Method and device for generating model for auditing bidding and tendering documents

Abstract

The invention discloses a method and device for generating a model for reviewing bidding documents, and the method comprises the steps: obtaining a training sample data set, and carrying out the i-round training of a large language model injected into a low-rank adapter through the training sample data set, a gradient vector generated by each piece of training sample data for the low-rank adapter in the last j round of training is obtained; according to the comprehensive training gradient of each piece of training sample data, clustering the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain each sample gradient group; screening out target training sample data according to the similarity between each comprehensive training gradient in the sample gradient group and a preset reference gradient corresponding to the examination conclusion label; training the large language model by using the target training sample data to obtain a target model for reviewing the bidding and tendering document; according to the method, the bidding document review model which is adaptive to the bidding scene and is accurate and efficient in screening can be generated.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for generating models for reviewing bidding documents. Background Technology

[0002] In the bidding and tendering field, the review of the compliance and reasonableness of bidding documents is a core element in ensuring the fair and efficient conduct of the activity. With the expansion of bidding and tendering business and the increasing complexity of documents, traditional manual review methods are no longer sufficient to meet the demand for rapid review of massive amounts of documents. Furthermore, manual review is susceptible to subjective factors, resulting in low accuracy in identifying hidden risks such as "implicit exclusivity clauses, conflicting qualification requirements, and contradictory technical parameters." This has led to long-standing disputes and compliance risks in the industry caused by oversights in bid document review.

[0003] To address the aforementioned issues, Large Language Models (LLMs) are widely used in the construction of intelligent review models for bidding documents. The quality of the training data directly determines the performance of the review model. Currently, the techniques for selecting training data for LLMs mainly fall into two categories: one relies on expert-defined textual statistical features (such as length, perplexity, and lexical complexity) to construct static scoring rules, iteratively calculating the "gain" of samples on the existing dataset to prioritize samples that enhance diversity. However, this method deviates significantly from the core performance of the model in bidding review scenarios (such as risk identification accuracy and cross-project generalization ability). Optimizing these metrics cannot ensure that the selected data systematically improves the model's review effectiveness. The other category is based on trainable LLM methods, utilizing the model's own capabilities for dynamic evaluation, such as the IFD method. The ratio of "conditional answer score" to "direct answer score" is used to assess the difficulty of following instructions. The Nuggets framework compares the scoring differences of the model under "zero sample" and "single sample" settings to identify high-value data. The MoDS method iteratively filters data from three dimensions: quality, coverage, and necessity. However, these methods generally require repeated calls or fine-tuning of large-scale pre-trained models, resulting in huge computational overhead and time costs. Moreover, their effectiveness is highly dependent on high-quality reward models or evaluators. The construction of such evaluation models is inherently difficult and costly, and performance defects will be directly transmitted to the screening results, making it difficult to achieve efficient and automated large-scale deployment.

[0004] Furthermore, existing methods often employ full-parameter fine-tuning during model training, which not only further increases computational costs but also easily leads to model overfitting, reducing its generalization ability across different bidding projects.

[0005] Therefore, there is an urgent need for a review model generation method that is adapted to bidding scenarios and can accurately and efficiently select data. This method can solve the problems of disconnect between training sample data selection and review tasks, high implementation costs, and insufficient model generalization ability in existing technologies, and provide high-quality data support and efficient training path for the construction of intelligent review models for bidding documents. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for generating a model for reviewing bidding documents. This invention was completed with the support of the project titled "Key Technologies for Construction and Application of Large Models of Building Engineering" and project number "2024YFC3811200". It can generate a bidding document review model that is adapted to bidding scenarios and can screen accurately and efficiently.

[0007] According to one aspect of the present invention, a method for generating a model for reviewing tender documents is provided, the method comprising: Obtain the original sample dataset in the bidding and tendering field, and obtain the training sample dataset from the original sample dataset according to a preset sampling ratio; wherein, the sample data includes: bidding and tendering documents, preset review rules, and review conclusion labels, and the review conclusion labels represent the compliance status of the bidding and tendering documents under the preset review rules; The large language model with the low-rank adapter injected is trained i times using the training sample dataset, and the gradient vector generated by each training sample data in the reciprocal j-th training round is obtained for the low-rank adapter; wherein, in each training round, only the parameters in the low-rank adapter are adjusted, and the parameters in the large language model are not adjusted. The comprehensive training gradient of each training sample is calculated based on the j gradient vectors generated in the jth round of training for each training sample. Then, the training sample data with the same review conclusion label and their corresponding comprehensive training gradients are clustered to obtain sample gradient groups for each type of review conclusion label. For each type of review conclusion label, the sample gradient grouping is used to select target training sample data from the sample gradient group based on the similarity between each comprehensive training gradient in the sample gradient group and the preset benchmark gradient corresponding to the review conclusion label. The large language model is trained using the target training sample data to obtain a target model for reviewing bidding documents.

[0008] Optionally, obtaining the original sample dataset in the bidding and tendering field, and obtaining the training sample dataset from the original sample dataset according to a preset sampling ratio, includes: Steps for using the ratio: Obtain multiple preset sampling ratios, and use each sampling ratio in ascending order; Model training steps: Obtain a sample dataset from the original sample dataset according to the currently used sampling ratio, and train the preset base model using the sample dataset to obtain the result model; Model validation steps: Validate the resulting model using a preset validation dataset. If the validation passes, set the sampled dataset as the training dataset. If the validation fails, use the next sampling ratio and re-execute the model training steps.

[0009] Optionally, the step of training the large language model with the injected low-rank adapter using the training sample dataset for i rounds, and obtaining the gradient vector generated by each training sample data with respect to the low-rank adapter in the reciprocal j rounds of training, includes: Obtain a test sample dataset with the same amount of data as the training sample dataset from the original sample dataset; The large language model with low-rank adapter injected is trained i times using the training sample dataset, and j intermediate models are obtained in the reciprocal j rounds of training. The j training gradient vectors of each training sample data with respect to the j intermediate models are calculated using the forward and backward propagation algorithms. The test sample dataset is input into the j intermediate models respectively, and the forward and backpropagation algorithms are used to calculate the j test gradient vectors of each test sample dataset with respect to the j intermediate models.

[0010] Optionally, the step of calculating the comprehensive training gradient of each training sample data based on the j gradient vectors generated in the reciprocal j rounds of training, and clustering the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain sample gradient groupings for each type of review conclusion label includes: Calculate the average training gradient vectors of j training gradient vectors generated in the last j rounds of training for each training sample data, and input the average training gradient into the projection algorithm to obtain the comprehensive training gradient of each training sample data; Calculate the average value of the j test gradient vectors generated by each test sample data in the reciprocal j rounds of training, and input the average value into the projection algorithm to obtain the comprehensive test gradient of each test sample data; Cluster the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain the training sample gradient grouping of each type of review conclusion label. Cluster the test sample data with the same review conclusion label and the corresponding comprehensive test gradient to obtain test sample gradient groups for each type of review conclusion label.

[0011] Optionally, the sample gradient grouping for each type of review conclusion label involves selecting target training sample data from the sample gradient group based on the similarity between each comprehensive training gradient within the sample gradient group and the preset benchmark gradient corresponding to the review conclusion label, including: The average comprehensive test gradient is calculated based on each comprehensive test gradient in the test sample gradient group of each type of review conclusion label, and is used as the benchmark gradient of the corresponding type of review conclusion label. Iterate through the training sample gradient groups of each type of review conclusion label, calculate the cosine similarity between each comprehensive training gradient in the currently traversed training sample gradient group and the baseline gradient of the corresponding type of review conclusion label, and set the training sample data corresponding to the comprehensive training gradient in the currently traversed training sample gradient group whose cosine similarity is greater than the preset similarity threshold as the target training sample data.

[0012] Optionally, training the large language model using the target training sample data to obtain a target model for reviewing bidding documents includes: The background information, project information, and content information of a bidding document are extracted from the training sample data of each target, and the review rules are preset, with risk labels and no risk labels. The background information, project information, content information, and preset review rules of the bidding documents in each target training sample data are used as input to the large language model, and the risky label and the risk-free label are used as the standard output of the large language model to train the large language model to obtain the target model.

[0013] To achieve the above objectives, the present invention also provides an apparatus for generating a model for reviewing bidding documents, the apparatus comprising: The acquisition module is used to acquire the original sample dataset in the bidding and tendering field, and to acquire the training sample dataset from the original sample dataset according to a preset sampling ratio; wherein, the sample data includes: bidding and tendering documents, preset review rules, and review conclusion labels, and the review conclusion labels represent the compliance status of the bidding and tendering documents under the preset review rules; The training module is used to train the large language model injected with the low-rank adapter using the training sample dataset for i rounds, and to obtain the gradient vector generated by each training sample data in the reciprocal j rounds of training with respect to the low-rank adapter; wherein, in each round of training, only the parameters in the low-rank adapter are adjusted, and the parameters in the large language model are not adjusted. The calculation module is used to calculate the comprehensive training gradient of each training sample data based on the j gradient vectors generated in the jth round of training for each training sample data, and to cluster the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain sample gradient groupings of various review conclusion labels. The grouping module is used to group samples for each type of review conclusion label. Based on the similarity between each comprehensive training gradient in a sample gradient group and the preset benchmark gradient corresponding to the review conclusion label, the target training sample data is selected from the sample gradient group. The generation module is used to train the large language model using the target training sample data to obtain a target model for reviewing bidding documents.

[0014] Optionally, the training module is specifically used for: Obtain a test sample dataset with the same amount of data as the training sample dataset from the original sample dataset; The large language model with low-rank adapter injected is trained i times using the training sample dataset, and j intermediate models are obtained in the reciprocal j rounds of training. The j training gradient vectors of each training sample data with respect to the j intermediate models are calculated using the forward and backward propagation algorithms. The test sample dataset is input into the j intermediate models respectively, and the forward and backpropagation algorithms are used to calculate the j test gradient vectors of each test sample dataset with respect to the j intermediate models.

[0015] To achieve the above objectives, the present invention also provides a computer device, which specifically includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above for generating a model for reviewing tender documents.

[0016] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above for generating a model for reviewing tender documents.

[0017] The present invention provides a method and apparatus for generating a model for reviewing bidding documents. This method acquires an original sample dataset from the bidding field and obtains a training sample dataset from the original dataset according to a preset sampling ratio. It employs a parameter-efficient low-rank adapter for initial model training and gradient acquisition, requiring only a small number of parameter updates, enabling rapid completion of multiple iterations and gradient analysis with limited computing resources. Training sample data with the same review conclusion label and their corresponding comprehensive training gradients are clustered. For each group of sample gradients corresponding to a review conclusion label, target training sample data is selected from the group based on the similarity between each comprehensive training gradient within the sample gradient group and the preset benchmark gradient corresponding to that review conclusion label. This selection based on gradient similarity within the same review conclusion category ensures the consistency of the selected samples in teaching specific review skills and effectively filters out noise and contradictory samples. The target training sample data is used to train the large language model to obtain a target model for reviewing bidding documents. The entire selection process closely aligns with the characteristics of data structuring, rule complexity, and diversified review dimensions in the bidding field, ultimately producing a "core review knowledge base" rich in high-quality samples. The model trained based on this knowledge base has a decision-making process that better reflects professional review logic, and the results are more credible and traceable, providing reliable technical support for the fairness, efficiency and compliance of bidding activities. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is an optional flowchart illustrating a method for generating a model for reviewing tender documents, as provided in Embodiment 1. Figure 2 This is a flowchart illustrating the process of filtering gradient vectors provided in Example 1. Figure 3 This is a schematic diagram of an optional component structure of the apparatus for generating a model for reviewing tender documents, as provided in Embodiment 2. Figure 4 This is a schematic diagram of an optional hardware structure for the computer device provided in Embodiment 3. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0020] Example 1 This invention provides a method for generating a model for reviewing bidding documents, such as... Figure 1 As shown, the method specifically includes the following steps: Step S101: Obtain the original sample dataset in the bidding and tendering field, and obtain the training sample dataset from the original sample dataset according to the preset sampling ratio; wherein, the sample data includes: bidding and tendering documents, preset review rules, and review conclusion labels, and the review conclusion labels represent the compliance status of the bidding and tendering documents under the preset review rules.

[0021] In this embodiment, the original sample data includes the following parts: background information, used to define the roles, objectives, and basic specifications of the review task; bidding project information, including key project metadata such as project name, project number, bidding unit, and project type; bidding document content, including the specific bidding document text to be reviewed, covering core chapters such as technical specifications, qualification requirements, and evaluation methods; review rule information, generally including structured review basis and guidance, specifically including review rule name: the standardized identifier of the rule, and review rule details: a detailed description and scope of application of the rule, the review rule being: review steps, clarifying the operational process for standard execution; review basis, listing relevant laws, regulations, policy documents, or industry standards; negative cases, showing typical examples of errors that violate the rule; positive cases, providing standard reference samples that conform to the specifications; the task and output format is the specific review task and expected structure output of the bidding document review model: risk level, risk description, rectification suggestions, and review conclusion labels that identify the review sample data: "risky" or "no risk", representing the compliance status of the bidding document under the preset review rules, used for supervised learning and subsequent gradient analysis. A portion of the original sample dataset of bidding documents is selected as the training sample dataset. This sample dataset organically integrates domain knowledge, legal basis, and case guidance, providing the model with rich contextual information and judgment criteria, ensuring the accuracy and interpretability of the review results.

[0022] Step S102: Use the training sample dataset to train the large language model with the injected low-rank adapter for i rounds, and obtain the gradient vector generated by each training sample data in the reciprocal j rounds of training for the low-rank adapter; wherein, in each round of training, only the parameters in the low-rank adapter are adjusted, and the parameters in the large language model are not adjusted.

[0023] In this embodiment, the bidding document text is characterized by its complex structure, dense use of technical terms, strong logical connections between clauses, and compliance requirements implicit in deep semantics. Existing methods, such as the Adam model, utilize adaptive learning rate optimization algorithms, combining the ideas of momentum and gradient second-moment estimation. However, they struggle to accurately identify high-quality samples crucial for the model's learning of "compliance judgment logic," easily leading to problems such as weak generalization ability, low recognition rate for specific risk types (e.g., bid rigging), and uninterpretable decision-making basis in the trained review model. Therefore, this embodiment employs a large language model with an injected LoRA adapter (low-rank adapter). During training, only a small number of low-rank parameters in the LoRA adapter are updated, while all original parameters of the large language model are frozen. The large language model is trained for i rounds using the training sample dataset. The gradient vector is calculated based on the output results, and the parameters of the LoRA adapter are adjusted after each round of training using the gradient vector and stochastic gradient descent algorithm to obtain a large language model with stable LoRA adapter parameters. After the large language model is relatively stable, the gradient vectors from each of the last j rounds are used. Here, i and j are both positive integers, and i > j. Specifically, in this embodiment, the gradient vectors from the last three rounds are taken. Training only the low-rank adapter parameters greatly reduces computational and storage overhead, enabling rapid iteration of large language models and adaptation to the bidding and tendering field. Collecting the gradients from the last three rounds of training allows us to obtain the fine-tuning requirements of each sample for the "review logic" when the model is in a relatively stable state. This gradient signal has less noise and is more representative, providing high-quality feature input for subsequent value assessment.

[0024] Step S103: Calculate the comprehensive training gradient of each training sample data based on the j gradient vectors generated in the jth round of training for each training sample data, and cluster the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain sample gradient groups of various review conclusion labels.

[0025] In this embodiment, multiple gradient vectors generated during the training of the large language model using the sample data from the last three rounds are used to calculate the comprehensive training gradient for each sample data. The review conclusion label for each sample data is either "risky" or "risk-free." In the bidding scenario, the "risky" category includes bidding document samples that violate review rules (corresponding to negative case features), while the "risk-free" category includes bidding document samples that comply with regulations (corresponding to positive case features). The review conclusion label and comprehensive training gradient of each sample data are clustered to obtain training sample gradient groups for the "risky" and "risk-free" label classes of the training sample data. Furthermore, a set of test sample datasets is used to train the large language model generated from the last three rounds of training, obtaining test gradient vectors as a control group for the training sample dataset, thereby generating sample gradient groups for the "risky" and "risk-free" label classes of the test sample data. This classification process ensures that the model can learn high-quality, representative optimization directions in both types of tasks, avoiding interference from class imbalance in gradient analysis. Gradient vector fusion smooths out random fluctuations during training, obtaining a more stable influence representation for each sample. Grouping by "review conclusion label" ensures that all subsequent comparisons and screenings are conducted under the same semantic goal and optimization direction. This effectively prevents the gradient signals of "risk-free" samples and "risky" samples from interfering with each other or being misjudged due to opposite goals, and greatly improves the accuracy and robustness of the evaluation process.

[0026] Step S104: For each type of review conclusion label, the sample gradient is grouped, and the target training sample data is selected from the sample gradient group based on the similarity between each comprehensive training gradient in the sample gradient group and the preset benchmark gradient corresponding to the review conclusion label.

[0027] In this embodiment, gradient groups are defined for both "risky" and "risk-free" training and test samples. The similarity between the overall training gradient within each group and a preset benchmark gradient is determined to be higher than a similarity threshold. If so, the sample data corresponding to the highly similar overall training gradient is used as the target sample data, resulting in "risky" test sample data, "risky" training sample data, "risk-free" test sample data, and "risk-free" training sample data. The selected sample data is then stored in a database. By calculating the similarity between the sample gradient and the "benchmark gradient direction," high-quality samples whose optimization guidance direction is most consistent with the overall consensus direction of the category can be accurately identified. This is equivalent to selecting the most representative and clearly signaling core cases for teaching a specific review skill (such as identifying bid-rigging or confirming compliance) to the model, thereby automatically constructing a high-purity "core review knowledge base."

[0028] Step S105: Train the large language model using the target training sample data to obtain a target model for reviewing bidding documents.

[0029] In this embodiment, high-value "target training sample data" that has been intelligently filtered is used for final training, enabling the model to directly focus on learning the most essential and critical review patterns, avoiding interference from low-quality and redundant data. This not only significantly accelerates the model's convergence speed but also fundamentally ensures that the final target model has stronger discrimination accuracy, generalization ability, and reliability.

[0030] In this embodiment, the original sample dataset in the bidding and tendering field is acquired, and a training sample dataset is obtained from the original sample dataset according to a preset sampling ratio. A parameter-efficient low-rank adapter is used for preliminary model training and gradient acquisition, requiring only a very small number of parameter updates, enabling rapid completion of multiple iterations and gradient analysis under limited computing resources. The training sample data with the same review conclusion label and the corresponding comprehensive training gradient are clustered. For each group of sample gradients with review conclusion labels, target training sample data is selected from the sample gradient group based on the similarity between each comprehensive training gradient in the sample gradient group and the preset benchmark gradient corresponding to the review conclusion label. The selection is based on gradient similarity within the same review conclusion category, ensuring the consistency of the selected samples in teaching specific review skills and effectively filtering out noise and contradictory samples. The target training sample data is used to train the large language model to obtain the target model for reviewing bidding and tendering documents. The entire selection process closely follows the characteristics of data structuring, rule complexity, and diversified review dimensions in the bidding and tendering field, and the final output is a "core review knowledge base" rich in high-quality samples. The model trained based on this knowledge base has a decision-making process that better reflects professional review logic, and the results are more credible and traceable, providing reliable technical support for the fairness, efficiency and compliance of bidding activities.

[0031] Specifically, step S101, which involves obtaining the original sample dataset in the bidding and tendering field and obtaining the training sample dataset from the original sample dataset according to a preset sampling ratio, includes: Step A1: Obtain multiple preset sampling ratios and use each sampling ratio in ascending order; Step A2: Obtain a sampled dataset from the original sample dataset according to the currently used sampling ratio, and train the preset base model using the sampled dataset to obtain the result model; Step A3: Validate the resulting model using a preset validation dataset. If the validation passes, set the sampled dataset as the training dataset. If the validation fails, use the next sampling ratio and re-execute step A3.

[0032] In this embodiment, considering the massive scale, high training cost, and frequent updates of bidding document sample data, a threshold increment method is used to determine the minimum feasible training sample dataset for efficiently obtaining a high-quality gradient model. Specifically, a set of sampling ratios from small to large is preset, such as 5%, 8%, 10%, 12%, and 15%. Starting from the smallest sampling ratio, sample data is randomly collected from the original sample set at a rate of 5% as the sampling sample dataset. The preset base model is trained using this sample dataset to obtain the resulting model. A validation sample dataset is then obtained to validate the output of the resulting model. If the resulting model meets the model performance standard, the current sampling sample dataset is used as the training sample data; if the resulting model does not meet the model performance standard, sample data is collected from the original sample dataset using the next sampling ratio. The preset base model is then trained again using the sampling sample dataset at an 8% sampling ratio until the trained model meets the model performance standard. By iteratively validating from a small ratio, the minimum effective training set size can be automatically found while ensuring the model meets basic performance requirements. This method avoids the waste of resources or insufficient performance caused by blindly selecting the amount of data based on experience. It is especially suitable for bidding fields with massive amounts of data and high annotation costs, and achieves preliminary optimization and cost control in the data preparation stage.

[0033] Specifically, step S102, which involves training the large language model with the injected low-rank adapter using the training sample dataset for i rounds and obtaining the gradient vector generated by each training sample data with respect to the low-rank adapter in the reciprocal j rounds of training, includes: Step B1: Obtain a test sample dataset from the original sample dataset with the same amount of data as the training sample dataset; Step B2: Use the training sample dataset to train the large language model with the low-rank adapter for i rounds, and obtain j intermediate models obtained in the last j rounds of training, and use the forward and back propagation algorithms to calculate j training gradient vectors for each training sample data with respect to the j intermediate models; The forward propagation algorithm is the prediction result calculated normally by the large language model using the input data. The backpropagation algorithm is based on the prediction result of the forward propagation algorithm, calculates the error between each prediction result and the review conclusion label, and uses the loss function to calculate the gradient vector.

[0034] Step B3: Input the test sample dataset into the j intermediate models respectively, and use the forward and backpropagation algorithms to calculate the j test gradient vectors of each test sample dataset with respect to the j intermediate models.

[0035] In this embodiment, a test sample dataset with the same data volume as the training sample dataset is obtained and used as a control group for the training sample dataset. The training sample data is input into the large language model, and the gradient vector corresponding to each training sample data is calculated using the forward propagation algorithm, backpropagation algorithm, and loss function. Then, the parameters of the last few layers in the low-rank adapter are adjusted based on the gradient vectors and the stochastic gradient descent algorithm. A preset number of training rounds are performed to obtain a parameter-stable large language model containing the low-rank adapter. Three training gradient vectors corresponding to each training sample data are collected from the last three training rounds. Similarly, the test sample dataset is input into the three intermediate large language models trained in the last three rounds to obtain three test gradient vectors corresponding to each test sample data. The training gradient vectors generated during training and the test gradient vectors calculated on the test sample dataset are collected. These gradient data are related to the stochastic gradient descent algorithm, whose non-adaptive, batch-updating characteristics differ significantly from the Adam adaptive optimizer. This embodiment employs the stochastic gradient descent algorithm, which can more directly reflect the original impact of the sample itself on the low-rank adapter parameters, avoiding the normalization interference of the adaptive learning rate on the gradient magnitude. This provides higher quality and more discriminative feature basis for subsequent gradient-based data evaluation and screening.

[0036] Specifically, in step S103, the comprehensive training gradient of each training sample is calculated based on the j gradient vectors generated in the reciprocal j rounds of training. Then, training sample data with the same review conclusion label and their corresponding comprehensive training gradients are clustered to obtain sample gradient groups for each type of review conclusion label. This includes: Step C1: Calculate the training average of the j training gradient vectors generated by each training sample data in the last j rounds of training, and input the training average into the projection algorithm to obtain the comprehensive training gradient of each training sample data; Among them, fusion uses the mean method, and compression uses the fast Johnson-Lindenstrauss (JL) projection method. The JL method can embed a set of points in a high-dimensional space into a low-dimensional space while approximately preserving the Euclidean distance between the points.

[0037] Step C2: Calculate the test average of the j test gradient vectors generated in the reciprocal j rounds of training for each test sample data, and input the test average into the projection algorithm to obtain the comprehensive test gradient for each test sample data; Step C3: Cluster the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain the training sample gradient grouping of each type of review conclusion label; Step C4: Cluster the test sample data with the same review conclusion label and the corresponding comprehensive test gradient to obtain test sample gradient groups for each type of review conclusion label.

[0038] In this embodiment, for training sample data, the three training gradient vectors generated during the last three rounds of training for each training sample data are fused and compressed to obtain the comprehensive training gradient corresponding to each training sample data. The comprehensive training gradient, review conclusion label, and sample data ID of each training sample data are associated and stored in the database. Similarly, the test sample gradient of each test sample data is obtained, and the comprehensive test gradient, review conclusion label, and sample data ID of each test sample data are associated and stored in the database. The comprehensive training gradients of "risky" training sample data are clustered into one group; the comprehensive test gradients of "risky" test sample data are clustered into another group; the comprehensive test gradients of "risk-free" training sample data are clustered into another group; and the comprehensive test gradients of "risk-free" test sample data are clustered into yet another group. The fusion of multiple gradients through "average value" is explicitly achieved, improving feature stability; a "projection algorithm" is introduced for dimensionality reduction, significantly reducing subsequent computational complexity and storage overhead while retaining the key directional information of the gradient vectors. Meanwhile, the same fusion and dimensionality reduction were performed on the test data, and the data were clearly grouped, laying the foundation for accurate comparison in two balanced and comparable low-dimensional feature spaces.

[0039] Specifically, in step S104, the sample gradient grouping for each type of review conclusion label involves selecting target training sample data from the sample gradient group based on the similarity between each comprehensive training gradient within the sample gradient group and the preset benchmark gradient corresponding to the review conclusion label. This includes: Step D1: Calculate the average comprehensive test gradient based on each comprehensive test gradient in the test sample gradient grouping of each type of review conclusion label, and use it as the benchmark gradient of the corresponding type of review conclusion label. Step D2: Iterate through the training sample gradient groups of each type of review conclusion label in turn, calculate the cosine similarity between each comprehensive training gradient in the currently iterated training sample gradient group and the baseline gradient of the corresponding type of review conclusion label, and set the training sample data corresponding to the comprehensive training gradient in the currently iterated training sample gradient group whose cosine similarity is greater than the preset similarity threshold as the target training sample data.

[0040] In this implementation, such as Figure 2As shown, a stable large language model trained based on training sample data is trained for three rounds using both training set data (i.e., training sample dataset) and test set data (i.e., test sample dataset). Three training gradient vectors are collected for each training sample, and three test gradient vectors are collected for each test sample. These three training gradient vectors are then fused and compressed to obtain the comprehensive training gradient for each training sample. The training sample is then grouped according to its review conclusion label and comprehensive training gradient, resulting in two groups of training sample gradients. The same process is applied to the test sample data, resulting in two groups of test sample gradients. The average comprehensive test gradient within each group is calculated and used as the baseline gradient for that group. Within the same label class, the cosine similarity between each comprehensive training gradient and the baseline gradient is calculated. A cosine similarity closer to 1 indicates a greater consistency between the optimization direction of the training sample and the optimization direction of the test set for that class, generally signifying higher sample quality and representativeness. Within the same label class, the cosine similarity values ​​are sorted in descending order. The training sample data corresponding to the top 20% of the comprehensive training gradients constitute the high-quality refined data (i.e., the target training sample data) selected in this iteration. This selection process also preserves the complete data structure of the samples (tender project information, review rule details, review steps, review basis, negative / positive cases, etc.), ensuring that the refined training sample dataset not only has high gradient quality but also contains rich domain knowledge and case guidance, further improving the model's review accuracy and generalization ability.

[0041] Specifically, step S105, which involves training the large language model using the target training sample data to obtain a target model for reviewing bidding documents, includes: Step E1: Extract the background information, project information, and content information of a bidding document from the training sample data of each target, preset the review rules, and identify risky and non-risky labels; Step E2: Use the background information, project information, content information, and preset review rules of the bidding documents in each target training sample data as input to the large language model, and use the risky label and the risk-free label as the standard output of the large language model to train the large language model to obtain the target model.

[0042] In this embodiment, the selected target training sample dataset is used as a new training set to retrain the large language model injected into the LoRA adapter. This training can be further fine-tuned based on the large language model saved in S102, or the parameters in the LoRA adapter can be reinitialized and training can begin from scratch. To address the high computational cost and long cycle of directly training the large model 32B in the intelligent bidding review scenario, this embodiment proposes an accelerated screening strategy based on model scaling. The core of this strategy lies in utilizing the high transferability of the representational features in the target training sample data: First, the model 7B with a smaller parameter size and efficient training is used as the basic filter, and data fine filtering and performance verification are completed on it; the high-quality finely filtered dataset that has been verified can be directly transferred to train the target model 32B with a larger parameter size and stronger performance. By decoupling the data screening and large-scale training stages, this method significantly reduces the overall time and computational cost of data cleaning while ensuring the performance of the final model.

[0043] In this embodiment, the overall solution achieved the following technical effects: (1) It realized an efficient and adaptive high-quality training sample data screening mechanism: This invention abandons the data selection method that relies on fixed, empirical thresholds, and innovatively evaluates and determines the data screening threshold by dynamically evaluating and determining the LoRA light fine-tuning results on a small-scale validation set (such as 20% random samples). This method can automatically adapt the optimal screening criteria according to the characteristics of different tasks and sample datasets, ensuring the relevance and high quality of the selected training sample data, thereby fundamentally improving the model training efficiency and final performance. (2) It optimized the gradient-based data selection strategy and focused on key model knowledge: The key to this embodiment is that when evaluating the importance of data, the gradient vectors in the later stage of model training are mainly selected. This design is based on the insight that the gradient of deep network can better reflect the essential characteristics of the task and high-level semantics, avoiding the interference of unstable gradients and low-level general features in the early training, so that the data screening process can more accurately lock the most critical and information-rich samples for improving the target task. (3) Ensures semantic consistency and computational stability of data evaluation: When performing comprehensive gradient aggregation and similarity calculation, this embodiment introduces a strategy of segmentation by category / label, stipulating that gradient addition and sample similarity measurement are only performed within the same semantic category. This effectively prevents gradient cancellation or misjudgment of similarity caused by semantic differences between samples of different categories, greatly improving the accuracy and robustness of data importance evaluation, and providing a more consistent target training sample dataset for model learning. (4) Improves model generalization ability and stability of feature selection: This embodiment makes a key improvement in the gradient-based data selection framework: using the gradient generated by the stochastic gradient descent optimizer instead of the commonly used Adam optimizer gradient as the feature representation for calculating the importance and similarity of sample data. This choice stems from a deep understanding of optimization dynamics: Adam, due to its adaptive learning rate characteristics, tends to optimize the model to the "sharp" minimum of the loss function, and its gradient direction changes drastically, which may introduce noise and reduce the stability of the data features.

[0044] Example 2 This invention provides an apparatus for generating a model for reviewing bidding documents, such as... Figure 3 As shown, the device specifically includes the following components: The acquisition module 301 is used to acquire the original sample dataset in the bidding and tendering field, and to acquire the training sample dataset from the original sample dataset according to a preset sampling ratio; wherein, the sample data includes: bidding and tendering documents, preset review rules, and review conclusion labels, and the review conclusion labels represent the compliance status of the bidding and tendering documents under the preset review rules; Training module 302 is used to train the large language model injected with the low-rank adapter using the training sample dataset for i rounds, and to obtain the gradient vector generated by each training sample data in the reciprocal j rounds of training with respect to the low-rank adapter; wherein, in each round of training, only the parameters in the low-rank adapter are adjusted, and the parameters in the large language model are not adjusted. The calculation module 303 is used to calculate the comprehensive training gradient of each training sample data based on the j gradient vectors generated in the jth round of training for each training sample data, and to cluster the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain sample gradient groupings of various review conclusion labels. Grouping module 304 is used to group samples for each type of review conclusion label. Based on the similarity between each comprehensive training gradient in a sample gradient group and the preset benchmark gradient corresponding to the review conclusion label, target training sample data is selected from the sample gradient group. The generation module 305 is used to train the large language model using the target training sample data to obtain a target model for reviewing bidding documents.

[0045] Optionally, the acquisition module 301 is specifically used for: Steps for using the ratio: Obtain multiple preset sampling ratios, and use each sampling ratio in ascending order; Model training steps: Obtain a sample dataset from the original sample dataset according to the currently used sampling ratio, and train the preset base model using the sample dataset to obtain the result model; Model validation steps: Validate the resulting model using a preset validation dataset. If the validation passes, set the sampled dataset as the training dataset. If the validation fails, use the next sampling ratio and re-execute the model training steps.

[0046] Optionally, the training module 302 is specifically used for: Obtain a test sample dataset with the same amount of data as the training sample dataset from the original sample dataset; The large language model with low-rank adapter injected is trained i times using the training sample dataset, and j intermediate models are obtained in the reciprocal j rounds of training. The j training gradient vectors of each training sample data with respect to the j intermediate models are calculated using the forward and backward propagation algorithms. The test sample dataset is input into the j intermediate models respectively, and the forward and backpropagation algorithms are used to calculate the j test gradient vectors of each test sample dataset with respect to the j intermediate models.

[0047] Optionally, the calculation module 303 is specifically used for: Calculate the average training gradient vectors of j training gradient vectors generated in the last j rounds of training for each training sample data, and input the average training gradient into the projection algorithm to obtain the comprehensive training gradient of each training sample data; Calculate the average value of the j test gradient vectors generated by each test sample data in the reciprocal j rounds of training, and input the average value into the projection algorithm to obtain the comprehensive test gradient of each test sample data; Cluster the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain the training sample gradient grouping of each type of review conclusion label. Cluster the test sample data with the same review conclusion label and the corresponding comprehensive test gradient to obtain test sample gradient groups for each type of review conclusion label.

[0048] Optionally, the grouping module 304 is specifically used for: The average comprehensive test gradient is calculated based on each comprehensive test gradient in the test sample gradient group of each type of review conclusion label, and is used as the benchmark gradient of the corresponding type of review conclusion label. Iterate through the training sample gradient groups of each type of review conclusion label, calculate the cosine similarity between each comprehensive training gradient in the currently traversed training sample gradient group and the baseline gradient of the corresponding type of review conclusion label, and set the training sample data corresponding to the comprehensive training gradient in the currently traversed training sample gradient group whose cosine similarity is greater than the preset similarity threshold as the target training sample data.

[0049] Optionally, the generation module 305 is specifically used for: The background information, project information, and content information of a bidding document are extracted from the training sample data of each target, and the review rules are preset, with risk labels and no risk labels. The background information, project information, content information, and preset review rules of the bidding documents in each target training sample data are used as input to the large language model, and the risky label and the risk-free label are used as the standard output of the large language model to train the large language model to obtain the target model.

[0050] Example 3 This embodiment also provides a computer device, such as a smartphone, tablet computer, laptop computer, desktop computer, rack server, blade server, tower server, or cabinet server (including a standalone server or a server cluster composed of multiple servers), etc., capable of executing programs. Figure 4As shown, the computer device 40 in this embodiment includes, but is not limited to, a memory 401 and a processor 402 that are communicatively connected to each other via a system bus. It should be noted that... Figure 4 Only a computer device 40 with components 401-402 is shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0051] In this embodiment, the memory 401 (i.e., the readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 401 may be an internal storage unit of the computer device 40, such as the hard disk or memory of the computer device 40. In other embodiments, the memory 401 may also be an external storage device of the computer device 40, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 40. Of course, the memory 401 may include both the internal storage unit and its external storage device of the computer device 40. In this embodiment, the memory 401 is typically used to store the operating system and various application software installed on the computer device 40. In addition, the memory 401 may also be used to temporarily store various types of data that have been output or will be output.

[0052] In some embodiments, processor 402 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. This processor 402 is typically used to control the overall operation of computer device 40.

[0053] Specifically, in this embodiment, the processor 402 is used to execute a program stored in the memory 401 for generating a model for reviewing bidding documents. When the program for generating a model for reviewing bidding documents is executed, it performs the following steps: Obtain the original sample dataset in the bidding and tendering field, and obtain the training sample dataset from the original sample dataset according to a preset sampling ratio; wherein, the sample data includes: bidding and tendering documents, preset review rules, and review conclusion labels, and the review conclusion labels represent the compliance status of the bidding and tendering documents under the preset review rules; The large language model with the low-rank adapter injected is trained i times using the training sample dataset, and the gradient vector generated by each training sample data in the reciprocal j-th training round is obtained for the low-rank adapter; wherein, in each training round, only the parameters in the low-rank adapter are adjusted, and the parameters in the large language model are not adjusted. The comprehensive training gradient of each training sample is calculated based on the j gradient vectors generated in the jth round of training for each training sample. Then, the training sample data with the same review conclusion label and their corresponding comprehensive training gradients are clustered to obtain sample gradient groups for each type of review conclusion label. For each type of review conclusion label, the sample gradient grouping is used to select target training sample data from the sample gradient group based on the similarity between each comprehensive training gradient in the sample gradient group and the preset benchmark gradient corresponding to the review conclusion label. The large language model is trained using the target training sample data to obtain a target model for reviewing bidding documents.

[0054] For a detailed description of the above method steps, please refer to Example 1. This example will not be repeated here.

[0055] Example 4 This embodiment also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, app store, etc., which stores a computer program. When the computer program is executed by a processor, it implements the following method steps: Obtain the original sample dataset in the bidding and tendering field, and obtain the training sample dataset from the original sample dataset according to a preset sampling ratio; wherein, the sample data includes: bidding and tendering documents, preset review rules, and review conclusion labels, and the review conclusion labels represent the compliance status of the bidding and tendering documents under the preset review rules; The large language model with the low-rank adapter injected is trained i times using the training sample dataset, and the gradient vector generated by each training sample data in the reciprocal j-th training round is obtained for the low-rank adapter; wherein, in each training round, only the parameters in the low-rank adapter are adjusted, and the parameters in the large language model are not adjusted. The comprehensive training gradient of each training sample is calculated based on the j gradient vectors generated in the jth round of training for each training sample. Then, the training sample data with the same review conclusion label and their corresponding comprehensive training gradients are clustered to obtain sample gradient groups for each type of review conclusion label. For each type of review conclusion label, the sample gradient grouping is used to select target training sample data from the sample gradient group based on the similarity between each comprehensive training gradient in the sample gradient group and the preset benchmark gradient corresponding to the review conclusion label. The large language model is trained using the target training sample data to obtain a target model for reviewing bidding documents.

[0056] For a detailed description of the above method steps, please refer to the first embodiment. This embodiment will not repeat the details here.

[0057] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0058] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0059] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0060] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method of generating a model for reviewing a tender document, characterized by, The method comprises: obtaining an original sample data set in the field of bidding and tendering, and obtaining a training sample data set from the original sample data set according to a preset sampling ratio; wherein the sample data comprises a bidding and tendering document, a preset review rule and a review conclusion label, and the review conclusion label represents the compliance state of the bidding and tendering document under the preset review rule; using the training sample data set to perform i rounds of training on a large language model injected with a low-rank adapter, and obtaining a gradient vector generated by each training sample data in the last j rounds of training on the low-rank adapter; wherein only the parameters in the low-rank adapter are adjusted in each round of training, and the parameters in the large language model are not adjusted; calculating a comprehensive training gradient of each training sample data according to the j gradient vectors generated by each training sample data in the last j rounds of training, and clustering the training sample data and the corresponding comprehensive training gradient with the same review conclusion label to obtain a sample gradient grouping of each review conclusion label; for each sample gradient grouping of the review conclusion label, filtering out target training sample data from the sample gradient grouping according to the similarity between each comprehensive training gradient in the sample gradient grouping and a preset reference gradient corresponding to the review conclusion label; using the target training sample data to train the large language model to obtain a target model for reviewing bidding and tendering documents.

2. The method of claim 1, wherein, The method comprises: proportion using step: obtaining a plurality of preset sampling ratios, and using each sampling ratio in turn according to the order from small to large; model training step: obtaining a sampling sample data set from the original sample data set according to the currently used sampling ratio, and training a preset base model through the sampling sample data set to obtain a result model; model verification step: verifying the result model by using a preset verification data set, if the verification is passed, setting the sampling sample data set as the training sample data set, if the verification is not passed, using the next sampling ratio and re-executing the model training step.

3. The method of claim 1, wherein, The method comprises: obtaining a test sample data set with the same data volume as the training sample data set from the original sample data set; using the training sample data set to perform i rounds of training on a large language model injected with a low-rank adapter, and obtaining j intermediate models obtained in the last j rounds of training, and calculating j training gradient vectors of each training sample data with respect to the j intermediate models by using a forward algorithm and a back propagation algorithm; inputting the test sample data set into the j intermediate models respectively, and calculating j test gradient vectors of each test sample data set with respect to the j intermediate models by using a forward algorithm and a back propagation algorithm.

4. The method of claim 3, wherein, The j gradient vectors generated in the inverse j round of training are used to calculate a comprehensive training gradient of each training sample data, and training sample data and corresponding comprehensive training gradients with the same review conclusion label are clustered to obtain sample gradient groups of each review conclusion label, including: The training average value of the j training gradient vectors generated in the inverse j round of training is calculated, and the training average value is input into a projection algorithm to obtain the comprehensive training gradient of each training sample data; The test average value of the j test gradient vectors generated in the inverse j round of training is calculated, and the test average value is input into a projection algorithm to obtain the comprehensive test gradient of each test sample data; Training sample data and corresponding comprehensive training gradients with the same review conclusion label are clustered to obtain training sample gradient groups of each review conclusion label; Test sample data and corresponding comprehensive test gradients with the same review conclusion label are clustered to obtain test sample gradient groups of each review conclusion label.

5. The method of generating a model for reviewing a bid document of claim 4, wherein, For each sample gradient group of each review conclusion label, a target training sample data is selected from the sample gradient group according to the similarity between each comprehensive training gradient in the sample gradient group and a preset reference gradient corresponding to the review conclusion label, including: An average comprehensive test gradient is calculated according to each comprehensive test gradient in the test sample gradient group of each review conclusion label to serve as a reference gradient of the corresponding review conclusion label; The cosine similarity between each comprehensive training gradient in the training sample gradient group and the reference gradient of the corresponding review conclusion label is calculated by sequentially traversing the training sample gradient group of each review conclusion label, and the training sample data corresponding to the comprehensive training gradient with a cosine similarity greater than a preset similarity threshold in the currently traversed training sample gradient group is set as the target training sample data.

6. The method of generating a model for reviewing a bid document of claim 1, wherein, The target model for reviewing bidding documents is obtained by training the large language model using the target training sample data, including: The background information, project information, content information of a bidding document, a preset review rule, a risk label, and a no-risk label are parsed from each target training sample data; The background information, project information, content information of a bidding document, and a preset review rule in each target training sample data are used as the input of the large language model, and the risk label and the no-risk label are used as the standard output of the large language model, and the large language model is trained to obtain the target model.

7. An apparatus for generating a model for reviewing a bid document, the apparatus comprising: The device includes: An acquisition module is configured to acquire an original sample data set in the bidding field and acquire a training sample data set from the original sample data set according to a preset sampling ratio; wherein the sample data includes a bidding document, a preset review rule, and a review conclusion label, and the review conclusion label represents the compliance status of the bidding document under the preset review rule; The training module is configured to perform i rounds of training on the large language model injected with the low-rank adapter by using the training sample dataset, and obtain gradient vectors generated by the low-rank adapter for each training sample data in the last j rounds of training; in each round of training, only the parameters in the low-rank adapter are adjusted, and the parameters in the large language model are not adjusted. The calculation module is configured to calculate a comprehensive training gradient of each training sample data according to j gradient vectors generated by each training sample data in the last j rounds of training, and cluster the training sample data with the same review conclusion label and the corresponding comprehensive training gradient to obtain a sample gradient grouping of each review conclusion label. The grouping module is configured to, for the sample gradient grouping of each review conclusion label, filter out target training sample data from the sample gradient grouping according to a similarity between each comprehensive training gradient in the sample gradient grouping and a preset reference gradient corresponding to the review conclusion label. The generation module is configured to train the large language model by using the target training sample data to obtain a target model for reviewing bidding documents.

8. The apparatus for generating a model for reviewing a bid document of claim 7, wherein, The training module is specifically configured to: obtain a test sample dataset with the same data volume as the training sample dataset from the original sample dataset; perform i rounds of training on the large language model injected with the low-rank adapter by using the training sample dataset, and obtain j intermediate models obtained in the last j rounds of training, and j training gradient vectors of each training sample data with respect to the j intermediate models by using a forward algorithm and a back propagation algorithm; input the test sample dataset into the j intermediate models respectively, and calculate j test gradient vectors of each test sample dataset with respect to the j intermediate models by using the forward algorithm and the back propagation algorithm.

9. A computer device comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the method in any one of claims 1 to 6 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 6.

Images