from problem results to root cause data. Finally, by automatically generating visualized accountability reports and formatted audit reports, it directly transforms the technical tracing results into audit evidence that meets regulatory requirements. The entire solution makes the previously invisible fine-tuning impact path fully transparent, searchable, and verifiable, significantly improving audit efficiency while ensuring tracing accuracy, fundamentally meeting the financial industry's rigid requirements for interpretability and compliance auditing throughout the model's entire lifecycle. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating the financial large-scale model fine-tuning interpretability auditing method provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the matrix snapshot generation process provided in an embodiment of the present invention; Figure 3 A schematic diagram of the traceability analysis process provided in an embodiment of the present invention; Figure 4 A schematic diagram illustrating the process of generating attribution reports provided in this embodiment of the invention; Figure 5 A schematic diagram illustrating the process of generating audit reports provided in this embodiment of the invention; Figure 6 A schematic diagram of the financial large-scale model fine-tuning interpretability auditing device provided in an embodiment of the present invention; Figure 7 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

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

[0014] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0015] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0016] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0017] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0018] The present invention will be described in detail below with reference to a specific embodiment of a credit approval model audit.

[0019] Specific Scenario: A commercial bank's intelligent credit approval model recently encountered a problem when processing a batch of loan applications from small and medium-sized enterprise (SME) owners in a specific region (e.g., Region A). The risk management team discovered that the model frequently produced rejection decisions. Because this phenomenon may imply regional discrimination, the bank, in accordance with regulatory requirements, needs to conduct an internal audit to determine whether such bias was introduced or amplified during the model's fine-tuning process. The audit must generate a traceable and explainable audit report to clarify the root cause of the problem and assign responsibility.

[0020] Please see Figure 1This invention provides a method for auditing the interpretability of fine-tuning of a large financial model, comprising steps S1-S5: S1. During the fine-tuning of the large financial model based on the LoRA adapter, the specific influence of each training sample on the LoRA adapter parameters is recorded in real time, and a snapshot of the time-series influence matrix containing the relationship between the training samples and the LoRA adapter is generated and saved.

[0021] In this embodiment, when constructing the credit approval model, historical approval data is used to fine-tune the general financial model using LoRA. When each training sample (e.g., a structured record containing customer ID, income, years of operation, industry, region, and final approval result label) is input into the model for forward and backward propagation, the matrix snapshot generation module begins its work. Specifically, the sample-adaptor association tracker in the module accurately calculates the contribution of the sample to the low-rank increment matrix of each LoRA adapter layer within the model during gradient descent updates. It doesn't simply record changes in loss values, but generates a fine-grained influence weight vector from the sample to each adapter by calculating gradient weighted sums, quantifying the degree to which the sample drives each adapter parameter to update in a specific direction. Next, the hierarchical influence propagation analyzer analyzes how this parameter increment propagates and interacts through the model's forward computation graph. For example, it tracks how a change in an adapter parameter affects the query-key matching score in the next attention mechanism, thereby altering the distribution of feature representations. Then, the token-level output mapper establishes the causal relationship between this underlying parameter perturbation and the final text generated by the model. For example, it analyzes and records that when a small change in an adapter parameter occurs due to the current sample, the probability of the corresponding token's logits increases measurably when the model decodes and generates keywords representing negative decisions (such as rejection, high risk). Finally, the aforementioned association weights, propagation path analysis, and token mapping relationships are packaged with information such as the current model parameter state, training epoch, and sample metadata to form a complete impact snapshot, which is then persistently stored. This mechanism of saving snapshots by epoch is like establishing a continuous and tamper-proof logbook for the entire learning process of a large financial model, laying a solid data foundation for accurate auditing and tracing at any subsequent point in time.

[0022] In a specific embodiment, such as Figure 2 As shown, the real-time recording of the specific impact of each training sample on the LoRA adapter parameters, generating and saving a snapshot of the time-series influence matrix containing the correlation between training samples and the LoRA adapter, includes steps S11-S14: S11. Quantize and record the influence weight of each training sample on the LoRA adapter parameters.

[0023] In this embodiment, the quantization process can specifically employ a weighted integral method based on gradients calculated during training. For the current training sample, the gradient vectors generated during backpropagation for each LoRA adapter parameter are captured. Then, based on the actual update step size of the optimizer (e.g., Adam), these gradients are integrated to calculate the net change of the sample in each parameter dimension, ultimately converging into a multi-dimensional influence weight tensor. This influence weight tensor clearly indicates whether the sample enhances or weakens the responsiveness of a specific adapter in the model to a specific type of feature (e.g., geographic information). This transforms the abstract sample influence into storable and comparable numerical evidence.

[0024] S12. Analyze the propagation path and intensity of the LoRA adapter parameter changes in each layer of the financial big data model.

[0025] In this embodiment, the propagation path analysis involves a localized virtual perturbation propagation within the neural network structure of the large-scale financial model. Using the parameter changes calculated in step S11 as the initial perturbation, a forward simulation is performed through the model's forward computation graph to observe how this perturbation is propagated, amplified, or attenuated layer by layer between different network layers (such as multi-head attention layers and feedforward network layers). For example, it can be traced that a change in a certain adapter parameter causes the model's attention layer to negative words to increase by 10% when processing fields involving region A, and this effect is further amplified in subsequent layers. The analysis results form a directed graph of influence propagation, where nodes represent network layers and edges represent the paths and strength coefficients of influence propagation. Its purpose is to reveal the intermediate chain of how micro-parameter changes trigger changes in macro-model behavior.

[0026] S13. Based on the propagation path and strength, establish a correlation mapping between the changes in the LoRA adapter parameters and the tokens that trigger audit requirements in the output of the financial big model.

[0027] In this embodiment, the correlation mapping is established by constructing a causal link model. The feature representation perturbation of the final layer (typically the output projection layer) obtained in step S12 is correlated with the logits probability changes of each token in the model vocabulary. Specifically, it identifies which parameter changes contribute most to the increase in the probability of specific key tokens in the output (e.g., the rejection token leading to the audit in this example) by calculating cross-attention or gradient correlation, and scores this contribution. For example, the mapping result might show that 65% of the causal factors leading to the increase in the rejection token probability during fine-tuning can be attributed to changes in several specific dimensions of the 7th layer LoRA adapter. This achieves a direct, quantitative causal connection from the adjustment of underlying parameters to the final controversial decision output.

[0028] S14. Save a snapshot of the time-series influence matrix containing the influence weights, propagation paths and association mappings according to the fine-tuning training rounds.

[0029] In this embodiment, the saving operation involves serializing the series of data structures generated in steps S11 to S13 (influence weight tensors, directed graphs of propagation paths, and token association mapping tables) along with metadata such as timestamps, training round numbers, and model checkpoints, and storing them in a specially designed time-series database. An independent snapshot file is generated at the end of each round. This design not only completely preserves the immediate impact of each sample in each training round, but more importantly, it retains the trajectory of the impact's evolution over time. This allows for rewinding to any historical training moment to view the model's state at that time and how it was shaped by the data, providing a complete spatiotemporal context for accurate tracing.

[0030] In a specific embodiment, the step of recording the specific impact of each training sample on the LoRA adapter parameters in real time, and generating and saving a snapshot of the time-series influence matrix containing the correlation between training samples and the LoRA adapter, further includes: Save a snapshot of the time-series influence matrix according to the fine-tuning training rounds.

[0031] In this embodiment, by ensuring that a complete influence matrix is ​​archived for each training round, a continuous and unbroken chain of evidence can be provided when facing post-audit. This is of crucial analytical value for determining whether biases in the audit model are concentratedly introduced in a specific round or gradually accumulated over long-term training.

[0032] S2. Based on the snapshot of the time-series influence matrix, establish a multi-level index and fast retrieval mechanism from the token that triggers audit requirements in the output results of the financial big model to the fine-tuning training samples.

[0033] In this embodiment, the retrieval mechanism generates a module that reads all saved snapshots of the time-series influence matrix. It first extracts and aggregates the token association mapping data established in step S13, forming a global inverted index. The first level of this index maps each token that may trigger an audit (e.g., rejection, approval, high risk, low credit, etc.) to all LoRA adapter modules and specific parameter dimensions that significantly influence its probability of occurrence. The second level index, based on the influence weights saved in step S11, reverse-maps each LoRA adapter module (and its parameter dimensions) to all training sample IDs and influence strengths that significantly influence it. Simultaneously, a third level index is established, constructing a nearest neighbor retrieval system based on the semantic features of the training samples (e.g., vectorized text descriptions or structured features). This multi-level index structure enables millisecond-level precise location of the source sample from hundreds of millions of potential relationships, tracing back from the output token. This solution transforms massive amounts of unstructured influence data into an efficient, queryable knowledge graph, providing underlying support for real-time auditing.

[0034] S3. Receive an audit traceability query request for the output results of the financial big model, and perform a reverse impact traceability analysis from the output results of the financial big model to the training samples based on the multi-level index and fast retrieval mechanism to obtain the analysis results.

[0035] In a specific embodiment, such as Figure 3 As shown, the step of performing a reverse impact analysis from the output of the large financial model to the training samples to obtain the analysis results includes steps S31-S34: S31. Starting with the token that triggers the audit requirement output by the financial big model, query the multi-level index to locate one or more LoRA adapter modules that contribute the most to it, and use them as key LoRA adapter modules.

[0036] In this embodiment, auditors input a specific case to be reviewed via an interface, such as a loan application from a technology company in region A that was rejected by the financial big data model. First, the financial big data model's output for this case is parsed to identify the key token that triggered the audit request (rejection in this example). Then, the traceability analysis module uses this token as the query key to retrieve the first-level index, quickly obtaining a list of LoRA adapter modules sorted by contribution. For example, the query results show that adapter module G, responsible for regional feature fusion, and adapter module H, responsible for risk keyword understanding, contributed more than 70% to the rejection decision and are therefore marked as key LoRA adapter modules. This solution reduces complex model decisions to a few interpretable, relatively well-defined internal modules, greatly simplifying the problem.

[0037] S32. Quantify the contribution weights of each LoRA adapter module located to the output of the financial large model.

[0038] In this embodiment, instead of simply adopting the ranking in the index, a more precise attribution calculation is performed. It may employ methods similar to integral gradients or SHAP values ​​to virtually ablate or perturb the parameters of the key adapter modules identified in step S31 for the current input case, observing changes in output probabilities to quantify the specific percentage contribution of each module to the final rejection decision. For example, it might be calculated that module G contributed 45% and module H contributed 28%. This precise quantification makes the responsibility analysis more objective and convincing, providing a basis for subsequently determining the primary and secondary order of root cause data.

[0039] S33. Based on the contribution weight and the association between the training samples and the LoRA adapter, retrieve and determine one or more training samples that have the greatest impact on the LoRA adapter module.

[0040] In this embodiment, based on the results of steps S31 and S32, the key adapter module (such as module G) with the highest contribution is selected as the primary tracing target. It queries the second-level index to find the top K training samples with the highest weighting on the parameters of module G across all rounds of historical fine-tuning. Simultaneously, the tracing results are weighted by combining the module's contribution weight. For example, the system might retrieve three training samples whose weighting on module G during fine-tuning far exceeds that of other samples. By examining the metadata of these samples, auditors discover that they all contain annotations with negative stereotypes about customers in region A (such as inductive descriptions of historical default cases in that region). This step directly traces from the functional modules within the model to specific external training data, completing the leap from "why" to "what."

[0041] S34. Based on semantic similarity, retrieve other potentially influential samples related to the identified training samples to supplement the tracing scope.

[0042] In this embodiment, to ensure the comprehensiveness of the audit and prevent the omission of samples with similar biases but weaker impacts, a third-level index is utilized. It uses the identified root sample (such as the three samples from the previous step) as the query vector and performs a nearest neighbor search in its semantic feature space to find all training samples that are highly similar in features. For example, five more samples might be found; although their direct impact weight on module G is not high, their content all involves negative qualitative descriptions of specific regions or industries. The system will add these samples as potential impact samples to the analysis report. This approach can reveal patterns of bias rather than isolated points, helping financial institutions to systematically evaluate and clean up training data.

[0043] S4. Based on the analysis results, generate a visual responsibility attribution report showing the complete impact propagation path from the problem output to the root training sample.

[0044] In this embodiment, the analysis results obtained in step S3 are integrated and visualized through the attribution report generation module. It constructs an interactive chart that clearly starts with the rejection decision being audited, connecting sequentially to key LoRA adapter modules (labeled with contribution weights) via arrows, then further connecting to specific root training samples (labeled with influence weights), and finally linking to other potentially similar samples with weaker connections. The chart is accompanied by textual explanations illustrating the logic behind each step. For example, the chart visually demonstrates that the current rejection decision primarily stems from adapter module G (regional bias, contributing 45%), and that the bias in module G is mainly injected by the three training samples numbered X, Y, and Z during the fifth round of fine-tuning. This visualization report presents the complex causal chains, originally deeply embedded within the model, to auditors and regulators in an intuitive and easily understandable way, significantly lowering the barrier to comprehension.

[0045] In a specific embodiment, such as Figure 4 As shown, the generation of a visual responsibility attribution report that demonstrates the complete impact propagation path from the problem output to the root training sample includes: S41. Display the output of the financial big model in a graphical manner, tracing the multi-level influence propagation chain back to the root training sample through the key LoRA adapter module.

[0046] In this embodiment, the graphical representation uses a layered flowchart or Sankey diagram. The top layer shows the input cases and the problem output (rejection). The middle layer is a schematic diagram of the model's internal structure, where key LoRA adapter modules are highlighted, accompanied by arrows pointing from the output to them and their weights. The bottom layer contains abbreviated information cards of the root training samples (such as sample ID, key features, and training epoch), with thicker arrows pointing from key modules to these samples, representing stronger influence. The entire chain comprehensively covers the entire process from the user-visible output to the model's internal black box, and then to tracing back to the external data source, making the source of responsibility for the model's behavior immediately clear.

[0047] S42. In the influence propagation chain, mark the contribution weight of the LoRA adapter module and the influence degree of the training samples.

[0048] In this embodiment, the contribution of each key adapter module node to the problem output is indicated as a percentage next to its node in the graph. The influence weight (possibly a standardized score) of each root training sample node on the corresponding key module parameter is indicated next to its node. This dual labeling not only identifies who is responsible but also indicates the extent of their responsibility, making the division of responsibility more precise and providing direct data support for subsequent model adjustments (whether to adjust module weights or delete sample data).

[0049] S43. Based on the aforementioned impact propagation chain and annotation information, generate a textual analysis report that clearly defines the source of responsibility for the output of the financial big data model. The source of responsibility includes at least specific training samples, sample characteristics, and fine-tuning operation batches.

[0050] In this embodiment, the textual analysis report systematically summarizes the findings from the visualization charts. It explicitly states that, upon auditing, the specific biases in the current large-scale financial model are primarily attributable to training samples (responsible entities) numbered X, Y, and Z. These samples collectively exhibit characteristics that strongly correlate region A with high risk (responsibility characteristics), and these effects mainly occur during the fifth round of fine-tuning training (responsibility point). The report may also analyze whether this bias is reinforced or weakened in subsequent training. This report transforms the results of the technology traceability into a clear and decision-making accountability document.

[0051] S5. Based on the visualized responsibility attribution report and the pre-set financial compliance checklist, generate a formatted compliance audit report that meets financial regulatory requirements.

[0052] In a specific embodiment, such as Figure 5 As shown, the generation of a formatted compliance audit report that meets financial regulatory requirements includes steps S51-S54: S51. Integrate the aforementioned visual responsibility attribution report and related text analysis report into a structured audit trajectory according to the timeline and operational logic.

[0053] In this embodiment, an audit report generation module creates a timeline view that records the following sequentially from the initiation of the audit task: received cases to be audited, executed trace queries, located key modules, discovered root cause samples, and generated visual analysis. Simultaneously, it embeds every key operation performed by technical personnel on the interface (such as adjusting trace thresholds or viewing sample details) as an operation log into the timeline. This makes the entire audit process transparent and verifiable, meeting regulatory requirements for the standardization of audit procedures.

[0054] S52. Based on the audit trajectory, automatically determine whether there are biases, errors or non-compliance in the output of the financial big data model, and associate it with the specific training data responsible party.

[0055] In this embodiment, the built-in rule engine automatically determines the audit results based on predefined indicators (such as whether the root sample contains protected features and whether the bias contribution exceeds a threshold). For example, if the output contains potential regional discriminatory bias, reaching a medium risk level, it will automatically associate the suggested responsible party (such as the data annotation team or a business department) based on the annotator ID or data source department that may be recorded in the sample metadata. This achieves a seamless transition from technical problem discovery to clear management responsibility.

[0056] S53. Invoke the pre-set compliance checklist and verify item by item whether the key nodes in the audit trajectory comply with financial regulatory requirements.

[0057] In this embodiment, a configurable compliance checklist is loaded, the contents of which are based on regulatory documents. It checks item by item: whether the complete fine-tuning process is recorded; whether decision traceability is achieved; whether potential discriminatory biases are identified; whether problematic data is located; and whether audit logs are maintained. For each checklist item, a verification conclusion is automatically generated based on yes / no evidence provided by the current audit trajectory. This is essentially an automatic assessment of whether the audit itself is compliant.

[0058] S54. Integrate the audit trail, liability determination results, and compliance verification results to generate and export a complete compliance audit report document in the specified format.

[0059] In this embodiment, the structured audit trail (including visual charts), automated liability determination report, and compliance clause verification form are formatted and synthesized according to fixed templates required by financial institutions or regulatory agencies. The final product is a detailed, complete, and formatted formal audit report document that can be directly used for internal archiving or submitted for regulatory review. This solution transforms complex technical audit work into a standardized, high-quality compliance deliverable, improving audit efficiency from days to minutes.

[0060] In a specific embodiment, the financial big data model fine-tuning interpretability auditing method further includes: During the process of receiving the audit traceability query request, executing the reverse impact traceability analysis, generating the visualized responsibility attribution report and the formatted compliance audit report, access control based on role permissions, operation log recording and automatic desensitization of sensitive information are performed.

[0061] In this embodiment, considering that the audit process involves sensitive model internal parameters, training data, and customer information, a secure access control module intervenes throughout the entire process. For example, only users with the auditor role can initiate traceability queries and view the complete root sample data; financial large model developers can only see partial aggregation results. All user login, query, and export operations are recorded in detail by the operation logger, forming an irrefutable audit log. Simultaneously, when reports or interfaces need to display specific training sample content or customer case information, the data anonymization processor automatically identifies sensitive fields such as ID card numbers, mobile phone numbers, and precise addresses, and replaces them with asterisks or masks, ensuring that while meeting audit requirements, data privacy protection regulations are strictly followed. This series of measures ensures the secure and compliant operation of the system of this invention in a highly sensitive financial environment.

[0062] like Figure 6 As shown, this embodiment of the invention also provides a financial large-scale model fine-tuning interpretability auditing device, comprising: The matrix snapshot generation module 10 is used to record the specific influence of each training sample on the LoRA adapter parameters in real time during the fine-tuning of the large financial model based on the LoRA adapter, and to generate and save a snapshot of the time-series influence matrix containing the relationship between the training samples and the LoRA adapter. The retrieval mechanism generation module 20 is used to establish a multi-level index and fast retrieval mechanism based on the time-series influence matrix snapshot, from the token that triggers audit requirements in the output results of the financial big model to the fine-tuning training samples. The traceability analysis module 30 is used to receive audit traceability query requests for the output results of the financial big model, and based on the multi-level index and fast retrieval mechanism, to perform reverse influence traceability analysis from the output results of the financial big model to the training samples, and obtain analysis results. The attribution report generation module 40 is used to generate a visual attribution report based on the analysis results, showing the complete impact propagation path from the problem output to the root training sample; The audit report generation module 50 is used to generate a formatted compliance audit report that meets financial regulatory requirements based on the visualized responsibility attribution report and the pre-set financial compliance checklist.

[0063] In a specific embodiment, the matrix snapshot generation module 10 is specifically used for: Quantitatively record the weight of each training sample's influence on the LoRA adapter parameters; Analyze the propagation path and intensity of the LoRA adapter parameter changes in each layer of the financial big data model; Based on the propagation path and strength, establish a correlation mapping between the changes in the LoRA adapter parameters and the tokens that trigger audit requirements in the output of the financial big model; Save a snapshot of the time-series influence matrix containing the influence weights, propagation paths, and association mappings for each fine-tuning training round.

[0064] In a specific embodiment, the traceability analysis module 30 is specifically used for: Starting with the token that triggers the audit requirement output by the financial big model, query the multi-level index to locate one or more LoRA adapter modules that contribute the most to it, and use them as key LoRA adapter modules. The contribution weights of each LoRA adapter module identified by the quantification to the output of the large financial model are determined. Based on the contribution weights and the association between the training samples and the LoRA adapter, one or more training samples that have the greatest impact on the LoRA adapter module are retrieved and identified. Semantic similarity is used to retrieve other potentially influential samples that are related to the identified training samples to supplement the tracing scope.

[0065] In a specific embodiment, the attribution report generation module 40 is specifically used for: The output of the financial model is displayed graphically, and the multi-level influence propagation chain is traced back to the root training sample through the key LoRA adapter module. In the influence propagation chain, the contribution weight of the LoRA adapter module and the degree of influence of the training samples are marked; Based on the aforementioned impact propagation chain and annotation information, a textual analysis report is generated that clearly defines the source of responsibility for the output of the financial big data model. The source of responsibility includes at least specific training samples, sample characteristics, and fine-tuning operation batches.

[0066] In a specific embodiment, the audit report generation module 50 is specifically used for: The visualized responsibility attribution report and related textual analysis reports are integrated into a structured audit trail according to the timeline and operational logic; Based on the audit trajectory, the system automatically determines whether the output of the financial big data model contains biases, errors, or non-compliance, and links it to the specific party responsible for the training data. The pre-set compliance checklist is invoked to verify whether the key nodes in the audit trajectory comply with financial regulatory requirements item by item; Integrate the audit trail, liability determination results, and compliance verification results to generate and export a complete compliance audit report document in the specified format.

[0067] In a specific embodiment, the financial large-scale model fine-tuning interpretability auditing device further includes: a secure access control module, specifically used for: During the process of receiving the audit traceability query request, executing the reverse impact traceability analysis, generating the visualized responsibility attribution report and the formatted compliance audit report, access control based on role permissions, operation log recording and automatic desensitization of sensitive information are performed.

[0068] In a specific embodiment, the matrix snapshot generation module 10 is further configured to: Save a snapshot of the time-series influence matrix according to the fine-tuning training rounds.

[0069] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned financial large model fine-tuning interpretability audit device and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0070] The aforementioned financial big data model fine-tuning interpretability auditing device can be implemented as a computer program, which can, for example... Figure 7 It runs on the computer device shown.

[0071] Please see Figure 7 , Figure 7 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server. The terminal can be an electronic device with communication functions, such as a smartphone, tablet, laptop, desktop computer, personal digital assistant, or wearable device. The server can be a standalone server or a server cluster composed of multiple servers.

[0072] See Figure 7 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0073] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it causes the processor 502 to perform a financial large-scale model fine-tuning interpretability auditing method.

[0074] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0075] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a financial big model fine-tuning interpretability auditing method.

[0076] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0077] The processor 502 is used to run a computer program 5032 stored in the memory to perform the following steps: S1. During the fine-tuning of the large financial model based on LoRA adapter, the specific impact of each training sample on the LoRA adapter parameters is recorded in real time, and a snapshot of the time-series impact matrix containing the relationship between the training samples and the LoRA adapter is generated and saved. S2. Based on the snapshot of the time-series influence matrix, establish a multi-level index and fast retrieval mechanism from the token that triggers audit requirements in the output results of the financial big model to the fine-tuning training samples; S3. Receive an audit traceability query request for the output results of the financial big model, and based on the multi-level index and fast retrieval mechanism, perform a reverse impact traceability analysis from the output results of the financial big model to the training samples to obtain the analysis results; S4. Based on the analysis results, generate a visual responsibility attribution report showing the complete impact propagation path from the problem output to the root cause training sample; S5. Based on the visualized responsibility attribution report and the pre-set financial compliance checklist, generate a formatted compliance audit report that meets financial regulatory requirements.

[0078] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, 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, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0079] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program may be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0080] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform the following steps: S1. During the fine-tuning of the large financial model based on LoRA adapter, the specific impact of each training sample on the LoRA adapter parameters is recorded in real time, and a snapshot of the time-series impact matrix containing the relationship between the training samples and the LoRA adapter is generated and saved. S2. Based on the snapshot of the time-series influence matrix, establish a multi-level index and fast retrieval mechanism from the token that triggers audit requirements in the output results of the financial big model to the fine-tuning training samples; S3. Receive an audit traceability query request for the output results of the financial big model, and based on the multi-level index and fast retrieval mechanism, perform a reverse impact traceability analysis from the output results of the financial big model to the training samples to obtain the analysis results; S4. Based on the analysis results, generate a visual responsibility attribution report showing the complete impact propagation path from the problem output to the root cause training sample; S5. Based on the visualized responsibility attribution report and the pre-set financial compliance checklist, generate a formatted compliance audit report that meets financial regulatory requirements.

[0081] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code.

[0082] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0083] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0084] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0085] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0086] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0087] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0088] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fine-tuning the interpretability audit of a large financial model, characterized in that, include: During the fine-tuning of the large financial model based on LoRA adapter, the specific impact of each training sample on the LoRA adapter parameters is recorded in real time, and a snapshot of the time-series impact matrix containing the relationship between the training samples and the LoRA adapter is generated and saved. Based on the snapshot of the time-series influence matrix, a multi-level index and fast retrieval mechanism is established from the token that triggers audit requirements in the output of the financial big model to the fine-tuning training samples. Receive audit traceability query requests for the output results of the financial big model, and perform reverse impact traceability analysis from the output results of the financial big model to the training samples based on the multi-level index and fast retrieval mechanism to obtain the analysis results; Based on the analysis results, a visual responsibility attribution report is generated, showing the complete impact propagation path from the problem output to the root training sample; Based on the visualized responsibility attribution report and the pre-set financial compliance checklist, a formatted compliance audit report that meets financial regulatory requirements is generated.

2. The financial large-scale model fine-tuning interpretability auditing method according to claim 1, characterized in that, The process involves real-time recording of the specific impact of each training sample on the LoRA adapter parameters, generating and saving a snapshot of the time-series impact matrix containing the correlation between training samples and the LoRA adapter, including: Quantitatively record the weight of each training sample's influence on the LoRA adapter parameters; Analyze the propagation path and intensity of the LoRA adapter parameter changes in each layer of the financial big data model; Based on the propagation path and strength, establish a correlation mapping between the changes in the LoRA adapter parameters and the tokens that trigger audit requirements in the output of the financial big model; Save a snapshot of the time-series influence matrix containing the influence weights, propagation paths, and association mappings for each fine-tuning training round.

3. The financial large-scale model fine-tuning interpretability auditing method according to claim 1, characterized in that, The analysis involves tracing the reverse impact from the output of the large financial model to the training samples, yielding the following results: Starting with the token that triggers the audit requirement output by the financial big model, query the multi-level index to locate one or more LoRA adapter modules that contribute the most to it, and use them as key LoRA adapter modules. The contribution weights of each LoRA adapter module identified by the quantification to the output of the large financial model are determined. Based on the contribution weights and the association between the training samples and the LoRA adapter, one or more training samples that have the greatest impact on the LoRA adapter module are retrieved and identified. Semantic similarity is used to retrieve other potentially influential samples that are related to the identified training samples to supplement the tracing scope.

4. The financial large-scale model fine-tuning interpretability auditing method according to claim 3, characterized in that, The generated, visual responsibility attribution report, which shows the complete impact propagation path from the problem output to the root training sample, includes: The output of the financial model is displayed graphically, and the multi-level influence propagation chain is traced back to the root training sample through the key LoRA adapter module. In the influence propagation chain, the contribution weight of the LoRA adapter module and the degree of influence of the training samples are marked; Based on the aforementioned impact propagation chain and annotation information, a textual analysis report is generated that clearly defines the source of responsibility for the output of the financial big data model. The source of responsibility includes at least specific training samples, sample characteristics, and fine-tuning operation batches.

5. The financial large-scale model fine-tuning interpretability auditing method according to claim 1, characterized in that, The generation of a formatted compliance audit report that meets financial regulatory requirements includes: The visualized responsibility attribution report and related textual analysis reports are integrated into a structured audit trail according to the timeline and operational logic; Based on the audit trajectory, the system automatically determines whether the output of the financial big data model contains biases, errors, or non-compliance, and links it to the specific party responsible for the training data. The pre-set compliance checklist is invoked to verify whether the key nodes in the audit trajectory comply with financial regulatory requirements item by item; Integrate the audit trail, liability determination results, and compliance verification results to generate and export a complete compliance audit report document in the specified format.

6. The financial large-scale model fine-tuning interpretability auditing method according to claim 1, characterized in that, Also includes: During the process of receiving the audit traceability query request, executing the reverse impact traceability analysis, generating the visual responsibility attribution report and the formatted compliance audit report, access control based on role permissions, operation log recording and automatic desensitization of sensitive information are performed.

7. The financial large-scale model fine-tuning interpretability auditing method according to claim 1, characterized in that, The method of recording the specific impact of each training sample on the LoRA adapter parameters in real time, generating and saving a snapshot of the time-series influence matrix containing the correlation between training samples and the LoRA adapter, also includes: Save a snapshot of the time-series influence matrix according to the fine-tuning training rounds.

8. A financial large-scale model fine-tuning interpretability auditing device, characterized in that, include: The matrix snapshot generation module is used to record the specific impact of each training sample on the LoRA adapter parameters in real time during the fine-tuning of the large financial model based on the LoRA adapter, and to generate and save a snapshot of the time-series impact matrix containing the relationship between the training samples and the LoRA adapter. The retrieval mechanism generation module is used to establish a multi-level index and fast retrieval mechanism based on the time-series influence matrix snapshot, from the token that triggers audit requirements in the output results of the financial big model to the fine-tuning training samples. The traceability analysis module is used to receive audit traceability query requests for the output results of the financial big model, and based on the multi-level index and fast retrieval mechanism, to perform reverse impact traceability analysis from the output results of the financial big model to the training samples, and obtain the analysis results. The attribution report generation module is used to generate a visual attribution report based on the analysis results, showing the complete impact propagation path from the problem output to the root training sample; The audit report generation module is used to generate a formatted compliance audit report that meets financial regulatory requirements based on the visualized responsibility attribution report and the pre-set financial compliance checklist.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the financial large model fine-tuning interpretability auditing method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the financial large model fine-tuning interpretability auditing method as described in any one of claims 1 to 7.

Images