LoRA fine tuning-based achievement transformation large model task decomposition method and system

Abstract

The invention belongs to the technical field of artificial intelligence task decomposition, and particularly relates to an achievement transformation large model task decomposition method and system based on LoRA fine tuning. According to the method, a LoRA (Low-Rank Adaptation) technology is adopted to carry out efficient parameter fine tuning on a large achievement transformation model, and four core parts including achievement transformation task data set construction, task step decomposition, relation prediction and joint optimization training are included. The achievement conversion task data set is constructed by collecting multi-step task samples of typical scenes such as technology evaluation, market analysis and patent layout, and labeling execution dependency relationships among steps. And fine tuning is performed on the result transformation large model based on the LoRA technology, two functions of task step decomposition and inter-step relationship prediction are realized at the same time, and step decomposition loss and relationship prediction loss are calculated respectively. Through joint optimization of the two loss functions, the model after fine tuning can output a task step sequence with a reasonable front-back execution relation. Through the LoRA parameter efficient fine tuning technology, the achievement transformation large model can quickly adapt to the new task under the condition of a small amount of data, the training time and the consumption of computing resources are reduced, meanwhile, the accuracy and the flexibility of task decomposition are improved, and the method is suitable for various achievement transformation scenes.

Description

Background Art

[0002] With the rapid development of artificial intelligence technology, large language models have shown great potential in the field of scientific and technological achievement transformation, especially in task decomposition. The process of scientific and technological achievement transformation involves multiple links such as technology assessment, market analysis, patent layout, production planning, etc. Each link contains multiple execution steps, and there are complex dependencies between these steps. By reasonable task decomposition, the efficiency of achievement transformation can be improved, the implementation difficulty can be reduced, and the transformation success rate can be enhanced. Therefore, how to improve the task decomposition ability of large models in the scenario of achievement transformation has become a key issue.

[0003] Currently, there are mainly four types of technical routes for task decomposition in achievement transformation: 1) Rule-based methods, which decompose achievement transformation tasks through predefined templates and rules; 2) Search data-based methods, which mine task steps using users' query behaviors in the process of achievement transformation; 3) Crowdsourcing-based methods, which construct a task decomposition dataset for achievement transformation through manual annotation; 4) Large language model-based methods, which utilize the task decomposition knowledge contained in pre-trained models. Among them, the large language model-based method has received extensive attention due to its powerful knowledge acquisition and reasoning capabilities, and has shown unique advantages in this complex scenario of scientific and technological achievement transformation.

[0004] However, the inventors found that the existing large language model-based achievement transformation task decomposition methods have the following problems: 1) Traditional fine-tuning methods require updating all model parameters, resulting in huge computational resource consumption when adapting to achievement transformation tasks; 2) When fine-tuning all parameters, the parameter space is too large, and the training efficiency is low for different types of achievement transformation tasks; 3) There is a lack of targeted parameter adjustment strategies, making it difficult to accurately grasp the temporal dependencies between subtasks in the process of achievement transformation; 4) The cost of model fine-tuning is high, and it is difficult to quickly adapt to the task decomposition requirements of different technical fields. Especially in the achievement transformation scenario involving multiple technical field intersections, these problems are more prominent. Therefore, there is an urgent need for a fine-tuning method with high parameter efficiency, fast training, and the ability to accurately identify task temporal relationships to improve the application effect of large models in achievement transformation task decomposition. Summary of the Invention

[0005] To solve at least one of the above technical problems in the background art, the present invention provides a method and system for task decomposition of an achievement transformation large model based on LoRA fine-tuning. Through the parameter-efficient low-rank adaptation technology, it realizes the characteristics of low computational resource consumption, high training efficiency, accurate prediction of temporal relationships, and strong domain adaptability, and is more in line with actual application scenarios such as scientific and technological achievement transformation.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] Construction of the achievement transformation task dataset. The present invention first constructs a high-quality achievement transformation task decomposition dataset, and the construction process of this dataset includes:

[0008] (1) Collect task samples, and collect representative multi-step task descriptions from typical achievement transformation scenarios such as technology assessment, market analysis, patent layout, and production planning as training samples;

[0009] (2) Step decomposition annotation. Domain experts perform detailed step decomposition on each achievement transformation task. The decomposition process needs to follow the following principles: a. Each step should be specific and executable, such as patent retrieval, technology route assessment, etc.; b. The step granularity should be moderate, neither too general nor too fragmented; c. The step description should be standardized and professional, using unified technical terms; d. Ensure the integrity of the step set, covering the entire process of achievement transformation.

[0010] (3) Temporal relationship annotation. Perform dependency annotation on the decomposed step set. The dependency relationships are divided into the following categories: a. Strict sequential relationship: For example, technology assessment must be completed before business planning; b. Parallel relationship: For example, market research and patent retrieval can be carried out simultaneously; c. Optional sequential relationship: The order of some preparatory work can be adjusted flexibly. Through these relationships, a complete achievement transformation execution process is constructed.

[0011] Task step decomposition module. Using the achievement transformation large model as the base, and performing efficient parameter fine-tuning through LoRA, which is used to decompose tasks into specific executable step sequences. The training configuration includes:

[0012] (1) Structure of the pre-trained model: The pre-trained large language model structure Llama-2-7B adopted in the present invention is used as the achievement transformation large model, and it can also be any other structural model.

[0013] (2) Fine-tuning configuration: Perform LoRA fine-tuning on the Query and Value matrices of the Attention layer, with the rank value set to 8 and the alpha value set to 16.

[0014] (3) Training strategy: Use the AdamW optimizer, with a learning rate of 3e-4, and dynamically adjust the learning rate using the validation set.

[0015] (4) Optimization objective: Minimize the cross-entropy loss between the predicted step sequence and the true step sequence.

[0016] Task step decomposition module. Using the achievement transformation large model as the base, and performing efficient parameter fine-tuning through LoRA, which is used to decompose tasks into specific executable step sequences. The training configuration includes: Input form:

[0017] (1) Structure of the pre-trained model: The same pre-trained large language model in the sampling and task step decomposition module.

[0018] (2) Input form: Take the task step pair as the input and predict the dependency type between them.

[0019] (3) Fine-tuning strategy: Use the same LoRA configuration as the step decomposer, but specifically optimize the loss function.

[0020] (4) Optimization objective: Minimize the binary cross-entropy loss between the predicted dependency relationship and the true dependency relationship.

[0021] Joint optimization training. Jointly optimize the loss functions of step decomposition and relationship prediction, and design a unified optimization objective through weighted combination. Use the same set of LoRA parameters to implement the end-to-end model training and inference process, evaluate the overall performance through the validation set, and dynamically adjust the parameters.

[0022] The second aspect of the present invention provides a task decomposition system for the achievement transformation large model based on LoRA fine-tuning, including:

[0023] An achievement transformation task dataset construction module for constructing a task decomposition training dataset for the science and technology achievement transformation scenario; it includes multi-step task descriptions such as technology evaluation, market analysis, patent layout, etc., as well as corresponding step sequences and execution dependency relationship annotations; a task step decomposition module for decomposing the transformation task into specific executable steps through the achievement transformation large model with LoRA fine-tuning and calculating the step decomposition loss; a relationship prediction module for predicting the execution dependency relationship between steps based on the same achievement transformation large model with LoRA fine-tuning and calculating the relationship prediction loss; a joint optimization module for uniformly calculating the weighted combination of the step decomposition loss and the relationship prediction loss, and efficiently fine-tuning the achievement transformation large model through a single set of low-rank parameters to improve the task decomposition ability of the model in different technical fields

[0024] The third aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the steps in the task decomposition method of the achievement transformation large model based on LoRA fine-tuning as described above.

[0025] The fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the task decomposition method of the achievement transformation large model based on LoRA fine-tuning as described above.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] (1) The present invention uses LoRA technology to perform parameter-efficient fine-tuning on the achievement transformation large model. Through low-rank matrix decomposition and adaptive updates, it significantly reduces the consumption of computing resources and improves the training efficiency.

[0028] (2) The present invention designs the task decomposition as a joint task of step generation and relationship prediction, and improves the accuracy of decomposition and the prediction quality of temporal relationships through unified LoRA parameter optimization.

[0029] (3) The present invention has good domain adaptability, can handle the task decomposition requirements of different fields including scientific and technological achievement transformation, and does not require retraining the entire model for a new field.

[0030] (4) The method of the present invention is easy to deploy and expand, can be quickly applied to practical scenarios such as scientific and technological achievement transformation and project management, and provides strong support for the collaborative innovation of industry, university and research.

[0031] Advantages of additional aspects of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0033] Figure 1 is a flowchart of the task decomposition method of the achievement transformation large model based on LoRA fine-tuning in an embodiment of the present invention;

[0034] Figure 2 is a flowchart of the construction of the achievement transformation task dataset in an embodiment of the present invention;

[0035] Figure 3 is a flowchart of the training of the task step decomposer in an embodiment of the present invention;

[0036] Figure 4 is a flowchart of the training steps of the task step relationship predictor in an embodiment of the present invention;

[0037] Figure 5 is a flowchart of the joint optimization training based on low-rank parameters in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] The present invention will be further described below in conjunction with the drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0040] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0041] Reference Figure 1 This embodiment provides a method for decomposing a large model task for achievement transformation based on LoRA fine-tuning, which includes:

[0042] S101: Construction of a dataset for achievement transformation tasks. Collect samples of transformation tasks in typical scenarios such as technology evaluation, market analysis, patent layout, and production planning. Domain experts will decompose the steps and annotate the execution dependencies to build a training dataset.

[0043] S102: Task step decomposition training, based on the large model of achievement transformation, fine-tuning LoRA by designing a low-rank parameter matrix, decomposing complex transformation tasks into specific executable step sequences, and calculating the step decomposition loss.

[0044] S103: Step relationship prediction training, based on the same achievement transformation model and shared LoRA parameters, predicts the execution dependency between step pairs and calculates the relationship prediction loss.

[0045] S104: Joint optimization training, weighted combination of step decomposition loss and relationship prediction loss, optimizing the objective function through unified LoRA parameters, and outputting a task step sequence with a reasonable execution order.

[0046] Reference Figure 2 , the flow chart of the steps for constructing the dataset for the achievement transformation task, and the detailed description of the dataset construction process:

[0047] Step 200: Collect samples of achievement transformation tasks. Collect typical task descriptions from the field of scientific and technological achievement transformation to ensure the diversity and representativeness of the samples. Step 200 can be broken down into the following:

[0048] Step 200-1: Determine the scope of task collection, including typical scenarios such as technology evaluation, market analysis, patent layout, and production planning;

[0049] Step 200-2: Develop task complexity criteria to ensure that the tasks include multiple key transformation links;

[0050] Step 200-3: Set the target number of samples, N = 1000, to ensure that the scale of the dataset meets the training requirements;

[0051] Step 200-4: Normalize the task description to unify technical terms and format standards.

[0052] Step 201: Decompose and annotate the transformation steps. Domain experts decompose and annotate each achievement transformation task. Step 201 can be refined as follows:

[0053] Step 201-1: Develop professional step decomposition criteria to ensure that the steps are specific and executable;

[0054] Step 201-2: Split each transformation task into links;

[0055] Step 201-3: Ensure the professionalism and integrity of the step description;

[0056] Step 201-4: Conduct expert review and correction of the decomposition results;

[0057] Step 201-5: Annotate the technical associations between steps.

[0058] Step 202: Perform dependency annotation. Determine the execution dependencies between transformation steps. Step 202 can be refined as follows:

[0059] Step 202-1: Define the types of dependencies (e.g., technical evaluation must precede business planning, etc.);

[0060] Step 202-2: Annotate the execution order between step pairs;

[0061] Step 202-3: Verify the feasibility of the dependencies;

[0062] Step 202-4: Establish a complete step execution sequence.

[0063] Refer to Figure 3 , the training flow chart of the task step decomposer, to elaborate on the decomposer training process in detail:

[0064] Step 300: Load the task step decomposition training dataset D_train, set a random seed to randomly sample from the training set, and train in batches;

[0065] Step 301: Configure the base model. Select the achievement transformation large model as the base and configure the LoRA parameters. Step 301 can be refined as follows:

[0066] Step 301-1: Select the base model M as Llama-2-7B, with the input dimension d = 1024;

[0067] Step 301-2: Sample the random seed set in Step 300, initialize the parameters of the LoRA weight matrix W ∈ Rd×r according to the random seed, and freeze the remaining parameters of the pre-trained large language model, keeping them unchanged;

[0068] Step 301-3: Set the LoRA parameters (r = 8, α = 16);

[0069] Step 301-4: Determine the fine-tuning layer set L = {Attn_Q, Attn_V};

[0070] Step 302: Training parameter configuration. Set the hyperparameters and optimization strategies related to training. The said Step 302 can be refined as follows:

[0071] Step 302-1: Configure the AdamW optimizer, with the learning rate η = 3e-4;

[0072] Step 302-2: Set the cosine annealing scheduler, with the period T = 3;

[0073] Step 302-3: Configure the training batch size B = 8;

[0074] Step 302-4: Set the gradient accumulation step k = 4.

[0075] Step 303: Calculate the cross-entropy loss L(θ) of the step decomposition.

[0076] Refer to Figure 4 , the training flow chart of the task step relationship predictor, for a detailed description of the training process of the relationship predictor:

[0077] Step 400: Load the task relationship prediction training set D_train corresponding to the task step decomposition training data set, randomly sample from the training set using the random seed in Step 300, and train in batches;

[0078] Step 401: Prediction module configuration. Use the same result transformation large model and LoRA parameters as in the step decomposition, and configure the step for the input format <s1, s2>. The said Step 401 can be refined as follows:

[0079] Step 401-2: Share the LoRA weight matrix W ∈ Rd×r of the step decomposition module, and keep the pre-trained model parameters frozen;

[0080] Step 401-3: Define the relationship type set R = {seq, par, opt};

[0081] Step 401-4: Reuse LoRA parameter configuration (r = 8, α = 16);

[0082] Step 401-5: Share the fine-tuning layer set L = {Attn_Q, Attn_V};

[0083] Step 402: Training parameter configuration. Reuse the training strategy of the step decomposition module. Step 402 can be refined as follows:

[0084] Step 402-1: Reuse the AdamW optimizer with a learning rate η = 3e-4;

[0085] Step 402-2: Reuse the cosine annealing scheduler with a period T = 3;

[0086] Step 402-3: Keep the training batch size B = 8;

[0087] Step 402-4: Keep the gradient accumulation step k = 4;

[0088] Step 403: Calculate the multi-class cross-entropy loss L_rel(θ) for relationship prediction;

[0089] Refer to Figure 5 for a detailed description of the joint optimization training process, which includes:

[0090] Step 500: Load the training dataset D_train, randomly sample from the training set using the previous random seed, and train in batches;

[0091] Step 501: Joint optimization configuration. Based on the same set of LoRA parameters, jointly optimize the loss functions of step decomposition and relationship prediction. Step 501 can be refined as follows:

[0092] Step 501-1: Define the joint loss function L_total = λ 1 *L(θ) + λ 2 *L_rel(θ), where λ 1 and λ 2 are weight coefficients, set to 0.6 and 0.4 respectively;

[0093] Step 501-1: Reuse the previous large model configuration for result conversion;

[0094] Step 501-2: Reuse the previous LoRA parameters and training configuration.

[0095] Step 502: Joint training and optimization. Perform joint optimization training based on a single LoRA parameter. Step 502 can be refined as follows:

[0096] Step 502-1: Calculate the weighted combined joint loss L_total;

[0097] Step 502-2: Update the shared LoRA parameters through back propagation;

[0098] Step 502-3: Save model checkpoints regularly to prevent training interruptions;

[0099] Step 502-4: Verify the overall performance of the model on the two tasks of step decomposition and relationship prediction.

[0100] Step 503: Determine whether the number of training iterations reaches a threshold or the joint loss tends to be stable. If so, the optimization training ends. Otherwise, continue to optimize the LoRA parameters;

[0101] Embodiment 2

[0102] This embodiment provides a large model task decomposition system for achievement transformation based on LoRA fine-tuning, which includes:

[0103] The module for building a dataset for the task of transforming achievements is used to collect samples of transformation tasks in typical scenarios such as technology evaluation, market analysis, patent layout, and production planning. Domain experts will decompose the steps and annotate the execution dependencies to build a standardized training dataset.

[0104] Task step decomposition module: Based on the large model of achievement transformation, the LoRA technology is used to efficiently fine-tune parameters and decompose the task into a sequence of specific steps. The optimization goal of step decomposition is to minimize the cross entropy loss L_task(θ).

[0105] Relationship prediction module: Based on the same large model of achievement conversion and LoRA parameters, it predicts the temporal dependencies between steps; the optimization goal of relationship prediction is to minimize the cross entropy loss L_rel(θ) of relationship classification.

[0106] Joint optimization training module: Through a single set of LoRA parameters, the step decomposition and relationship prediction are jointly optimized to output task steps with a reasonable execution order; among which, the goal of the joint optimization is to minimize the weighted loss function L_total=λ1*L_task(θ)+λ2*L_rel(θ).

[0107] Embodiment 3

[0108] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps in the task decomposition method of the large model of achievement transformation based on LoRA fine-tuning as described above are implemented.

[0109] Embodiment 4

[0110] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the method for task decomposition of the achievement transformation large model based on LoRA fine-tuning as described above.

[0111] Those skilled in the art of technology should recognize that the embodiments of the present invention can take the form of a method, a system, or a computer program product. Therefore, the present invention can be implemented by hardware, software, or a combination of both. In addition, the present invention can take the form of a computer program product on a computer-usable storage medium (including but not limited to disk memories and optical memories, etc.), which contains one or more computer-usable program codes.

[0112] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A large-scale model task decomposition technology for achievement transformation based on LoRA fine-tuning, characterized in that: It includes data set construction, task step decomposition, relationship prediction and joint optimization training. Among them, data set construction is used to collect samples of achievement transformation tasks and perform step decomposition and execution dependency annotation. Task step decomposition is used to efficiently fine-tune the parameters of the achievement transformation large model through LoRA technology, and decompose the transformation task into a specific step sequence. Relationship prediction is used to predict the execution dependency between steps based on shared LoRA parameters. Joint optimization training is used to optimize the same set of low-rank parameters and output a transformation step sequence with a reasonable execution order.

2. A process of constructing a data set based on the task decomposition technology of the large model of LoRA fine-tuning achievement transformation, characterized in that: Task samples are collected from typical achievement transformation scenarios such as technology evaluation, market analysis, patent layout, and production planning, and domain experts perform step decomposition and execution dependency annotation. The dataset contains task descriptions, step sequences, and their execution relationships. Decomposition standards and dependency types are formulated based on the characteristics of achievement transformation.

3. A process of decomposing modules of a large model task decomposition technology training step based on LoRA fine-tuning, characterized in that: The training batch is constructed by random sampling. The step decomposition module is based on the large model of achievement transformation and uses a low-rank parameter matrix for efficient fine-tuning. The input is the transformation task description and the output is a specific execution step sequence. According to the characteristics of the achievement transformation task, configure parameters such as LoRA's rank size and learning rate.

4. A process for training a relationship prediction module based on the task decomposition technology of a large model for the transformation of LoRA fine-tuning, characterized in that: The training batch is constructed by random sampling. The relation prediction module is based on the same large model and LoRA parameters, with the input as step pairs and the output as execution dependencies. The low-rank parameter configuration of the shared step decomposition module ensures the consistency of the prediction results.

5. A process of joint optimization training of large model task decomposition technology based on LoRA fine-tuning, characterized in that: The step decomposition loss L(θ) and the relationship prediction loss L_rel(θ) are jointly optimized, and the joint loss function is L_total=λ1*L(θ)+λ2*L_rel(θ), where λ1 and λ2 are weight coefficients. Set optimization strategies and number of iterations based on the characteristics of the results transformation tasks.

6. A process of applying the large model task decomposition technology for achievement transformation based on LoRA fine-tuning, characterized in that: In practical applications, the model receives description information of the achievement transformation task, and simultaneously completes step decomposition and dependency prediction through shared low-rank parameters, providing execution guidance for the transformation process.

7. A process of decomposing modules of a large model task decomposition technology training step based on LoRA fine-tuning, characterized in that: According to claim 3, the training step decomposes the module and dynamically adjusts the learning rate and optimization parameters until the model reaches the expected performance.

8. A process for training a relationship prediction module based on the task decomposition technology of a large model for the transformation of LoRA fine-tuning, characterized in that: According to claim 4, the relationship prediction module is trained, and the same optimization strategy and judgment criteria are reused to ensure the synergy with the step decomposition module.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the task decomposition method of a large model for achievement transformation based on LoRA fine-tuning are implemented as described in any one of claims 1 to 8.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps in the task decomposition method of a large model for achievement transformation based on LoRA fine-tuning as described in any one of claims 1-8.

Images