Multi-field data weighted mixing method for fine tuning of large mine model

Abstract

The invention relates to the technical field of fine adjustment of large mine models, in particular to a multi-field data weighted mixing method, device and equipment for fine adjustment of a large mine model and a computer storage medium. According to the method, the domain target weight and the sample multi-homing probability are accurately combined in the batch construction stage, controllable adjustment of the contribution proportion of each domain is achieved, and it is ensured that low-resource or key domains get full attention in training; in the loss calculation stage, multi-field loss is subjected to weighted fusion by using the attribution weight, so that the model learning process better fits the real multi-field distribution of the sample. Finally, according to the method, the performance balance and the overall generalization ability of the fine-tuned large mine model in each sub-field task are remarkably improved, and the adaptability to a complex multi-label scene is enhanced.

Description

Technical Field

[0001] This invention relates to the field of fine-tuning technology for large-scale mine models, and in particular to a method, apparatus, equipment, and computer storage medium for multi-domain data weighted mixing for fine-tuning of large-scale mine models. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, general-purpose large models have achieved remarkable results in fields such as natural language processing, question answering, and knowledge graphs. However, when applied to vertical industry scenarios, they often exhibit insufficient performance, especially in the field of mining safety where data is scarce and highly dependent on domain knowledge. To improve the practical application effect of large models in the mining industry, it is necessary to adapt general-purpose large models to specific mining tasks through fine-tuning.

[0003] In the mining industry, data often exhibits multi-domain characteristics, with a single data point potentially encompassing multiple sub-domains such as "underground mining," "support equipment," and "gas control." This multi-label, multi-domain nature poses a challenge to traditional fine-tuning methods: existing methods typically assume each sample belongs to only a single domain and sample only based on the number of samples during batch construction, making it difficult to precisely control the contribution ratio of each domain during training. This can easily lead to underperformance of the model in some domains or unbalanced training results.

[0004] Some existing domain reweighting methods attempt to alleviate inter-domain imbalance by assigning weights to different domains, thereby increasing the sampling probability of small-sample domain data. However, these methods still have shortcomings in multi-domain scenarios: (1) The sample may involve multiple fields at the same time, and a single sampling probability is difficult to fully reflect its contribution in each field; (2) Batch construction relies heavily on random sampling, making it difficult to strictly guarantee that each domain achieves the expected target weight distribution; Therefore, in the post-training fine-tuning task of large mining models, there is an urgent need for a multi-domain data mixing method that can simultaneously consider the multi-domain correlation of samples, accurately construct training batches based on domain target weights, and support dynamic adaptive weighting, so as to improve the performance of the model in each subdomain while maintaining the overall generalization ability. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problem that in the mining scenario, a single sample often involves multiple fields at the same time, making it difficult to accurately represent its affiliation with a single label, thus leading to insufficient data utilization and training bias.

[0006] To address the aforementioned technical problems, this invention provides a multi-domain data weighted mixing method for fine-tuning large-scale mine models, comprising: Based on a multi-domain, multi-label training dataset for mining, the probability of each training sample belonging to multiple domains is calculated by manually annotating by domain experts, predicting by domain classifiers, or combining data statistical features and text similarity. This generates the original multi-domain belonging weight vector and then performs smooth normalization to obtain the smooth normalized belonging weight vector. Based on the smooth normalized set of attribution weight vectors and the pre-defined batch target weights for each domain, the sampling probability of each training sample in the current training batch is calculated. Based on the sampling probability, the current batch training data is generated by random sampling from the multi-domain multi-label training dataset of the mine, so that the probability of each training sample being selected into the current training batch is proportional to its sampling probability. For each training sample in the current batch of training data, calculate the weighted loss of each training sample based on its smoothed normalized home weight vector and the loss of each domain, and calculate the total loss of the current batch based on the weighted loss of all training samples. The large mining model is fine-tuned by calculating gradients through backpropagation and using an optimizer to iteratively update model parameters to minimize the total loss of the current batch.

[0007] Preferably, the method of calculating the affiliation probability of each training sample in multiple domains based on the multi-domain, multi-label training dataset of the mine, and generating the original multi-domain affiliation weight vector by manually annotating the dataset by domain experts, predicting with a domain classifier, or combining data statistical features and text similarity, includes: Based on expert knowledge, each training sample is pre-labeled with its domain affiliation probability as its affiliation weight, ensuring that the sum of the N domain affiliation weights for the same training sample is 1, or Using a pre-trained domain classifier, each training sample is classified using multi-label classification. The normalized confidence scores of each class are used as their respective weights. Extract the textual or statistical features of each training sample, calculate its similarity with representative features of each domain, and use the normalized similarity as the attribution weight. Finally, an original multi-domain attribution weight vector is generated for each training sample.

[0008] Preferably, the step of generating the current batch of training data by random sampling from the multi-domain, multi-label training dataset of the mine based on sampling probability, such that the probability of each training sample being selected into the current training batch is proportional to its sampling probability, includes: Based on sampling probability, a predetermined number of samples are drawn from the multi-domain, multi-label training dataset of the mine using random sampling without replacement to form the current batch of training data. In this sampling process, the probability of each training sample being selected is proportional to its sampling probability.

[0009] Preferably, the step of calculating the weighted loss of each training sample in the current batch of training data based on its smoothed normalized home weight vector and the loss of each domain, and calculating the total loss of the current batch based on the weighted loss of all training samples includes: Each training sample in the current batch of training data is input into the general large model to be fine-tuned, and forward propagation is performed to obtain the model prediction result of each training sample. For each domain, the loss value of each training sample in that domain is calculated. Based on the loss value of each training sample in each domain and the smoothed normalized assignment weight vector of each training sample, calculate the weighted loss of each training sample. The total loss for this training batch is obtained by summing the weighted losses of each training sample in the current batch of training data.

[0010] Preferably, the step of fine-tuning the large mining model by calculating the gradient through backpropagation and iteratively updating the model parameters using an optimizer to minimize the total loss of the current batch includes: With the total loss of the current batch as the optimization objective, the gradient of the loss function with respect to all trainable parameters of the model is calculated through the backpropagation algorithm, and the pre-selected optimizer updates the model parameters based on the calculated gradient in order to reduce the total loss of the current batch. The next round of forward propagation, loss calculation, and parameter updates is performed using new training batches until the preset training rounds or convergence conditions are met, resulting in a specialized large-scale mining model adapted to the needs of the mining industry, which has been fine-tuned by weighted mixing of multi-domain data.

[0011] This invention also provides a multi-domain data weighted mixing device for fine-tuning large-scale mine models, comprising: The module for obtaining the smoothed and normalized attribution weight vector is used to calculate the attribution probability of each training sample in multiple domains based on a multi-domain and multi-label training dataset of a mine. This is achieved through manual annotation by domain experts, prediction by domain classifiers, or by combining data statistical features and text similarity. The module generates the original multi-domain attribution weight vector and then performs smoothing and normalization processing to obtain the smoothed and normalized attribution weight vector. The batch training data determination module is used to calculate the sampling probability of each training sample in the current training batch based on the smooth normalized set of attribution weight vectors and the preset batch target weights of each domain. Based on the sampling probability, the current batch training data is generated by random sampling from the multi-domain multi-label training dataset of the mine, so that the probability of each training sample being selected into the current training batch is proportional to its sampling probability. The batch total loss determination module is used to calculate the weighted loss of each training sample in the current batch of training data based on its smoothed normalized home weight vector and the loss of each domain, and to calculate the total loss of the current batch based on the weighted loss of all training samples. The model training module is used to calculate gradients through backpropagation and use the optimizer to iteratively update model parameters to minimize the total loss of the current batch, thus fine-tuning the large mining model.

[0012] Preferably, the method of calculating the affiliation probability of each training sample in multiple domains based on the multi-domain, multi-label training dataset of the mine, and generating the original multi-domain affiliation weight vector by manually annotating the dataset by domain experts, predicting with a domain classifier, or combining data statistical features and text similarity, includes: Based on expert knowledge, each training sample is pre-labeled with its domain affiliation probability as its affiliation weight, ensuring that the sum of the N domain affiliation weights for the same training sample is 1, or Using a pre-trained domain classifier, each training sample is classified using multi-label classification. The normalized confidence scores of each class are used as their respective weights. Extract the textual or statistical features of each training sample, calculate its similarity with representative features of each domain, and use the normalized similarity as the attribution weight. Finally, an original multi-domain attribution weight vector is generated for each training sample.

[0013] This invention also provides a multi-domain data weighted mixing device for fine-tuning large-scale mine models, comprising: Memory, used to store computer programs; A processor is used to implement the steps of the above-described method for fine-tuning a large-scale mine model using multi-domain data weighted mixing when executing the computer program.

[0014] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for fine-tuning a large-scale mine model using multi-domain data weighted mixing.

[0015] The technical solution of the present invention has the following advantages compared with the prior art: The multi-domain data weighted mixing method for fine-tuning large-scale mining models described in this invention precisely combines domain target weights with sample multi-attribution probabilities during the batch construction phase, enabling controllable adjustment of the contribution ratio of each domain and ensuring that low-resource or critical domains receive sufficient attention during training. During the loss calculation phase, the attribution weights are used to weight and fuse multi-domain losses, making the model learning process more closely reflect the real multi-domain distribution of the samples. Ultimately, this method significantly improves the performance balance and overall generalization ability of the fine-tuned large-scale mining model across sub-domain tasks and enhances its adaptability to complex multi-label scenarios. This invention, through soft attribution modeling and dynamic weight mixing of multi-domain, multi-label mining samples, effectively solves the problems of uneven data utilization and training skew caused by fixed single-domain partitioning and simple random sampling in traditional fine-tuning methods. Attached Figure Description

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the implementation of a multi-domain data weighted mixing method for fine-tuning large-scale mine models provided by this invention. Figure 2 This is a structural block diagram of a multi-domain data weighted mixing device for fine-tuning large-scale mining models provided in an embodiment of the present invention. Detailed Implementation

[0017] The core of this invention is to provide a multi-domain data weighted mixing method, apparatus, equipment, and computer storage medium for fine-tuning large-scale mining models, which effectively solves the problems of uneven data utilization and training skew caused by fixed sample single-domain division and simple random sampling in traditional fine-tuning methods.

[0018] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please refer to Figure 1. Figure 1 The flowchart illustrates the implementation of a multi-domain data weighted mixing method for fine-tuning large-scale mine models provided by this invention; the specific operation steps are as follows: S101: Based on the multi-domain multi-label training dataset of mine, the probability of each training sample belonging to multiple domains is calculated by manual annotation by domain experts, prediction by domain classifiers, or by combining data statistical features and text similarity. The original multi-domain belonging weight vector is generated and then smoothed and normalized to obtain the smoothed and normalized belonging weight vector. S102: Based on the smooth normalized set of attribution weight vectors and the pre-set batch target weights of each domain, calculate the sampling probability of each training sample in the current training batch, and based on the sampling probability, generate the current batch training data by random sampling from the multi-domain multi-label training dataset of the mine, so that the probability of each training sample being selected into the current training batch is proportional to its sampling probability. S103: For each training sample in the current batch of training data, calculate the weighted loss of each training sample based on its smoothed normalized home weight vector and the loss of each domain, and calculate the total loss of the current batch based on the weighted loss of all training samples. S104: Fine-tuning the large mining model by calculating gradients through backpropagation and iteratively updating model parameters using the optimizer to minimize the total loss of the current batch.

[0020] Based on the above embodiments, this embodiment will provide a detailed description of step S101: In some embodiments, based on a multi-domain, multi-label training dataset from a mining area, the probability of each training sample belonging to multiple domains is calculated for each sample through methods such as manual annotation by domain experts, prediction by domain classifiers, or a combination of data statistical features and text similarity, generating an original multi-domain belonging weight vector including: Using a mining dataset containing multiple training samples as input, calculate the original affiliation weights for each sample i within N predefined domains (e.g., "underground mining," "support equipment," "gas control," etc.). This process is implemented using one of the following three methods: Based on expert knowledge, each training sample is pre-labeled with its domain affiliation probability as its affiliation weight, ensuring that the sum of the N domain affiliation weights for the same training sample is 1, or Using a pre-trained domain classifier, each training sample is classified using multi-label classification. The normalized confidence scores of each class are used as their respective weights. Extract the textual or statistical features of each training sample, calculate its similarity with representative features of each domain (such as domain keywords and typical sample feature vectors), and use the normalized similarity as the attribution weight. Finally, an original multi-domain attribution weight vector is generated for each training sample:

[0021] The variables are described below:

[0022] Attribution probability It can be determined directly by domain experts through manual annotation, or the confidence level of samples in various domains can be predicted with the help of domain classifiers or pre-trained models. It can also automatically generate the corresponding probability distribution by combining the statistical characteristics of the data and the similarity of the text content.

[0023] Example: When N=4, the probability of a sample being assigned is [0.5,0,0,0.5], which means that the sample contributes a weight of 0.5 to both the 1st and 4th domains.

[0024] In some embodiments, to handle data noise or label uncertainty and prevent training imbalance caused by some domain weights being too large or too small, the original multi-domain weight vector is modified. Perform smoothing and normalization: Based on the original multi-domain attribution weight vector and a preset smoothing coefficient The original weights for each domain are smoothed. Smoothing aims to prevent some weights from being 0 or 1 (e.g., samples are completely assigned to a certain domain) to enhance training stability.

[0025] To ensure that the sum of the weights in each domain remains 1 after smoothing, the smoothed weight vector is normalized:

[0026]

[0027]

[0028] Represents the d-th domain For smoothing

[0029] .

[0030] Based on the above embodiments, this embodiment will provide a detailed description of step S102: In some embodiments, calculating the sampling probability of each training sample in the current training batch based on the smoothed normalized set of attribution weight vectors and the pre-defined batch target weights for each domain includes: Based on the needs of mining operations or expert experience, assign a batch target weight to each domain that reflects the relative importance or contribution ratio desired for that domain during model training, rather than the number of samples, to ensure that each domain receives a reasonable contribution during training.

[0031] Give higher weight to key or critical areas of focus to ensure they receive sufficient attention and make a greater contribution during training.

[0032] Based on the smoothed normalized set of attribution weight vectors of all training samples and the pre-set batch target weights of each domain, the total contribution of each training sample to each domain according to its attribution weight and the sum of the total contributions of all training samples are calculated. Calculate the sampling probability of each training sample in the current training batch based on the sum of the total contribution of each training sample to each domain according to its affiliation weight and the total contribution of all training samples:

[0033] in:

[0034] The molecule reflects the "total contribution" of sample i to each target domain according to its affiliation weight; The denominator is the sum of the total contributions of all samples, used to normalize the contribution of each sample into a probability. Therefore, This represents the probability that sample i is selected for the next training batch. The sample is in a high-importance domain (high...). The attribution weight (high) The higher the ), the higher its sampling probability. The larger. This is used to resample samples belonging to low-representation or high-priority domains, thereby making the training data distribution closer to the expected domain mixing ratio.

[0035] In some embodiments, training data for the current batch is generated by random sampling from a multi-domain, multi-label training dataset of a mine, based on sampling probability, such that the probability of each training sample being selected into the current training batch is proportional to its sampling probability, including: Based on sampling probability, a predetermined number of samples (e.g., batch size B) are drawn from the multi-domain, multi-label training dataset of the mine using random sampling without replacement (or sampling with replacement, depending on the implementation). These samples constitute the current batch of training data. During this sampling process, the probability of each training sample being selected is proportional to its sampling probability. In this way, statistically speaking, the proportion of total contribution from each domain represented by the samples in each training batch L will approximate the predetermined target weight distribution.

[0036] Based on the above embodiments, this embodiment will provide a detailed description of step S103: In some embodiments, for each training sample in the current batch of training data, the weighted loss of each training sample is calculated based on its smoothed normalized attribution weight vector and the loss of each domain, and the total loss of the current batch is calculated based on the weighted loss of all training samples, including: Each training sample in the current batch of training data is input into the general large model to be fine-tuned, and forward propagation is performed to obtain the model prediction result of each training sample. For each domain (or for each domain subtask under the multi-task learning setting), the loss value (e.g., cross-entropy loss, mean squared error, etc.) related to each training sample in that domain is calculated. Based on the loss value of each training sample in each domain and the smoothed normalized assignment weight vector of each training sample, the weighted loss of each training sample is calculated. This weighted loss reflects the contribution of sample i to the loss of its respective domain, and the weights are determined by its assignment weights. Decide:

[0037]

[0038] :

[0039]

[0040] Total number of samples in the batch dataset The total loss for this training batch is obtained by summing the weighted losses of each training sample in the current batch of training data.

[0041] Based on the above embodiments, this embodiment will provide a detailed description of step S104: In some embodiments, fine-tuning the large mining model involves calculating gradients through backpropagation and iteratively updating model parameters using an optimizer to minimize the total loss of the current batch. Using the total loss of the current batch as the optimization objective, the gradient of the loss function with respect to all trainable parameters of the model is calculated through the backpropagation algorithm. The pre-selected optimizer (such as Adam, SGD, etc.) updates the model parameters based on the calculated gradient to reduce the total loss of the current batch. (Using the updated model may recalculate the sampling probability or sample as originally planned.) The next round of forward propagation, loss calculation, and parameter updates will be performed using a new training batch. The goal of each update is to reduce the overall weighted loss, so that the model can be closer to the distribution of real samples during continuous training, thereby gradually improving its adaptability to mining industry tasks until the preset training rounds or convergence conditions are reached, resulting in a dedicated large-scale mining model that has been fine-tuned by multi-domain data weighted mixing and is adapted to the needs of the mining industry.

[0042] This invention achieves the following technical effects in fine-tuning of large-scale mine models through a multi-domain data weighted mixing method: (1) Improve multi-domain performance. By performing multi-domain attribution modeling on each sample and dynamically constructing training batches according to domain weights, this invention can effectively balance the contributions of samples from different sub-domains in training and significantly improve the fine-tuning performance of the model in each sub-domain.

[0043] (2) Enhance the generalization ability of the model. An adaptive weight adjustment mechanism is adopted to dynamically adjust the target weights according to the performance of each domain during training, so that the model can receive more attention in low-performance domains, thereby improving the overall generalization ability and reducing the bias caused by uneven data distribution or label uncertainty.

[0044] (3) Precise control of domain contribution. By using batch construction and weight allocation methods, the contribution ratio of each domain in each training batch can be accurately controlled, and even if the samples belong to multiple domains, the training weight of each domain can be effectively reflected.

[0045] (4) Adaptable to complex multi-domain scenarios. The method of this invention can be flexibly adapted to complex data with multi-label and multi-domain characteristics in the mining industry, and supports random or deterministic batch construction strategies, providing a reliable and controllable training scheme for practical applications.

[0046] Please refer to Figure 2 , Figure 2 This invention provides a structural block diagram of a multi-domain data weighted mixing device for fine-tuning large-scale mine models; the specific device may include: The module for obtaining the smoothed and normalized attribution weight vector is used to calculate the attribution probability of each training sample in multiple domains based on a multi-domain and multi-label training dataset of a mine. This is achieved through manual annotation by domain experts, prediction by domain classifiers, or by combining data statistical features and text similarity. The module generates the original multi-domain attribution weight vector and then performs smoothing and normalization processing to obtain the smoothed and normalized attribution weight vector. The batch training data determination module is used to calculate the sampling probability of each training sample in the current training batch based on the smooth normalized set of attribution weight vectors and the preset batch target weights of each domain. Based on the sampling probability, the current batch training data is generated by random sampling from the multi-domain multi-label training dataset of the mine, so that the probability of each training sample being selected into the current training batch is proportional to its sampling probability. The batch total loss determination module is used to calculate the weighted loss of each training sample in the current batch of training data based on its smoothed normalized home weight vector and the loss of each domain, and to calculate the total loss of the current batch based on the weighted loss of all training samples. The model training module is used to calculate gradients through backpropagation and use the optimizer to iteratively update model parameters to minimize the total loss of the current batch, thus fine-tuning the large mining model.

[0047] The multi-domain data weighted mixing device for fine-tuning large-scale mining models in this embodiment is used to implement the aforementioned multi-domain data weighted mixing method for fine-tuning large-scale mining models. Therefore, the specific implementation of the multi-domain data weighted mixing device for fine-tuning large-scale mining models can be found in the previous embodiment section of the multi-domain data weighted mixing method for fine-tuning large-scale mining models. For example, the smoothing and normalizing attribution weight vector acquisition module, the batch training data determination module, the batch total loss determination module, and the model training module are respectively used to implement steps S101, S102, S103, and S104 in the aforementioned multi-domain data weighted mixing method for fine-tuning large-scale mining models. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0048] A specific embodiment of the present invention also provides a multi-domain data weighted mixing device for fine-tuning large-scale mine models, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the aforementioned multi-domain data weighted mixing method for fine-tuning large-scale mine models.

[0049] A specific embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for fine-tuning a large-scale mine model using multi-domain data weighted mixing.

[0050] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0051] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0053] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0054] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A multi-domain data weighted mixing method for fine-tuning large-scale mine models, characterized in that, include: Based on a multi-domain, multi-label training dataset for mining, the probability of each training sample belonging to multiple domains is calculated by manually annotating by domain experts, predicting by domain classifiers, or combining data statistical features and text similarity. This generates the original multi-domain belonging weight vector and then performs smooth normalization to obtain the smooth normalized belonging weight vector. Based on the smooth normalized set of attribution weight vectors and the pre-defined batch target weights for each domain, the sampling probability of each training sample in the current training batch is calculated. Based on the sampling probability, the current batch training data is generated by random sampling from the multi-domain multi-label training dataset of the mine, so that the probability of each training sample being selected into the current training batch is proportional to its sampling probability. For each training sample in the current batch of training data, calculate the weighted loss of each training sample based on its smoothed normalized home weight vector and the loss of each domain, and calculate the total loss of the current batch based on the weighted loss of all training samples. The large mining model is fine-tuned by calculating gradients through backpropagation and using an optimizer to iteratively update model parameters to minimize the total loss of the current batch.

2. The multi-domain data weighted mixing method for fine-tuning large-scale mine models according to claim 1, characterized in that, The multi-domain, multi-label training dataset based on mining operations calculates the probability of each training sample belonging to multiple domains through methods such as manual annotation by domain experts, prediction by domain classifiers, or a combination of data statistical features and text similarity. This generates the original multi-domain belonging weight vector, which includes: Based on expert knowledge, each training sample is pre-labeled with its domain affiliation probability as its affiliation weight, ensuring that the sum of the N domain affiliation weights for the same training sample is 1, or Using a pre-trained domain classifier, each training sample is classified using multi-label classification. The normalized confidence scores of each class are used as their respective weights. Extract the textual or statistical features of each training sample, calculate its similarity with representative features of each domain, and use the normalized similarity as the attribution weight. Finally, an original multi-domain attribution weight vector is generated for each training sample.

3. The multi-domain data weighted mixing method for fine-tuning large-scale mine models according to claim 1, characterized in that, The calculation of the sampling probability of each training sample in the current training batch, based on the smoothed normalized set of attribution weight vectors and the pre-set batch target weights for each domain, includes: Based on the needs of mining operations or expert experience, set a batch target weight for each area that reflects the relative importance or contribution ratio that the area is expected to gain during model training. Based on the smoothed normalized set of attribution weight vectors of all training samples and the pre-set batch target weights of each domain, the total contribution of each training sample to each domain according to its attribution weight and the sum of the total contributions of all training samples are calculated. The sampling probability of each training sample in the current training batch is calculated based on the sum of the total contribution of each training sample to each domain according to its affiliation weight and the total contribution of all training samples.

4. The multi-domain data weighted mixing method for fine-tuning large-scale mine models according to claim 1, characterized in that, The method of generating the current batch of training data by random sampling from the multi-domain, multi-label training dataset of the mine, based on sampling probability, such that the probability of each training sample being selected into the current training batch is proportional to its sampling probability, includes: Based on sampling probability, a predetermined number of samples are drawn from the multi-domain, multi-label training dataset of the mine using random sampling without replacement to form the current batch of training data. In this sampling process, the probability of each training sample being selected is proportional to its sampling probability.

5. The multi-domain data weighted mixing method for fine-tuning large-scale mine models according to claim 1, characterized in that, For each training sample in the current batch of training data, the weighted loss of each training sample is calculated based on its smoothed normalized attribution weight vector and the loss of each domain. The total loss of the current batch is then calculated based on the weighted loss of all training samples, including: Each training sample in the current batch of training data is input into the general large model to be fine-tuned, and forward propagation is performed to obtain the model prediction result of each training sample. For each domain, the loss value of each training sample in that domain is calculated. Based on the loss value of each training sample in each domain and the smoothed normalized assignment weight vector of each training sample, calculate the weighted loss of each training sample. The total loss for this training batch is obtained by summing the weighted losses of each training sample in the current batch of training data.

6. The multi-domain data weighted mixing method for fine-tuning large-scale mine models according to claim 1, characterized in that, The step of fine-tuning the large mining model by calculating gradients through backpropagation and iteratively updating model parameters using an optimizer to minimize the total loss of the current batch includes: With the total loss of the current batch as the optimization objective, the gradient of the loss function with respect to all trainable parameters of the model is calculated through the backpropagation algorithm, and the pre-selected optimizer updates the model parameters based on the calculated gradient in order to reduce the total loss of the current batch. The next round of forward propagation, loss calculation, and parameter updates is performed using new training batches until the preset training rounds or convergence conditions are met, resulting in a specialized large-scale mining model adapted to the needs of the mining industry, which has been fine-tuned by weighted mixing of multi-domain data.

7. A multi-domain data weighted mixing device for fine-tuning large-scale mine models, characterized in that, include: The module for obtaining the smoothed and normalized attribution weight vector is used to calculate the attribution probability of each training sample in multiple domains based on a multi-domain and multi-label training dataset of a mine. This is achieved through manual annotation by domain experts, prediction by domain classifiers, or by combining data statistical features and text similarity. The module generates the original multi-domain attribution weight vector and then performs smoothing and normalization processing to obtain the smoothed and normalized attribution weight vector. The batch training data determination module is used to calculate the sampling probability of each training sample in the current training batch based on the smooth normalized set of attribution weight vectors and the preset batch target weights of each domain. Based on the sampling probability, the current batch training data is generated by random sampling from the multi-domain multi-label training dataset of the mine, so that the probability of each training sample being selected into the current training batch is proportional to its sampling probability. The batch total loss determination module is used to calculate the weighted loss of each training sample in the current batch of training data based on its smoothed normalized home weight vector and the loss of each domain, and to calculate the total loss of the current batch based on the weighted loss of all training samples. The model training module is used to calculate gradients through backpropagation and use the optimizer to iteratively update model parameters to minimize the total loss of the current batch, thus fine-tuning the large mining model.

8. The multi-domain data weighted mixing device for fine-tuning large-scale mine models according to claim 7, characterized in that, The multi-domain, multi-label training dataset based on mining operations calculates the probability of each training sample belonging to multiple domains through methods such as manual annotation by domain experts, prediction by domain classifiers, or a combination of data statistical features and text similarity. This generates the original multi-domain belonging weight vector, which includes: Based on expert knowledge, each training sample is pre-labeled with its domain affiliation probability as its affiliation weight, ensuring that the sum of the N domain affiliation weights for the same training sample is 1, or Using a pre-trained domain classifier, each training sample is classified using multi-label classification. The normalized confidence scores of each class are used as their respective weights. Extract the textual or statistical features of each training sample, calculate its similarity with representative features of each domain, and use the normalized similarity as the attribution weight. Finally, an original multi-domain attribution weight vector is generated for each training sample.

9. A multi-domain data weighted mixing device for fine-tuning large-scale mine models, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the multi-domain data weighted mixing method for fine-tuning a large-scale mine model as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the multi-domain data weighted mixing method for fine-tuning a large-scale mine model as described in any one of claims 1 to 7.

Images