Metal mine dispatching large model fine-tuning data screening method and device and storage medium

Abstract

The application relates to the technical field of artificial intelligence and intelligent mines, and provides a metal mine scheduling large model fine-tuning data screening method, a device and a storage medium, which comprises the following steps: based on the normalized quality score, the mutual information correlation degree score and the high-dimensional semantic consistency rearrangement evaluation score of a to-be-screened fine-tuning data set, and obtaining the comprehensive evaluation score of the to-be-screened fine-tuning data set; taking the comprehensive evaluation score as a quality evaluation dimension to obtain a candidate fine-tuning data set containing a multi-dimensional score label; obtaining a graph-level high-dimensional semantic topology feature vector of the candidate fine-tuning data set; performing dimension reduction processing on the graph-level high-dimensional semantic topology feature vector to obtain a dimension-reduced semantic feature vector; based on the semantic clustering result of the semantic feature vector, determining a plurality of scheduling fine-tuning data sets from the candidate fine-tuning data set, and generating a target quality scheduling fine-tuning data set with balanced structure distribution and covering multiple scheduling scenarios based on the plurality of scheduling fine-tuning data sets. In this way, the data quality of the fine-tuning data set is improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and smart mining technology, specifically to a method, apparatus and storage medium for fine-tuning data screening of a large-scale scheduling model for metal mines. Background Technology

[0002] With the development of artificial intelligence, Large Language Models (LLMs) have shown great potential in intelligent mine scheduling. However, the geological conditions of metal mining are complex, and its production scheduling involves close coordination among multiple devices and processes, with strong temporal and spatial constraints and rigid safety requirements. Directly applying general-purpose large models can easily lead to "illusions" and erroneous scheduling, causing serious safety hazards. Therefore, it is crucial to fine-tune large models using domain-specific data.

[0003] However, in the field of metal mine scheduling, there is currently a severe lack of high-quality, structured, dedicated fine-tuning datasets. Existing open-source mine corpora mostly focus on general safety regulations for coal mines or shallow production common sense, while deep scheduling knowledge related to metal mines is almost non-existent. Even if mining companies have accumulated some historical scheduling logs and business data, these data generally have the following problems: (1) severe unstructuredness, mostly manifested as natural language text, handwritten records, or scattered logs, lacking a unified instruction format and annotation; (2) high noise, mostly manifested as anomalies such as instruction execution deviations, record omissions, and timestamp errors; (3) significant homogenization, mostly manifested as highly repetitive samples of routine working conditions. These data quality problems make it impossible for the original business data to be directly used for fine-tuning of high-quality metal mine scheduling large models.

[0004] Furthermore, existing fine-tuning data selection and training methods suffer from issues such as simple sample bias and simplistic data topology when dealing with complex scheduling operations in metal mines. They neglect the complex network topology dependencies between various devices, personnel, and environmental entities in scheduling instructions, resulting in excessive data stacking in simple, routine scenarios, making it difficult to eliminate invalid and low-quality data, while lacking coverage of key complex scenarios. These problems lead to insufficient attention and learning weights for complex examples such as multi-device collaboration, making it difficult for fine-tuned models to handle complex decision-making scenarios.

[0005] In summary, how to overcome the scarcity and low quality of existing metal mine scheduling fine-tuning data, and the limitations of traditional data screening and fine-tuning methods, and provide a high-quality and efficient method for screening fine-tuning data for large-scale metal mine scheduling models, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This application aims to address the problem of poor data quality in the fine-tuning dataset of large-scale metal mine scheduling models in the prior art, and provides a method, apparatus, and storage medium for filtering fine-tuning data in large-scale metal mine scheduling models.

[0007] To solve the above problems, this application is implemented as follows:

[0008] Firstly, this application provides a method for filtering fine-tuning data in a large-scale metal mine scheduling model, including: Based on historical scheduling logs and business data of metal mines, scheduling data sets under different scheduling scenarios are extracted, and a scheduling alignment sample set adapted to fine-tuning of a general-purpose base large language model is constructed based on the scheduling data sets. The initial scheduling question-and-answer evaluation model is trained based on the aforementioned scheduling alignment sample set to obtain a scheduling question-and-answer quality evaluation model for quality verification of scheduling data. Based on the scheduling question-answering quality assessment model, the normalized quality score of the dataset to be screened and fine-tuned is obtained; based on autoregressive language modeling, the mutual information correlation score of the dataset to be screened and fine-tuned is obtained; and based on the pre-trained semantic representation model, the high-dimensional semantic consistency reordering evaluation score of the dataset to be screened and fine-tuned is obtained. The normalized quality score, the mutual information correlation score, and the high-dimensional semantic consistency reordering evaluation score are multiplicatively fused to generate a comprehensive evaluation score for the dataset to be screened for fine-tuning; the comprehensive evaluation score is used as a quality evaluation dimension and is successively incorporated into the corresponding dataset to be screened for fine-tuning, and a candidate fine-tuning dataset containing multi-dimensional scoring labels is formed by splicing and mapping. A metal mine scheduling topology dependency graph is constructed based on the text entity lexical units of the candidate fine-tuning dataset. The graph attention network is then used to aggregate node features of the metal mine scheduling topology dependency graph to obtain a graph-level high-dimensional semantic topology feature vector. The high-dimensional semantic topological feature vector at the graph level is subjected to dimensionality reduction processing to obtain the dimensionality-reduced semantic feature vector; the semantic feature vector is semantically clustered using an adaptive clustering algorithm to obtain the semantic clustering result; based on the semantic clustering result, multiple scheduling fine-tuning datasets are determined from the candidate fine-tuning dataset; and a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios is generated based on the multiple scheduling fine-tuning datasets.

[0009] Secondly, this application provides a data filtering device for fine-tuning a large-scale metal mine scheduling model, comprising: The module is used to extract scheduling data sets under different scheduling scenarios based on historical scheduling logs and business data of metal mines, and to build a scheduling alignment sample set adapted to fine-tuning of the general base large language model based on the scheduling data set. The training module is used to train the initial scheduling question-answering evaluation model based on the scheduling alignment sample set, so as to obtain a scheduling question-answering quality evaluation model for quality verification of scheduling data. The first acquisition module is used to acquire the normalized quality score of the dataset to be screened and fine-tuned based on the scheduling question-answering quality assessment model; acquire the mutual information correlation score of the dataset to be screened and fine-tuned based on autoregressive language modeling; and acquire the high-dimensional semantic consistency reordering assessment score of the dataset to be screened and fine-tuned based on a pre-trained semantic representation model. The second acquisition module is used to multiply and fuse the normalized quality score, the mutual information correlation score and the high-dimensional semantic consistency reordering evaluation score to generate a comprehensive evaluation score for the dataset to be screened and fine-tuned; the comprehensive evaluation score is used as a quality evaluation dimension and is successively incorporated into the corresponding dataset to be screened and fine-tuned, and a candidate fine-tuning dataset containing multi-dimensional scoring labels is formed by splicing and mapping. The aggregation module is used to construct a metal mine scheduling topology dependency graph based on the text entity lexical units of the candidate fine-tuning dataset, and to aggregate the node features of the metal mine scheduling topology dependency graph through a graph attention network to obtain a graph-level high-dimensional semantic topology feature vector. The generation module is used to perform dimensionality reduction processing on the graph-level high-dimensional semantic topology feature vector to obtain the dimensionality-reduced semantic feature vector; to perform semantic clustering on the semantic feature vector using an adaptive clustering algorithm to obtain the semantic clustering result; and to determine multiple scheduling fine-tuning datasets from the candidate fine-tuning dataset based on the semantic clustering result; and to generate a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios based on the multiple scheduling fine-tuning datasets. The target quality scheduling fine-tuning dataset is used to fine-tune the pre-trained general-purpose base language model to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling.

[0010] Thirdly, this application provides a terminal device including a processor and a memory, wherein the memory stores a program or instructions executable on the processor, and the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0011] Fourthly, this application provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0012] Fifthly, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the steps of the method described in the first aspect.

[0013] In a sixth aspect, this application provides a computer program product stored in a storage medium, which is executed by at least one processor to perform the steps of the method described in the first aspect.

[0014] Compared with existing technologies, this application has the following advantages: by constructing a scheduling alignment sample set adapted to fine-tuning of a general-purpose base large language model through a scheduling dataset, training a scheduling question-answering quality assessment model with a dynamic weight loss mechanism, and constructing a scoring system that includes normalized quality score, mutual information correlation score and high-dimensional semantic consistency reordering evaluation score, it can effectively avoid simple sample bias in the scheduling fine-tuning dataset and improve the data quality of the candidate fine-tuning dataset.

[0015] Furthermore, by constructing a topological dependency graph of the metal mine scheduling of the candidate fine-tuning dataset, and performing graph attention feature aggregation and dimensionality reduction clustering, the candidate fine-tuning dataset is locally optimized and sampled according to the comprehensive evaluation score. Finally, a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios is obtained, which provides a data foundation for constructing an efficient, high-quality, and highly safe large-scale metal mine scheduling model for fine-tuning. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a method for fine-tuning data filtering in a large-scale metal mine scheduling model according to an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a metal mine scheduling large model fine-tuning data filtering device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] The terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses. Additionally, the use of "and / or" in this application indicates at least one of the connected objects, such as A and / or B and / or C, representing seven possibilities: including A alone, B alone, C alone, and the presence of both A and B, both B and C, both A and C, and the presence of A, B, and C.

[0020] In this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0021] The following describes the data screening method for fine-tuning the large-scale metal mine scheduling model provided in this application.

[0022] See Figure 1 , Figure 1 This is a flowchart illustrating a method for fine-tuning data filtering in a large-scale metal mine scheduling model according to an embodiment of this application. Figure 1 The data filtering method for fine-tuning the large-scale metal mine scheduling model shown can be executed by terminal devices such as mobile phones and computers.

[0023] like Figure 1 As shown, the data filtering method for fine-tuning the large-scale metal mine scheduling model provided in this application may include the following steps: Step 101: Based on the historical scheduling logs and business data of the metal mine, extract scheduling data sets under different scheduling scenarios, and construct a scheduling alignment sample set adapted to the fine-tuning of the general base large language model based on the scheduling data sets.

[0024] In this embodiment, a set of scheduling data covering multi-equipment collaborative scheduling scenarios, crushing process simulation scheduling scenarios, personnel dispatch scheduling scenarios, spatiotemporal constraint scheduling scenarios, and safety rigid constraint scheduling scenarios can be obtained from the historical scheduling logs and business data of the metal mine.

[0025] Each scheduling data in the scheduling data set includes a scheduling task and the scheduling decision corresponding to the scheduling task.

[0026] In some embodiments, a scheduling task can be understood as metal mine problem information or scheduling instructions corresponding to the problem information; a scheduling decision can be understood as response information corresponding to the metal mine problem information or a corresponding scheduling reply. That is, the scheduling data in the scheduling data set of this application can be scheduling problem data or response information corresponding to scheduling problem data, without specific limitations here.

[0027] The aforementioned general-purpose base language model can be understood as a general-purpose base language model that has not been fine-tuned for the metal mine scheduling scenario. By constructing a scheduling alignment sample set that is adapted to the fine-tuning of the general-purpose base language model from the scheduling data set, the information output by the scheduling question-answering quality assessment model trained based on the scheduling alignment sample set can be adapted to the training process of the pre-trained general-purpose base language model.

[0028] In some embodiments, for any scheduling data in the scheduling data set Obtain standard professional scheduling strategies as positive samples. ; The scheduling data The input is fed into a general-purpose, base-based large language model that has not been fine-tuned for metal mine scheduling scenarios, generating generalized or non-standard scheduling responses as negative samples. ; The scheduling data positive samples With negative samples By splicing the data, a scheduling alignment sample set for metal mines is constructed. ,in The total number of samples constructed.

[0029] For example, in a crushing process simulation scheduling scenario, the scheduling data for "linkage scheduling command when jaw crusher is overloaded" is... Extract professional scheduling strategies from domain experts or standard operating procedures as positive samples. .

[0030] At the same time, the scheduling data The input is fed into a general-purpose, base-based large language model that has not been fine-tuned with data specific to metal mines. The generated scheduling responses, whether generalized, inaccurate, or containing "illusions," are collected as negative samples. By combining these three elements, a metal mine scheduling alignment sample set is constructed. ,in The total number of samples constructed.

[0031] Step 102: Train the initial scheduling question-and-answer evaluation model based on the scheduling alignment sample set to obtain a scheduling question-and-answer quality evaluation model for quality verification of scheduling data.

[0032] In this embodiment, the scheduling alignment sample set can be used as input to train the initial scheduling question-answering evaluation model until the model parameters of the initial scheduling question-answering evaluation model tend to stabilize, so as to obtain a scheduling question-answering quality evaluation model for quality verification of scheduling data.

[0033] In some embodiments, training an initial scheduling question-answering evaluation model based on the scheduling alignment sample set to obtain a scheduling question-answering quality evaluation model for quality verification of scheduling data includes: Input any sample from the scheduling alignment sample set into the initial scheduling question-answering evaluation model, and output the standard quality score for positive samples and the standard quality score for negative samples; Using a paired comparison probability framework, the preference probability that the positive sample is better than the negative sample is calculated; Based on the preference probability, a dynamic adaptive weight loss function based on difficult example mining is constructed to automatically reduce the gradient weight of high-confidence samples of simple tasks in metal mine scheduling during backpropagation. This forces the initial scheduling question-answering evaluation model to focus on complex scheduling difficulties in metal mine scheduling. The application scenarios of these complex scheduling difficulties include multi-device collaborative scheduling scenarios, spatiotemporal constraint scheduling scenarios, and safety rigid constraint scheduling scenarios. The model parameters of the initial scheduling question-and-answer evaluation model are adjusted based on the dynamic adaptive weight loss function to obtain a scheduling question-and-answer quality evaluation model for quality verification of scheduling data.

[0034] The aforementioned hard example mining can be understood as actively identifying samples that the model currently predicts incorrectly or with low prediction confidence during the training process, and allowing the model to focus on learning from these samples.

[0035] The aforementioned complex scheduling examples can be understood as high-difficulty samples involving multi-device collaboration, multiple spatiotemporal constraints, and approaching safety boundaries, making it difficult for the model to make correct scheduling decisions.

[0036] In this embodiment, the scheduling question-and-answer quality assessment model trained in the above manner possesses the core capability to automatically score the quality of metal mine scheduling data pairs. Moreover, by introducing a dynamic adaptive weight loss mechanism based on hard example mining, the scheduling question-and-answer quality assessment model significantly improves its discriminative power for complex scheduling scenarios such as multi-device collaboration and safety boundary approximation, and provides a reliable automated evaluation tool for the subsequent screening of high-quality fine-tuning data.

[0037] In some embodiments, the standard quality score of the above-mentioned positive samples can be defined as follows: The standard quality score for the aforementioned negative samples can be defined as follows: ,in These are the model network parameters for the initial scheduling question-answering evaluation model; the above preference probabilities can be defined as follows: ,in, The expression is:

[0038] The formula for calculating the above dynamic adaptive weight loss function is as follows:

[0039] in, This represents the total number of samples in a single training batch. and These are preset dynamic adjustment hyperparameters used to control the dynamic weight ratio of the complex samples in the total loss.

[0040] Step 103: Based on the scheduling question-answering quality assessment model, obtain the normalized quality score of the dataset to be screened and fine-tuned; based on autoregressive language modeling, obtain the mutual information correlation score of the dataset to be screened and fine-tuned; and based on the pre-trained semantic representation model, obtain the high-dimensional semantic consistency reordering evaluation score of the dataset to be screened and fine-tuned.

[0041] The aforementioned dataset to be screened and fine-tuned can be understood as the original dataset of training samples used to fine-tune the pre-trained general-purpose pedestal large language model.

[0042] In this embodiment, the data quality and complexity of the dataset to be screened and fine-tuned can be evaluated from three dimensions. By adopting a multi-dimensional joint scoring mechanism, the screening effect of the dataset to be screened and fine-tuned can be improved.

[0043] In some embodiments, obtaining a normalized quality score for the dataset to be screened and fine-tuned based on the scheduling question-answering quality assessment model; obtaining a mutual information correlation score for the dataset to be screened and fine-tuned based on autoregressive language modeling; and obtaining a high-dimensional semantic consistency re-ranking evaluation score for the dataset to be screened and fine-tuned based on a pre-trained semantic representation model includes: Obtain the fine-tuning dataset to be filtered, wherein each fine-tuning data in the fine-tuning dataset includes metal mine problem information and corresponding answer information; Obtain the initial quality score of the dataset to be screened and fine-tuned by the scheduling question-answering quality assessment model, and map the initial quality score to the desired dataset using the max-min normalization method. The interval is used to obtain the normalized quality score of the dataset to be screened and fine-tuned; Based on autoregressive language modeling, the conditional generation information entropy is calculated to generate the corresponding response information under the condition of the target metal mine problem information; the unconditional generation information entropy is calculated under the condition of no scheduling information, and the ratio of the conditional generation information entropy to the unconditional generation information entropy is calculated to obtain the mutual information correlation score of the dataset to be screened and fine-tuned; wherein, the target metal mine problem information condition is the metal mine problem information of any fine-tuned data in the dataset to be screened and fine-tuned. A pre-trained semantic representation model is used to extract dense feature vectors of the metal mine problem information and the response information, respectively. The cosine similarity between the dense feature vectors of the metal mine problem information and the response information is calculated, and the cosine similarity is linearly shifted and normalized to obtain the high-dimensional semantic consistency reordering evaluation score of the dataset to be screened and fine-tuned.

[0044] By normalizing the initial quality score, the dimensional differences in scores between different batches can be eliminated.

[0045] The aforementioned mutual information correlation score is used to measure the degree of dependence of the answer information on the target metal mine question information, effectively filtering out low information entropy content, such as filtering out low-information text such as "Okay, I understand".

[0046] The pre-trained semantic representation model mentioned above can be either the BAAI / bge-large-zh-v1.5 embedding model, which performs well in the industrial and engineering fields, or a Chinese semantic model finely tuned based on the RoBERTa architecture.

[0047] By applying linear translation normalization to the cosine similarity, it can also be translated and scaled to... The interval facilitates subsequent product fusion.

[0048] In this embodiment, the data quality and complexity of the dataset to be screened and fine-tuned are evaluated from three dimensions. By adopting a multi-dimensional joint scoring mechanism, the screening effect of the dataset to be screened and fine-tuned is improved.

[0049] In some embodiments, the aforementioned dataset to be screened and fine-tuned can be defined as The aforementioned information on metal mine issues can be defined as follows: The above answer information can be defined as , Indicates the number of data sets to be fine-tuned; The above normalized quality fraction can be defined as And its expression is:

[0050] in, and These are the lowest and highest initial quality scores in the current dataset to be filtered, respectively. The information entropy generated under the above conditions can be defined as follows: And its expression is:

[0051] in, For the answer information The sequence length, To answer the question in the information Each word element; The information entropy of the unconditionally generated information mentioned above can be defined as follows: ; The above mutual information correlation score can be defined as follows: And its expression is:

[0052] The above-mentioned high-dimensional semantic consistency reordering evaluation score can be defined as follows: And its expression is:

[0053] In the formula, This represents the dense feature vector of the aforementioned metal mine problem information. This is the dense feature vector of the above answer information.

[0054] Step 104: Multiply and fuse the normalized quality score, the mutual information correlation score, and the high-dimensional semantic consistency reordering evaluation score to generate a comprehensive evaluation score for the dataset to be screened and fine-tuned; use the comprehensive evaluation score as a quality evaluation dimension and incorporate it into the corresponding dataset to be screened and fine-tuned one by one, and form a candidate fine-tuning dataset containing multi-dimensional scoring labels by splicing and mapping.

[0055] In this embodiment, the comprehensive evaluation score generated by product fusion is used as a quality assessment dimension. It can be spliced ​​and mapped to form a candidate fine-tuning dataset containing multi-dimensional scoring labels. This allows the multi-dimensional scoring labels carried by the candidate fine-tuning dataset to evaluate the data quality and complexity of the dataset to be screened from multiple dimensions.

[0056] In some embodiments, the aforementioned comprehensive evaluation score can be defined as And its expression is:

[0057] The aforementioned candidate fine-tuning dataset containing multidimensional rating labels can be defined as follows: And its expression is:

[0058] In the formula, Information on issues related to metal mines. To answer the information.

[0059] In some embodiments, metal mine problem information can be understood as scheduling instructions, and response information can be understood as scheduling replies.

[0060] Step 105: Construct a metal mine scheduling topology dependency graph based on the text entity lexical units of the candidate fine-tuning dataset, and perform node feature aggregation on the metal mine scheduling topology dependency graph through a graph attention network to obtain a graph-level high-dimensional semantic topology feature vector.

[0061] In this embodiment, by extracting the graph-level high-dimensional semantic topological feature vector of the candidate fine-tuning dataset, it is possible to adapt to the collaborative logical constraints of equipment and spatiotemporal scheduling in metal mines.

[0062] In some embodiments, the step of constructing a metal mine scheduling topology dependency graph based on text entity lexical units of the candidate fine-tuning dataset, and then aggregating node features of the metal mine scheduling topology dependency graph using a graph attention network to obtain a graph-level high-dimensional semantic topology feature vector includes: Extract text entity words from the candidate fine-tuning dataset as nodes; Based on the production scheduling rules and spatiotemporal constraints of metal mines, directed edges are established between the nodes to construct a topological dependency graph for metal mine scheduling. A graph attention network is used to aggregate node features in the scheduling topology dependency graph of the metal mine to obtain a graph-level high-dimensional semantic topology feature vector that represents the overall logical structure of the scheduling data.

[0063] In this embodiment, by extracting the graph-level high-dimensional semantic topological feature vector of the candidate fine-tuning dataset, the correlation between devices, personnel and environmental entities within the scheduling sample can be effectively mined, which can adapt to the business needs of strong spatiotemporal constraints and multi-device collaborative dependence in the scheduling scenario of metal mines.

[0064] In some embodiments, the text entity lexical includes at least one of mining equipment, trackless transport vehicles, hoisting ore passes, and workers in a metal mine scheduling scenario.

[0065] In some embodiments, the aforementioned metal mine scheduling topology dependency graph can be defined as follows: ,in, For a set of nodes, Define the set of edges and assign an initial feature vector to each node; Using graph attention networks to apply topological dependency graphs to metal mine scheduling During the process of node feature aggregation, its first In the round of iteration, nodes eigenvectors Based on its local neighbor nodes The features are updated using attention-weighted methods, and the calculation formula is as follows:

[0066] in, The weight matrix is ​​a learnable matrix; For nodes with neighboring nodes The attention coefficient between different scheduling entities is used to dynamically measure the strength of business association between them. It is a non-linear activation function; go through After round-robin network iteration, the topology dependency graph of metal mine scheduling is obtained. Global pooling is performed on the feature vectors of all nodes to obtain a graph-level high-dimensional semantic topological feature vector representing the overall logical structure of the scheduling data. :

[0067] in, For the image The total number of nodes in the system.

[0068] Step 106: Perform dimensionality reduction processing on the graph-level high-dimensional semantic topology feature vector to obtain the dimensionality-reduced semantic feature vector; use an adaptive clustering algorithm to perform semantic clustering on the semantic feature vector to obtain the semantic clustering result; and based on the semantic clustering result, determine multiple scheduling fine-tuning datasets from the candidate fine-tuning datasets, and generate a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios based on the multiple scheduling fine-tuning datasets.

[0069] The target quality scheduling fine-tuning dataset is used to fine-tune the pre-trained general-purpose base language model to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling.

[0070] The aforementioned target quality scheduling fine-tuning dataset can be understood as a scheduling fine-tuning dataset with high data quality.

[0071] In this embodiment, by performing dimensionality reduction processing on the graph-level high-dimensional semantic topological feature vector, the computational complexity of high-dimensional graph features can be reduced and redundancy can be eliminated.

[0072] In some embodiments, the step of performing dimensionality reduction processing on the graph-level high-dimensional semantic topological feature vector to obtain a dimensionality-reduced semantic feature vector; using an adaptive clustering algorithm to semantically cluster the semantic feature vector to obtain a semantic clustering result; and based on the semantic clustering result, determining multiple scheduling fine-tuning datasets from the candidate fine-tuning datasets, and generating a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios based on the multiple scheduling fine-tuning datasets, including: The graph-level high-dimensional semantic topological feature vector is reduced in dimensionality using a feature orthogonal mapping technique based on principal component analysis to obtain a reduced semantic feature vector. An adaptive clustering algorithm is used to semantically cluster the semantic feature vectors to obtain semantic clustering results. Based on the semantic clustering results, the candidate fine-tuning dataset is divided into multiple sub-scheduling scenario clusters. Based on the comprehensive evaluation score, the sub-candidate fine-tuning datasets of each sub-scheduling scenario cluster are sorted in descending order to obtain the ranking result of each sub-scheduling scenario cluster. Based on the sorting results, the top-ranked sub-candidate fine-tuning datasets in each sub-scheduling scenario cluster are extracted to obtain multiple scheduling fine-tuning datasets. Perform a union operation on the multiple scheduling fine-tuning datasets to generate a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios.

[0073] The aforementioned candidate fine-tuning dataset can be understood as a set including N sub-candidate fine-tuning datasets.

[0074] Specifically, in the process of dimensionality reduction of graph-level high-dimensional semantic topological feature vectors, the graph-level high-dimensional semantic topological feature vectors can be calculated first. covariance matrix :

[0075] in, The total number of candidate fine-tuning datasets, It is the mean vector of all graph-level high-dimensional semantic topological feature vectors; For covariance matrix Perform eigenvalue decomposition and select the first... The projection matrix is ​​composed of the eigenvectors corresponding to the largest eigenvalues. Mapping the graph-level high-dimensional semantic topological feature vectors to a low-dimensional orthogonal feature subspace yields the dimensionality-reduced semantic feature vectors. :

[0076] In a low-dimensional orthogonal feature subspace, an adaptive clustering algorithm can be used to refine the dimensionality-reduced semantic feature vectors. Semantic clustering is performed to obtain semantic clustering results, and based on the semantic clustering results, the candidate fine-tuning dataset is divided into... Individual scheduling scenario clusters The objective function of the clustering algorithm is to minimize the sum of squared errors within each cluster.

[0077] in, For the first The cluster center vector of each cluster. This represents the square of the Euclidean distance between the reduced semantic feature vector and the cluster center of its respective cluster; Based on the comprehensive evaluation score, each sub-scheduling scenario cluster... The sub-candidate fine-tuning datasets are sorted in descending order to obtain each sub-scheduling scenario cluster. The sorting results; The preferred sampling threshold for each cluster can be set. Perform partial truncation sampling and extract the results from the ranking results where the overall evaluation score is higher than [the specified value]. The sub-candidate fine-tuning datasets are used to obtain multiple scheduling fine-tuning datasets; By performing a union operation on the optimization results of each cluster, a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios is generated. :

[0078] in, For the candidate scheduling fine-tuning dataset, For this data sample in the cluster According to The ranking of scores in descending order.

[0079] Each of the above sub-scheduling scenario clusters The preferred results include this sub-scheduling scenario cluster. The overall evaluation score in the ranking results is higher than The sub-candidate fine-tuning dataset.

[0080] In some embodiments, based on the target quality scheduling fine-tuning dataset, the pre-trained general-purpose base language model is fine-tuned to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling, including the following steps: Obtain a pre-trained general-purpose base language model, and freeze all pre-trained weight matrices of the general-purpose base language model. This ensures that the pre-trained weight matrix does not participate in gradient updates during backpropagation; In the self-attention mechanism network layer of the general-based large language model, a trainable low-rank dimensionality-reduced matrix is ​​injected via a bypass. and the increasing dimension matrix ,in The set low-rank hyperparameters, and satisfying ; Metal mine scheduling fine-tuning data pairs are extracted in batches from the target quality scheduling fine-tuning dataset. , The scheduling information in the target quality scheduling fine-tuning dataset, The response information corresponding to the scheduling information; As context input, when passing through the network layer injected with the low-rank matrix, its forward propagation calculation formula is:

[0081] in, This is the input feature vector of the current network layer. This is the output feature vector of the fine-tuned network layer. This is the parameter matrix that needs to be updated for the low-rank fine-tuning part; An autoregressive language model trained based on causal language modeling objectives is used to compute target scheduling information. Given the historical lexical sequence, generate the target scheduling response. The cross-entropy loss of each word in the text, and the overall fine-tuning of the objective function. Defined as:

[0082] in, For the target quality scheduling fine-tuning dataset, The total length of the target scheduling response. For the first in the response sequence Each word element; During the backpropagation process of training, only the low-rank reduced-dimensional matrix is ​​processed. With increasing dimension matrix The parameters are calculated and adaptively updated. After training convergence, information injection is completed to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling.

[0083] In this embodiment, by fine-tuning the pre-trained general-purpose base language model based on the target quality scheduling fine-tuning dataset obtained from the dataset to be screened, the large-scale metal mine scheduling model can meet the application requirements of complex scheduling scenarios in metal mines. Moreover, the efficient fine-tuning strategy of injecting low-rank matrix parameters through bypass significantly reduces the computational cost of localized training.

[0084] The data screening method for fine-tuning large-scale scheduling models in metal mines provided in this application constructs a scheduling alignment sample set adapted to fine-tuning of a general-based large language model through a scheduling data set, trains a scheduling question-answering quality assessment model with a dynamic weight loss mechanism, and constructs a scoring system that includes normalized quality score, mutual information correlation score, and high-dimensional semantic consistency reordering evaluation score. This method can effectively avoid simple sample bias in the scheduling fine-tuning dataset and improve the data quality of the candidate fine-tuning dataset.

[0085] Furthermore, by constructing a topological dependency graph of the metal mine scheduling of the candidate fine-tuning dataset, and performing graph attention feature aggregation and dimensionality reduction clustering, the candidate fine-tuning dataset is locally optimized and sampled according to the comprehensive evaluation score. Finally, a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios is obtained. The target quality scheduling fine-tuning dataset provides a data foundation for constructing an efficient, high-quality, and highly safe large-scale metal mine scheduling model fine-tuning.

[0086] See Figure 2 , Figure 2 This is a schematic diagram of the structure of a metal mine scheduling large model fine-tuning data filtering device provided in an embodiment of this application. Figure 2 As shown, the device 200 includes: The construction module 201 is used to extract scheduling data sets under different scheduling scenarios based on the historical scheduling logs and business data of the metal mine, and to construct a scheduling alignment sample set adapted to the fine-tuning of the general base large language model based on the scheduling data set. Training module 202 is used to train the initial scheduling question-answering evaluation model based on the scheduling alignment sample set to obtain a scheduling question-answering quality evaluation model for quality verification of scheduling data; The first acquisition module 203 is used to acquire the normalized quality score of the dataset to be screened and fine-tuned based on the scheduling question-answering quality assessment model; acquire the mutual information correlation score of the dataset to be screened and fine-tuned based on autoregressive language modeling; and acquire the high-dimensional semantic consistency reordering assessment score of the dataset to be screened and fine-tuned based on a pre-trained semantic representation model. The second acquisition module 204 is used to multiply and fuse the normalized quality score, the mutual information correlation score and the high-dimensional semantic consistency reordering evaluation score to generate a comprehensive evaluation score for the dataset to be screened and fine-tuned; the comprehensive evaluation score is used as a quality evaluation dimension and is successively incorporated into the corresponding dataset to be screened and fine-tuned, and a candidate fine-tuning dataset containing multi-dimensional scoring labels is formed by splicing and mapping. The aggregation module 205 is used to construct a metal mine scheduling topology dependency graph based on the text entity lexical units of the candidate fine-tuning dataset, and to perform node feature aggregation on the metal mine scheduling topology dependency graph through a graph attention network to obtain a graph-level high-dimensional semantic topology feature vector. The generation module 206 is used to perform dimensionality reduction processing on the graph-level high-dimensional semantic topology feature vector to obtain a dimensionality-reduced semantic feature vector; to perform semantic clustering on the semantic feature vector using an adaptive clustering algorithm to obtain a semantic clustering result; and to determine multiple scheduling fine-tuning datasets from the candidate fine-tuning dataset based on the semantic clustering result; and to generate a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios based on the multiple scheduling fine-tuning datasets. The target quality scheduling fine-tuning dataset is used to fine-tune the pre-trained general-purpose base language model to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling.

[0087] Optionally, the training module 202 is specifically used for: Input any sample from the scheduling alignment sample set into the initial scheduling question-answering evaluation model, and output the standard quality score for positive samples and the standard quality score for negative samples; Using a paired comparison probability framework, the preference probability that the positive sample is better than the negative sample is calculated; Based on the preference probability, a dynamic adaptive weight loss function based on difficult example mining is constructed to automatically reduce the gradient weight of high-confidence samples of simple tasks in metal mine scheduling during backpropagation. This forces the initial scheduling question-answering evaluation model to focus on complex scheduling difficulties in metal mine scheduling. The application scenarios of these complex scheduling difficulties include multi-device collaborative scheduling scenarios, spatiotemporal constraint scheduling scenarios, and safety rigid constraint scheduling scenarios. The model parameters of the initial scheduling question-and-answer evaluation model are adjusted based on the dynamic adaptive weight loss function to obtain a scheduling question-and-answer quality evaluation model for quality verification of scheduling data.

[0088] Optionally, the first acquisition module 203 is specifically used for: Obtain the fine-tuning dataset to be filtered, wherein each fine-tuning data in the fine-tuning dataset includes metal mine problem information and corresponding answer information; Obtain the initial quality score of the dataset to be screened and fine-tuned by the scheduling question-answering quality assessment model, and map the initial quality score to the desired dataset using the max-min normalization method. The interval is used to obtain the normalized quality score of the dataset to be screened and fine-tuned; Based on autoregressive language modeling, the conditional generation information entropy is calculated to generate the corresponding response information under the condition of the target metal mine problem information; the unconditional generation information entropy is calculated under the condition of no scheduling information, and the ratio of the conditional generation information entropy to the unconditional generation information entropy is calculated to obtain the mutual information correlation score of the dataset to be screened and fine-tuned; wherein, the target metal mine problem information condition is the metal mine problem information of any fine-tuned data in the dataset to be screened and fine-tuned. A pre-trained semantic representation model is used to extract dense feature vectors of the metal mine problem information and the response information, respectively. The cosine similarity between the dense feature vectors of the metal mine problem information and the response information is calculated, and the cosine similarity is linearly shifted and normalized to obtain the high-dimensional semantic consistency reordering evaluation score of the dataset to be screened and fine-tuned.

[0089] Optionally, the aggregation module 205 is specifically used for: Extract text entity words from the candidate fine-tuning dataset as nodes; Based on the production scheduling rules and spatiotemporal constraints of metal mines, directed edges are established between the nodes to construct a topological dependency graph for metal mine scheduling. A graph attention network is used to aggregate node features in the scheduling topology dependency graph of the metal mine to obtain a graph-level high-dimensional semantic topology feature vector that represents the overall logical structure of the scheduling data.

[0090] Optionally, the generation module 206 is specifically used for: The graph-level high-dimensional semantic topological feature vector is reduced in dimensionality using a feature orthogonal mapping technique based on principal component analysis to obtain a reduced semantic feature vector. An adaptive clustering algorithm is used to semantically cluster the semantic feature vectors to obtain semantic clustering results. Based on the semantic clustering results, the candidate fine-tuning dataset is divided into multiple sub-scheduling scenario clusters. Based on the comprehensive evaluation score, the sub-candidate fine-tuning datasets of each sub-scheduling scenario cluster are sorted in descending order to obtain the ranking result of each sub-scheduling scenario cluster. Based on the sorting results, the top-ranked sub-candidate fine-tuning datasets in each sub-scheduling scenario cluster are extracted to obtain multiple scheduling fine-tuning datasets. Perform a union operation on the multiple scheduling fine-tuning datasets to generate a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios.

[0091] Optionally, based on the target quality scheduling fine-tuning dataset, the pre-trained general-purpose base language model is fine-tuned to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling, including the following steps: Obtain a pre-trained general-purpose base language model, and freeze all pre-trained weight matrices of the general-purpose base language model. This ensures that the pre-trained weight matrix does not participate in gradient updates during backpropagation; In the self-attention mechanism network layer of the general-based large language model, a trainable low-rank dimensionality-reduced matrix is ​​injected via a bypass. and the increasing dimension matrix ,in The set low-rank hyperparameters, and satisfying ; Metal mine scheduling fine-tuning data pairs are extracted in batches from the target quality scheduling fine-tuning dataset. , The scheduling information in the target quality scheduling fine-tuning dataset, The response information corresponding to the scheduling information; As context input, when passing through the network layer injected with the low-rank matrix, its forward propagation calculation formula is:

[0092] in, This is the input feature vector of the current network layer. This is the output feature vector of the fine-tuned network layer. This is the parameter matrix that needs to be updated for the low-rank fine-tuning part; An autoregressive language model trained based on causal language modeling objectives is used to compute target scheduling information. Given the historical lexical sequence, generate the target scheduling response. The cross-entropy loss of each word in the text, and the overall fine-tuning of the objective function. Defined as:

[0093] in, For the target quality scheduling fine-tuning dataset, The total length of the target scheduling response. For the first in the response sequence Each word element; During the backpropagation process of training, only the low-rank reduced-dimensional matrix is ​​processed. With increasing dimension matrix The parameters are calculated and adaptively updated. After training convergence, information injection is completed to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling.

[0094] The metal mine scheduling large-scale model fine-tuning data filtering device provided in this application can achieve the functions described in this application. Figure 1 The various processes in the method embodiments, and the ways to achieve the same beneficial effects, will not be repeated here to avoid repetition.

[0095] like Figure 3 As shown, this application also provides a terminal device, including a processor 301 and a memory 302. The memory 302 stores a program or instructions that can run on the processor 301. When the program or instructions are executed by the processor 301, they implement the various steps of the above-described embodiment of the metal mine scheduling large model fine-tuning data filtering method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0096] It should be noted that the terminal device in this application can be a terminal or other devices besides a terminal. For example, the terminal device can be a mobile phone, tablet computer, laptop computer, etc., and this application does not make any specific limitation.

[0097] This application also provides a readable storage medium storing a program or instructions that, when executed by a processor, implements the various processes of the above-described embodiments of the metal mine scheduling large model fine-tuning data filtering method and achieves the same technical effect. To avoid repetition, these will not be described again here.

[0098] The processor is the processor in the terminal device described in the above embodiments. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (Read-Only Memory). Only memory (ROM), random access memory (RAM), magnetic disks or optical disks, etc.

[0099] This application also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described embodiments of the metal mine scheduling large model fine-tuning data screening method, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0100] It should be understood that the chip mentioned in this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0101] This application provides a computer program product stored in a storage medium. The program product is executed by at least one processor to implement the various processes of the above-described embodiment of the metal mine scheduling large model fine-tuning data screening method, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0102] 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. Without further limitations, 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. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0103] 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. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as a read-only memory). The device includes a number of instructions in a ROM (random access memory), RAM (magnetic disk), or optical disk to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0104] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for filtering fine-tuning data in a large-scale metal mine scheduling model, characterized in that, include: Based on historical scheduling logs and business data of metal mines, scheduling data sets under different scheduling scenarios are extracted, and a scheduling alignment sample set adapted to fine-tuning of a general-purpose base large language model is constructed based on the scheduling data sets. The initial scheduling question-and-answer evaluation model is trained based on the aforementioned scheduling alignment sample set to obtain a scheduling question-and-answer quality evaluation model for quality verification of scheduling data. Based on the aforementioned scheduling question-answering quality assessment model, the normalized quality score of the dataset to be screened and fine-tuned is obtained; Based on autoregressive language modeling, the mutual information correlation score of the dataset to be screened and fine-tuned is obtained; Furthermore, based on the pre-trained semantic representation model, the high-dimensional semantic consistency re-ranking evaluation score of the dataset to be screened and fine-tuned is obtained; The normalized quality score, the mutual information correlation score, and the high-dimensional semantic consistency reordering evaluation score are multiplicatively fused to generate a comprehensive evaluation score for the dataset to be screened for fine-tuning; the comprehensive evaluation score is used as a quality evaluation dimension and is successively incorporated into the corresponding dataset to be screened for fine-tuning, and a candidate fine-tuning dataset containing multi-dimensional scoring labels is formed by splicing and mapping. A metal mine scheduling topology dependency graph is constructed based on the text entity lexical units of the candidate fine-tuning dataset. The graph attention network is then used to aggregate node features of the metal mine scheduling topology dependency graph to obtain a graph-level high-dimensional semantic topology feature vector. The high-dimensional semantic topological feature vector at the graph level is subjected to dimensionality reduction processing to obtain the dimensionality-reduced semantic feature vector; the semantic feature vector is semantically clustered using an adaptive clustering algorithm to obtain the semantic clustering result; based on the semantic clustering result, multiple scheduling fine-tuning datasets are determined from the candidate fine-tuning dataset; and a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios is generated based on the multiple scheduling fine-tuning datasets. The target quality scheduling fine-tuning dataset is used to fine-tune the pre-trained general-purpose base language model to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling.

2. The method according to claim 1, characterized in that, The process of training the initial scheduling question-answering evaluation model based on the scheduling alignment sample set to obtain a scheduling question-answering quality evaluation model for quality verification of scheduling data includes: Input any sample from the scheduling alignment sample set into the initial scheduling question-answering evaluation model, and output the standard quality score for positive samples and the standard quality score for negative samples; Using a paired comparison probability framework, the preference probability that the positive sample is better than the negative sample is calculated; Based on the preference probability, a dynamic adaptive weight loss function based on difficult example mining is constructed to automatically reduce the gradient weight of high-confidence samples of simple tasks in metal mine scheduling during backpropagation. This forces the initial scheduling question-answering evaluation model to focus on complex scheduling difficulties in metal mine scheduling. The application scenarios of these complex scheduling difficulties include multi-device collaborative scheduling scenarios, spatiotemporal constraint scheduling scenarios, and safety rigid constraint scheduling scenarios. The model parameters of the initial scheduling question-and-answer evaluation model are adjusted based on the dynamic adaptive weight loss function to obtain a scheduling question-and-answer quality evaluation model for quality verification of scheduling data.

3. The method according to claim 2, characterized in that, The normalized quality score of the dataset to be screened and fine-tuned is obtained based on the scheduling question-answering quality assessment model; the mutual information correlation score of the dataset to be screened and fine-tuned is obtained based on autoregressive language modeling. And, based on the pre-trained semantic representation model, obtain the high-dimensional semantic consistency re-ranking evaluation score of the dataset to be screened and fine-tuned, including: Obtain the fine-tuning dataset to be filtered, wherein each fine-tuning data in the fine-tuning dataset includes metal mine problem information and corresponding answer information; Obtain the initial quality score of the dataset to be screened and fine-tuned by the scheduling question-answering quality assessment model, and map the initial quality score to the desired dataset using the max-min normalization method. The interval is used to obtain the normalized quality score of the dataset to be screened and fine-tuned; Based on autoregressive language modeling, the conditional generation information entropy is calculated to generate the corresponding response information under the condition of the target metal mine problem information; the unconditional generation information entropy is calculated under the condition of no scheduling information, and the ratio of the conditional generation information entropy to the unconditional generation information entropy is calculated to obtain the mutual information correlation score of the dataset to be screened and fine-tuned; wherein, the target metal mine problem information condition is the metal mine problem information of any fine-tuned data in the dataset to be screened and fine-tuned. A pre-trained semantic representation model is used to extract dense feature vectors of the metal mine problem information and the response information, respectively. The cosine similarity between the dense feature vectors of the metal mine problem information and the response information is calculated, and the cosine similarity is linearly shifted and normalized to obtain the high-dimensional semantic consistency reordering evaluation score of the dataset to be screened and fine-tuned.

4. The method according to any one of claims 1 to 3, characterized in that, The process involves constructing a metal mine scheduling topology dependency graph based on text entity lexical units from the candidate fine-tuning dataset, and then aggregating node features of the metal mine scheduling topology dependency graph using a graph attention network to obtain a graph-level high-dimensional semantic topology feature vector, including: Extract text entity words from the candidate fine-tuning dataset as nodes; Based on the production scheduling rules and spatiotemporal constraints of metal mines, directed edges are established between the nodes to construct a topological dependency graph for metal mine scheduling. A graph attention network is used to aggregate node features in the scheduling topology dependency graph of the metal mine to obtain a graph-level high-dimensional semantic topology feature vector that represents the overall logical structure of the scheduling data.

5. The method according to any one of claims 1 to 3, characterized in that, The process involves reducing the dimensionality of the graph-level high-dimensional semantic topological feature vector to obtain a dimensionality-reduced semantic feature vector; then, using an adaptive clustering algorithm, semantic clustering is performed on the semantic feature vector to obtain a semantic clustering result. Based on the semantic clustering result, multiple scheduling fine-tuning datasets are determined from the candidate fine-tuning dataset. Finally, a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios is generated based on these multiple scheduling fine-tuning datasets, including: The graph-level high-dimensional semantic topological feature vector is reduced in dimensionality using a feature orthogonal mapping technique based on principal component analysis to obtain a reduced semantic feature vector. An adaptive clustering algorithm is used to semantically cluster the semantic feature vectors to obtain semantic clustering results. Based on the semantic clustering results, the candidate fine-tuning dataset is divided into multiple sub-scheduling scenario clusters. Based on the comprehensive evaluation score, the sub-candidate fine-tuning datasets of each sub-scheduling scenario cluster are sorted in descending order to obtain the ranking result of each sub-scheduling scenario cluster. Based on the sorting results, the top-ranked sub-candidate fine-tuning datasets in each sub-scheduling scenario cluster are extracted to obtain multiple scheduling fine-tuning datasets. Perform a union operation on the multiple scheduling fine-tuning datasets to generate a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios.

6. The method according to any one of claims 1 to 3, characterized in that, Based on the target quality scheduling fine-tuning dataset, the pre-trained general-purpose base language model is fine-tuned to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling, including the following steps: Obtain a pre-trained general-purpose base language model, and freeze all pre-trained weight matrices of the general-purpose base language model. This ensures that the pre-trained weight matrix does not participate in gradient updates during backpropagation; In the self-attention mechanism network layer of the general-based large language model, a trainable low-rank dimensionality-reduced matrix is ​​injected via a bypass. and the increasing dimension matrix ,in The set low-rank hyperparameters, and satisfying ; Metal mine scheduling fine-tuning data pairs are extracted in batches from the target quality scheduling fine-tuning dataset. , The scheduling information in the target quality scheduling fine-tuning dataset, The response information corresponding to the scheduling information; As context input, when passing through the network layer injected with the low-rank matrix, its forward propagation calculation formula is: in, This is the input feature vector of the current network layer. This is the output feature vector of the fine-tuned network layer. This is the parameter matrix that needs to be updated for the low-rank fine-tuning part; An autoregressive language model trained based on causal language modeling objectives is used to compute target scheduling information. Given the historical lexical sequence, generate the target scheduling response. The cross-entropy loss of each word in the text, and the overall fine-tuning of the objective function. Defined as: in, For the target quality scheduling fine-tuning dataset, The total length of the target scheduling response. For the first in the response sequence Each word element; During the backpropagation process of training, only the low-rank reduced-dimensional matrix is ​​processed. With increasing dimension matrix The parameters are calculated and adaptively updated. After training convergence, information injection is completed to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling.

7. A data filtering device for fine-tuning a large-scale scheduling model of a metal mine, characterized in that, include: The module is used to extract scheduling data sets under different scheduling scenarios based on historical scheduling logs and business data of metal mines, and to build a scheduling alignment sample set adapted to fine-tuning of the general base large language model based on the scheduling data set. The training module is used to train the initial scheduling question-answering evaluation model based on the scheduling alignment sample set, so as to obtain a scheduling question-answering quality evaluation model for quality verification of scheduling data. The first acquisition module is used to acquire the normalized quality score of the dataset to be screened and fine-tuned based on the scheduling question-answering quality assessment model. Based on autoregressive language modeling, the mutual information correlation score of the dataset to be screened and fine-tuned is obtained; Furthermore, based on the pre-trained semantic representation model, the high-dimensional semantic consistency re-ranking evaluation score of the dataset to be screened and fine-tuned is obtained; The second acquisition module is used to multiply and fuse the normalized quality score, the mutual information correlation score and the high-dimensional semantic consistency reordering evaluation score to generate a comprehensive evaluation score for the dataset to be screened and fine-tuned; the comprehensive evaluation score is used as a quality evaluation dimension and is successively incorporated into the corresponding dataset to be screened and fine-tuned, and a candidate fine-tuning dataset containing multi-dimensional scoring labels is formed by splicing and mapping. The aggregation module is used to construct a metal mine scheduling topology dependency graph based on the text entity lexical units of the candidate fine-tuning dataset, and to aggregate the node features of the metal mine scheduling topology dependency graph through a graph attention network to obtain a graph-level high-dimensional semantic topology feature vector. The generation module is used to perform dimensionality reduction processing on the graph-level high-dimensional semantic topology feature vector to obtain the dimensionality-reduced semantic feature vector; to perform semantic clustering on the semantic feature vector using an adaptive clustering algorithm to obtain the semantic clustering result; and to determine multiple scheduling fine-tuning datasets from the candidate fine-tuning dataset based on the semantic clustering result; and to generate a target quality scheduling fine-tuning dataset with a balanced structure and covering multiple scheduling scenarios based on the multiple scheduling fine-tuning datasets. The target quality scheduling fine-tuning dataset is used to fine-tune the pre-trained general-purpose base language model to obtain a large-scale metal mine scheduling model suitable for metal mine scheduling.

8. The apparatus according to claim 7, characterized in that, The training module is specifically used for: Input any sample from the scheduling alignment sample set into the initial scheduling question-answering evaluation model, and output the standard quality score for positive samples and the standard quality score for negative samples; Using a paired comparison probability framework, the preference probability that the positive sample is better than the negative sample is calculated; Based on the preference probability, a dynamic adaptive weight loss function based on difficult example mining is constructed to automatically reduce the gradient weight of high-confidence samples of simple tasks in metal mine scheduling during backpropagation. This forces the initial scheduling question-answering evaluation model to focus on complex scheduling difficulties in metal mine scheduling. The application scenarios of these complex scheduling difficulties include multi-device collaborative scheduling scenarios, spatiotemporal constraint scheduling scenarios, and safety rigid constraint scheduling scenarios. The model parameters of the initial scheduling question-and-answer evaluation model are adjusted based on the dynamic adaptive weight loss function to obtain a scheduling question-and-answer quality evaluation model for quality verification of scheduling data.

9. The apparatus according to claim 8, characterized in that, The first acquisition module is specifically used for: Obtain the fine-tuning dataset to be filtered, wherein each fine-tuning data in the fine-tuning dataset includes metal mine problem information and corresponding answer information; Obtain the initial quality score of the dataset to be screened and fine-tuned by the scheduling question-answering quality assessment model, and map the initial quality score to the desired dataset using the max-min normalization method. The interval is used to obtain the normalized quality score of the dataset to be screened and fine-tuned; Based on autoregressive language modeling, the conditional generation information entropy of generating corresponding answer information is calculated under the condition of target metal mine problem information. Calculate the information entropy of unconditionally generated information under the condition of no scheduling information, and calculate the ratio of the information entropy of conditionally generated information to the information entropy of unconditionally generated information to obtain the mutual information correlation score of the dataset to be screened for fine-tuning; wherein, the target metal mine problem information condition is the metal mine problem information of any fine-tuning data in the dataset to be screened for fine-tuning. A pre-trained semantic representation model is used to extract dense feature vectors of the metal mine problem information and the response information, respectively. The cosine similarity between the dense feature vectors of the metal mine problem information and the response information is calculated, and the cosine similarity is linearly shifted and normalized to obtain the high-dimensional semantic consistency reordering evaluation score of the dataset to be screened and fine-tuned.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of 1 to 6.

Images