74 results about "Fine-tuning" patented technology

The invention relates to a large model parameter fine tuning method and device for electric power multi-modal data fusion and a medium. The method comprises the steps of obtaining electric power multi-modal data, constructing an electric power multi-modal feature vector and performing time sequence calibration, dividing the electric power multi-modal data into a plurality of sub-data sets according to a geographic position and an electric power equipment type, and distributing the sub-data sets to a plurality of computing nodes of a large model; calculating the local gradient of each calculation node, obtaining a parameter difference coefficient by calculating the gradient similarity between the adjacent calculation nodes, updating the weight coefficient of the calculation nodes, carrying out weighted aggregation on the local gradients of the plurality of calculation nodes according to the updated weight coefficient, obtaining a global gradient, and updating model parameters; and the parameter updating process is repeatedly executed until the model converges, and a large model after parameter fine tuning is obtained and used for outputting a power multi-modal data fusion result. Compared with the prior art, the method has the advantages that the problems of inconsistent power multi-modal data time sequence and unbalanced parameter aggregation are solved, and the model fine tuning efficiency and accuracy are improved.

The invention provides a GIS industry large language model parameter fine tuning method considering parameter adaptability difference, and relates to the field of geographic information science and deep learning, the method comprises the following steps: giving a GIS downstream task training data set and a pre-training large language model, and calculating the adaptability of each parameter of the model to the data set; calculating the adaptability of each level of the large language model according to the calculated adaptability of each parameter; based on the adaptability of each layer, distributing different trainable parameters for the model layers with different adaptability; and based on the distributed trainable parameters, performing fine tuning training on the pre-trained large language model to obtain a GIS industry large language model, and completing efficient fine tuning of the large language model parameters. According to the technical scheme, the heterogeneity of the large language model parameters in the GIS professional knowledge adaptation process is considered, and the characteristic is utilized to optimize the LoRA fine adjustment process so as to improve the performance of the large language model on GIS professional tasks.

The invention provides a rainfall runoff time sequence simulation method based on pre-training-fine tuning, and belongs to the field of hydrological simulation. The method comprises the following steps: firstly, acquiring a consistent hydrometeorological data set for global multi-basin pre-training and target basin fine tuning, constructing an LSTM runoff prediction model, performing global multi-basin pre-training to obtain a base model parameter weight, migrating the base model parameter weight to a target basin, freezing LSTM network layer parameters, and only finely tuning a regression output layer; and testing and evaluating in the target drainage basin. According to the method, firstly, transferable hydrological response is learned on global multi-basin large samples, and then lightweight fine tuning is carried out on the target basin, so that the prediction precision and cross-regional generalization ability of a data scarce region can be effectively improved, and rapid deployment is facilitated.

The invention discloses a large model fine tuning optimization method based on multi-strategy fusion, which comprises the following steps: designing a dynamic parameter selection mechanism, adaptively determining a parameter subset needing fine tuning according to a task demand and a model structure, and reducing unnecessary parameter updating calculation; constructing a dynamic low-rank decomposition framework, dynamically adjusting the rank of a low-rank matrix according to a model training state and data characteristics, and keeping key information while compressing a parameter scale; a self-adaptive task sensing mechanism is introduced, a fine adjustment strategy is automatically adjusted according to different task characteristics, and the adaptability of the model to various tasks is improved; and a mixed precision training method is adopted, so that the calculation complexity and the memory occupation are reduced on the premise of ensuring the model precision. According to the method, a parameter efficient fine tuning technology and a dynamic low-rank decomposition strategy are innovatively combined, and an adaptive task perception mechanism and a mixed precision training technology are introduced, so that the operand and resource requirements of model training are effectively reduced, and the fine tuning efficiency and the model performance are improved.

The invention discloses a hybrid adaptive enhanced fine tuning method, system and device, relates to the technical field of artificial intelligence, and is particularly suitable for a deep reasoning task of a large language model. The problems of response level length deviation, problem difficulty level deviation, insufficient exploration efficiency, insufficient sample utilization and the like existing in an existing reinforcement learning algorithm are solved. The method comprises the steps of data preprocessing, model parameter and reference strategy initialization, multiple response generation, award calculation, advantage calculation and correction, model strategy updating, sampling probability adjustment and iterative optimization. Wherein all potential useful samples are ensured to be fully utilized by introducing the correction advantages of a length normalization factor and a difficulty normalization factor and combining a hybrid cutting mechanism and an adaptive sampling strategy update model. Deviation is effectively eliminated, the model reasoning ability and training efficiency are improved, the long reasoning task exploration ability is enhanced, and diversified task requirements are met.

The invention discloses a large-model intelligent question setting system and method based on fuzzy mathematics fine tuning, and belongs to the technical field of intelligent question setting. Comprising a fuzzification processing module, a multi-dimensional reward feedback module, a rule-driven attribute reasoning module, an attribute accurate quantification module, a self-adaptive membership degree adjustment module, a strategy optimization and intra-cluster standardization module and a question generation and structured output module. According to objective fields such as post specifications, post levels, technical stack depth and team scales, hard indexes are converted into soft boundary semantics through five membership functions, and then quantitative attributes of surface test questions on difficulty, openness, prejudice and distinction are reasoned by using N rules without subjective wording. And the business result data is used for driving rewards to be updated online, and finally, daily automatic fine tuning is realized on the 32B large model, so that each subjective surface test question generated by the AI not only meets the objective requirements of posts, but also has explainable, auditory and iterable closed-loop capabilities.

The embodiment of the invention discloses a forgetting learning method and device for a large language model. The method comprises the steps that firstly, a first large language model and a forgetting sample set are obtained, the first large language model is initially a fine-tuning model obtained by fine-tuning a fine-tuning sample set, and the forgetting sample set is a subset of the fine-tuning sample set; then, for any first forgotten sample, a plurality of similar samples are determined based on a second large language model, and each similar sample and the first forgotten sample have similar semantics but different expressions; processing the first forgotten sample by using a first large language model to obtain a first hidden layer representation; then, training loss is determined, the training loss is negatively correlated to the distance between the first hidden layer representation and the hidden layer representation of each similar sample and is positively correlated to the distance between the first hidden layer representation and the random vector, and the hidden layer representation of each similar sample is obtained based on a fine tuning model; and then training the first large language model by using the training loss to realize forgetting learning.

The invention discloses a small language model fine tuning method based on an LLaMA Factory tool, and the method comprises the following steps: 1) carrying out the thinking chain distillation of an open source data set through a large model, and generating a training data set for the fine tuning of a small language model; 2) based on an LLaMA Factory tool, building a model fine tuning environment; 3) fine-tuning real-time monitoring and adjustment, in the fine-tuning process, monitoring loss of the verification set and the test set, and if the performance of the model is reduced or an over-fitting phenomenon occurs, adjusting training parameters and retraining; and 4) completing training and exporting the model: after the fine tuning process is completed, if the model reaches preset performance on the verification set, completing training, exporting the trained small language model, and deploying the small language model to practical application for reasoning. According to the method, targeted training is performed on the pre-trained small model through a fine tuning technology, specific task requirements can be efficiently met, and rapid migration of a new task can be realized only by using a small-scale task data set to adjust part of parameters of the model.

The application provides a pre-training-fine-tuning-based precipitation runoff time series simulation method, and belongs to the field of hydrological simulation. First, a consistent hydro-meteorological data set for global multi-basin pre-training and target basin fine-tuning is obtained, an LSTM runoff prediction model is constructed, a base model parameter weight is obtained through global multi-basin pre-training, the base model parameter weight is migrated to the target basin, the LSTM network layer parameter is frozen and only the regression output layer is fine-tuned, and test evaluation is carried out in the target basin. The application learns the transferable hydrological response on a large sample of global multi-basins first, and then performs light fine-tuning in the target basin, which can effectively improve the prediction accuracy and cross-region generalization ability in the data scarce area, and is convenient for rapid deployment.

The invention relates to a large model compression method based on continuous layer pruning and endpoint tuning, and the method comprises the steps: firstly introducing a learnable continuous interval soft mask, and building a differentiable hierarchical mask mechanism in a model in cooperation with a residual bypass; secondly, by minimizing the KL divergence between output distributions before and after pruning, the optimal pruning starting point and length are automatically learned, and adaptive selection of continuous layer segments is achieved; then, executing physical layer deletion according to the optimized interval parameters, and reconnecting the network structures before and after pruning; and finally, implementing an endpoint tuning strategy, only carrying out all-parameter fine tuning on key layers on two sides of the sheared interval, and recovering the model performance at the lowest calculation overhead. According to the method, through combination of differential interval search and end point directional optimization, accurate compression and high-performance maintenance of the depth dimension of the large model are realized, model storage occupation and reasoning delay are remarkably reduced, the model output reliability in a key task scene is guaranteed, and the method is suitable for large-scale popularization and application. The method is particularly suitable for efficient deployment of the large language model in a resource-constrained environment.

The application discloses a large language model parameter fine-tuning method and device based on manifold constraint multi-track adaptation and a medium. In view of the problems of unstable training and limited expression capacity of an existing low-rank adaptation technology, a plurality of parallel low-rank tracks are constructed, and a double random matrix is introduced to constrain the information flow between the tracks, so that the stability of gradient propagation is ensured. Meanwhile, the expression capacity of the adaptation module is enhanced under a limited parameter budget by dynamically fusing the outputs of the tracks through a dynamic gating mechanism. The method can realize stable, efficient and high-performance model fine-tuning by fine-tuning a small number of parameters, and reasoning has no additional overhead, and is particularly suitable for application scenarios with limited resources and high stability requirements, and has wide practical value.

The application belongs to the technical field of model fine-tuning, and specifically discloses a model fine-tuning method, system, terminal and medium for parameter update control in a large language model fine-tuning process. A fine-tuning model is established based on a large language model to be fine-tuned, input data and corresponding expected output data are obtained after pre-processing of training data, feature encoding and feature extraction are performed on the input data through a word vector encoding layer and a backbone neural network, intermediate feature results are updated and fused layer by layer using a multi-layer encoding structure, and model output results are obtained. Training loss is calculated based on the deviation between the model output results and the expected output data, the weight parameters in the backbone neural network are evaluated for importance according to the training loss, the weight parameters with a higher contribution degree to the training loss are selected as a target parameter set, and a parameter update operation is performed on the target parameter set. The application can improve the model fine-tuning efficiency and stability, and enhance the adaptation ability of the model in specific task scenarios.

The invention discloses electronic equipment for realizing a large language model parameter fine tuning method and a storage medium, and relates to the field of natural language processing. The fine tuning method comprises the steps that a pre-training model and a parameter matrix of the pre-training model are acquired, a parameter fine tuning module is arranged on a self-attention layer of the pre-training model, the parameter fine tuning module comprises LoRA modules of different task types and a router, each LoRA module comprises a general feature matrix and a specific feature matrix, and the general feature matrixes of the multiple LoRA modules are shared; obtaining a fine tuning data set; loading and freezing a parameter matrix of the pre-training model, and initializing parameters of a parameter fine tuning module; and performing fine tuning on the parameters of the parameter fine tuning module based on the fine tuning data set to obtain a fine-tuned large language model. Parameters of the MoE architecture are remarkably reduced, it is ensured that the model captures differences of various tasks to the maximum extent, and the cross-task generalization of the model is ensured.

The invention discloses a method and device for relieving catastrophic forgetting of a language model, equipment and a storage medium, and relates to the technical field of machine learning. According to the method, linear interpolation is carried out between the pre-training model parameter set and the fine-tuning model parameter set, so that a linear interpolation point of the model can still keep a relatively low loss value when the model learns new task knowledge; then the pre-training model parameter set and the fine-tuning model parameter set are added to form a fusion model parameter set, and the linear interpolation operation only performs one-time simple arithmetical operation after fine-tuning is completed without introducing any additional trainable parameters or changing the model structure, so that when the model adapts to a new task, the model can be quickly and accurately trained. Original general knowledge and performance on old tasks can be reserved to the maximum extent, and meanwhile complexity and expenditure of a model reasoning stage are not increased.

The invention belongs to the technical field of model fine tuning, and discloses a post-training fine tuning method and system for a diffusion model, and the method comprises the steps: carrying out the recursive structure analysis of a pre-obtained diffusion model based on a recursive likelihood ratio optimizer, obtaining recursive parameters, and determining the generation condition of the diffusion model through multi-scale prompt information; according to recursion parameters and generation conditions of the diffusion model, parameters are injected in the recursion process of the diffusion model, and the gradient of the diffusion model is estimated in combination with a gradient estimation method; and according to the estimated gradient of the diffusion model, updating parameters of the diffusion model by utilizing a model parameter updating formula so as to realize post-training adjustment of the diffusion model. The method has the characteristics of lower variance and higher sample efficiency, and effectively reduces the variance of gradient estimation by combining zero-order, half-order and first-order gradient estimation technologies.

The invention discloses a task processing method based on large language model fine tuning, electronic equipment and a medium. The method comprises the following steps: acquiring a training data set related to a downstream task; adding a low-rank adaptation module to all layers; setting a stratified sampling strategy: determining the relative importance of each intermediate layer according to the weight norm distribution of each layer during low-rank adaptation fine tuning of the large language model so as to set the sampling probability of each layer in the training process, enabling each intermediate layer to be uniformly sampled in a single training period through non-return sampling, and obtaining a stratified sampling result; the number of middle layers actually participating in parameter updating is reduced, and video memory occupation in the training process is greatly reduced on the premise that the fine tuning performance is not reduced; according to a stratified sampling strategy, layers needing to be unfrozen in the current training period are dynamically selected in the large language model, parameter updating is carried out on low-rank adaptation modules of the layers, and a training data set is traversed to carry out fine adjustment on the large language model; and the fine-tuned large language model is used for executing downstream tasks.

The application discloses a large model anti-forgetting fine-tuning method and device, computer equipment and medium. The fine-tuning method solves the problem of catastrophic forgetting by introducing an orthogonal penalty loss function. That is, by constructing an orthogonal penalty loss function, adding it to the original loss function to obtain a total loss function, and based on the total loss function, the model training learns in the direction of not forgetting the original knowledge during the model fine-tuning process. When the model learns new knowledge, the Lora fine-tuning technology is prone to cause forgetting of the original knowledge in the incremental pre-training process. At this time, the orthogonal penalty loss function will increase, and through back propagation for adjustment, the orthogonal relationship between the original parameter matrix is maintained, thereby reducing the catastrophic forgetting of the anti-forgetting large model after fine-tuning.

The invention discloses a rotor slip root cause traceability regulation and control method based on a large model and supervised fine tuning. The method comprises the following steps: constructing a slip text knowledge base based on slip parameter division rules, slip field cues and organized pre-declarations; the constructed slip text knowledge base is divided into a training set and a test set, and the structure of each piece of data comprises slip reason analysis, slip regulation and control measures and supplementary description; supervised fine tuning is carried out on the large language model by using the constructed slip text data set in combination with a quantized low-rank adapter technology; and matching the collected rotor working condition parameters with a slip parameter division rule to obtain corresponding rotor working condition parameter grades, and inputting the corresponding rotor working condition parameter grades into the large language model subjected to supervised fine tuning so as to determine a root cause causing slip and formulate regulation and control measures in a targeted manner. According to the invention, a reliable, credible and accurate rotor slip root cause traceability regulation result is obtained.

The invention discloses a multi-task large language model parameter efficient fine tuning method which comprises the following steps: acquiring data sets of different task types, and performing preprocessing and task embedding processing on the data sets; based on the pre-trained large language model, introducing a low-rank adaptive expert to construct a multi-task large language model through an expert mixing mode, a semantic perception routing mechanism and a task adaptive scaling mechanism; and training the constructed multi-task large language model through the data sets of different task types, so as to carry out parameter fine tuning on the low-rank matrix introduced when the low-rank adaptation expert is introduced and parameters in the two mechanisms, thereby obtaining the optimized multi-task large language model. According to the method, the adaptability, generalization performance and reasoning accuracy of the large language model in multi-task learning can be improved while low parameter overhead is kept.

The invention discloses a node fine-tuning method for avoiding bolt finite element contour distortion, aiming at the problem of contour distortion of an existing bolt precision finite element, according to the physical contour of a real bolt, the three-dimensional coordinate of each layer of finite element nodes is automatically fine-tuned, so that the finite element contour is perfectly fitted with the real contour. According to the method, the perfect contour of any thread can be generated under the condition that the contour is not distorted, and great help is provided for the accuracy of finite element calculation of the bolt. Meanwhile, the limitation that the precise finite element theory is only suitable for standard meter threads is broken through, and perfect generation of precise finite element models of special threads such as British Wheatstone threads, rectangular threads, trapezoidal threads, zigzag threads, pipe threads and various non-standard threads is finally achieved.

The application discloses a strong-weak large model cycle fine-tuning training method for fact-rich conversation content generation, and belongs to the technical field of artificial intelligence. The method comprises the following steps: performing one-time cleaning and cold start on an original fact-rich data set by using a strong large model; mixing the cleaned data and the original data according to a probability to form a mixed data set, and performing initial fine-tuning on a weak large model; performing performance evaluation on the fine-tuned model and calculating a performance value; when the performance value exceeds a dynamic threshold, triggering the weak large model to generate a new answer and updating a historical answer queue; performing multi-round training cycle fine-tuning based on the updated data set, and traversing multiple groups of parameter configurations; and finally saving a performance-optimal model. Through the strong-weak model decoupling cooperation and the cycle self-enhancement mechanism, the application effectively reduces the training cost, avoids overfitting, and improves the performance and generalization ability of the model in the fact-rich conversation generation.

The application discloses to the technical field of code completion, specifically a code completion method based on a multimodal two-stage fine-tuning, comprising the following specific steps: S1, constructing a code completion fine-tuning dataset D1: collecting source codes from public channels to construct an original code sample dataset D0; then generating three kinds of modal representation structures corresponding to each source code to construct a code completion fine-tuning dataset D1. The two-stage fine-tuning model proposed in the application combines prefix fine-tuning and adapter fine-tuning, not only reduces resource consumption, but also accurately adjusts the parameters of the model, and the type-code predictor is associated with the direct relationship between the specific code during prediction and the type label, so that it better adapts to the code completion task, and thus the accuracy and efficiency of code completion are improved.

The invention belongs to the technical field of large model parameter fine tuning, and particularly discloses a large model parameter fine tuning method and system based on a low-rank adaptation matrix, a terminal and a medium. Comprising the steps of obtaining an original parameter matrix of a to-be-fine-tuned model and determining an update matrix; low-rank decomposition is carried out on the updated matrix to obtain a low-rank matrix, and the rank value is far smaller than the dimension of an original matrix; in the training process, an original parameter matrix is frozen, only the low-rank matrix B and the low-rank matrix A are trained, and model parameters are updated based on the scaling factor. According to the method, the training parameter scale can be remarkably reduced while the stability of the pre-training model is kept, and the adaptation capability of a large model in a multi-task and resource-constrained environment is improved.

This invention provides a method and system for fine-tuning federated parameters for heterogeneous quantized large language models. The server broadcasts data to each client; each client utilizes metadata to map its local quantization scaling factor to a unified latent space independent of the quantization bit width using a normalization coefficient related to the quantization bit width, reconstructs the global update residual, and performs zero-order gradient estimation based on low-rank subspace perturbation on the quantization scaling factor while freezing low-bit integer weights. A scalar gradient estimate is calculated through forward inference, and a seed-gradient scalar is formed by combining a random seed with the scalar gradient estimate. The server groups and aggregates the seed-gradient scalar pairs according to the quantization bit width and updates the seed sampling probability distribution using a time-aware exponential moving average mechanism. This invention eliminates precision-related biases, reduces client memory usage, and compresses the amount of uploaded data.

The present application relates to a kind of gain modeling method of recommendation system when unobserved confounding exists, belong to information recommendation technical field, solve the problem that existing technology cannot accurately estimate gain when unobserved confounding exists.Method includes: first training set and second training set are constructed;The sample in first training set is the sample obtained from observational study;The sample in second training set is the sample obtained from randomized controlled trial;Each sample includes user article pair feature, the processing scheme of sample and whether the real outcome of user purchases article;Pre-training model is constructed, and pre-training model is used to predict the potential outcome of sample under different processing scheme;Based on first training set, pre-training model is trained to obtain trained pre-training model;Based on trained pre-training model, fine-tuning model is constructed, and based on second training set, fine-tuning model is trained to obtain the gain prediction model eliminating the influence of unobserved confounding. Accurate unbiased gain estimation is realized.

There is provided a method for unlearning to be performed by an unlearning apparatus, the method comprising: receiving to-be-forgotten target data as input; extracting a feature vector from the to-be-forgotten target data using a first layer that is a pre-attack layer of a pre-trained model; setting a second layer, which is a target attack layer of the pre-trained model, as a first model, and duplicating a second layer to generate a duplicated second layer to be set as a second model; generating a noisy feature vector by adding noise to the feature vector; fine-tuning one of the first model and the second model by applying a knowledge distillation technique to the first model and the second model using the feature vector and the noisy feature vector; and generating an unlearned model for the to-be-forgotten target data based on the first layer and the fine-tuned model.

A method comprising modifying an integral parameter with an anti-windup correction; modifying a derivative parameter with a bandlimited derivative; and modulating a control signal in accordance with a global optimum that corresponds to one or more operating conditions of a plant system.

A VLA model fine-tuning method based on structured stage and key frame supervision, containing five steps of basic model initialization, automatic label extraction, auxiliary architecture construction, joint training and inference execution, which can overcome the defects of structured operation supervision, long-range gain and low key frame prediction error. Zero artificial annotation, general adaptation automatically extracts stage / key frame label from demonstration gripper state, greatly improves the success rate of operation; long-term multi-key frame task gain is greater, no migration cost, no modification of basic VLA model architecture and inference process, plug and play, light and efficient, no performance loss auxiliary head is light MLP, the parameter amount of learning query token is extremely small, the training / inference speed is completely consistent with the basic model, the representation is accurate, the stage representation of long-term stable learning faithfully tracks the operation stage, the key frame prediction error is as low as 10 ‑4 -10 ‑5 orders of magnitude, and there is no error divergence in complex tasks.