Information generation model training method and information generation method

Abstract

This specification discloses an information generation model training method and an information generation method. The information generation model training method includes: obtaining sample data for a basic generation model, and constructing training instructions based on the sample data, and inputting the sample data and the training instructions into the basic generation model to obtain test results, and adjusting the training instructions according to the test results to obtain target training instructions, and compressing the model parameters of the basic generation model to obtain the information generation model to be trained; inputting the target training instructions and the sample data into the information generation model to determine the prediction results; and training the information generation model according to the prediction results. This solution can train an information generation model with high performance based on a small number of training samples, significantly reducing the demand for the number of samples for model performance, and effectively improving the accuracy of the model.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an information generation model training method and an information generation method. Background Art

[0002] With the rapid development of materials science and computing technology, the generation of material structural cells (i.e., the basic structural unit of materials) is of great significance in the fields of material design and performance optimization. Traditional material design methods often rely on a large number of experiments and experience, with long cycles and high costs. With the development of artificial intelligence technology in recent years, the application of large models has provided new ideas for material design.

[0003] However, existing models applied in the field of materials often require full training through a large amount of data resources, but the data in the field of materials science is relatively limited, and it is difficult for the model to achieve ideal performance based on limited data.

[0004] Therefore, how to ensure the reasoning performance of the model with limited sample data is an urgent problem to be solved. Summary of the invention

[0005] This specification provides an information generation model training method and an information generation method to partially solve the above-mentioned problems existing in the prior art.

[0006] This manual adopts the following technical solutions:

[0007] This specification provides an information generation model training method, including:

[0008] Acquire sample data for a basic generation model, and construct training instructions based on the sample data, wherein the sample data includes material structure data with different attribute information, and the training instructions are used to guide the basic generation model to generate material structure data with specified attributes, or guide the basic generation model to generate attribute information of specified material structure data;

[0009] Inputting the training instruction and the sample data into the basic generation model to obtain a first prediction result;

[0010] Determine a test result according to the first prediction result and the actual label corresponding to the sample data, adjust the training instruction according to the test result to obtain a target training instruction, and compress the model parameters of the basic generation model to obtain an information generation model to be trained;

[0011] Inputting the target training instruction and the sample data into the information generation model to determine a second prediction result through the information generation model;

[0012] According to the actual label corresponding to the second prediction result and the sample data, the loss value of the information generation model is determined, and the information generation model is trained according to the loss value.

[0013] Optionally, adjusting the training instruction according to the test result to obtain a target training instruction specifically includes:

[0014] According to the deviation between the first prediction result and the actual label corresponding to the sample data, determining information in the prediction result that does not meet the preset condition as adjustment reference information;

[0015] The training instruction is adjusted according to the adjustment reference information to obtain the target training instruction.

[0016] Optionally, compressing the model parameters of the basic generation model to obtain the information generation model to be trained specifically includes:

[0017] The weight matrix of the basic generation model is subjected to low-rank decomposition, and the basic generation model is adjusted based on the weight matrix obtained after the low-rank decomposition to obtain the information generation model.

[0018] Optionally, determining a loss value of the information generation model according to the prediction result and an actual label corresponding to the sample data, and training the information generation model according to the loss value specifically includes:

[0019] Determining a loss value of the information generation model according to the prediction result and the actual label corresponding to the sample data, and pre-training the information generation model according to the loss value to obtain a pre-trained information generation model;

[0020] The target training instruction of the pre-trained information generation model is adjusted, and the pre-trained information generation model is trained using the adjusted target training instruction.

[0021] Optionally, before inputting the target training instruction and the sample data into the information generation model, the method further includes:

[0022] Determining whether a preset information access condition is met according to the complexity of the target training instruction;

[0023] If yes, then obtaining preset reasoning guidance data, wherein the reasoning guidance data includes step description information corresponding to a plurality of reasoning steps, and is used to guide the reasoning steps performed by the information generation model in the process of determining the prediction result;

[0024] Inputting the target training instruction and the sample data into the information generation model to determine a second prediction result through the information generation model specifically includes:

[0025] The target training instruction, the sample data and the reasoning guidance data are input into the information generation model, so that each reasoning step is executed through the information generation model under the guidance of the description information of each step contained in the reasoning guidance data to obtain the prediction result.

[0026] Optionally, before inputting the target training instruction and the sample data into the information generation model, the method further includes:

[0027] Determine whether the preset information access conditions are met according to the complexity of the target training instruction;

[0028] If yes, then determining the attribute information corresponding to the material structure data generated by the information generation model according to the target training instruction, and searching for information matching the attribute information in a preset knowledge base as description enhancement information;

[0029] The target training instruction is updated using the description enhancement information.

[0030] This specification provides an information generation method, including:

[0031] Receiving an information generation instruction, wherein the information generation instruction is used to instruct a pre-trained information generation model to generate material structure data with specified attribute information, or to generate attribute information according to specified material structure data, and the information generation model is trained by the above method;

[0032] The generation instruction is input into the information generation model to generate material information through the information generation model.

[0033] This specification provides an information generation model training device, including:

[0034] A receiving module, used to obtain sample data for a basic generation model, and construct training instructions based on the sample data, wherein the sample data includes material structure data with different attribute information, and the training instructions are used to guide the basic generation model to generate material structure data with specified attributes, or guide the basic generation model to generate attribute information of specified material structure data;

[0035] A first prediction module, used for inputting the training instruction and the sample data into the basic generation model to obtain a first prediction result;

[0036] An adjustment module, configured to determine a test result according to the first prediction result and the actual label corresponding to the sample data, and to adjust the training instruction according to the test result to obtain a target training instruction, and to compress the model parameters of the basic generation model to obtain an information generation model to be trained;

[0037] A second prediction module, used for inputting the target training instruction and the sample data into the information generation model to determine a second prediction result through the information generation model;

[0038] A training module is used to determine the loss value of the information generation model according to the actual label corresponding to the second prediction result and the sample data, and train the information generation model according to the loss value.

[0039] This specification provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned information generation model training method and information generation method are implemented.

[0040] This specification provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned information generation model training method and information generation method when executing the program.

[0041] At least one of the above technical solutions adopted in this specification can achieve the following beneficial effects:

[0042] In the information generation model training method provided in the present specification, sample data for a basic generation model is obtained, and training instructions are constructed based on the sample data, and the sample data and the training instructions are input into the basic generation model to obtain test results, and the training instructions are adjusted according to the test results to obtain target training instructions, and the model parameters of the basic generation model are compressed to obtain the information generation model to be trained; the target training instructions and the sample data are input into the information generation model to determine the prediction results; and the information generation model is trained according to the prediction results.

[0043] It can be seen from the above method that this scheme can adjust the training instructions by testing the basic generation model, and then train the compressed information generation model through the target training instructions obtained after the adjustment. On the one hand, this scheme can effectively reduce the model parameters that need to be learned by compressing the basic generation model. On the other hand, based on the adjustment of the training instructions, the model's understanding ability and performance in generating content are improved, which significantly reduces the sample data required to achieve the ideal model performance, effectively guarantees the performance of the information generation model under limited sample data, and further improves the training efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings described herein are used to provide a further understanding of this specification and constitute a part of this specification. The illustrative embodiments and descriptions of this specification are used to explain this specification and do not constitute an improper limitation on this specification. In the drawings:

[0045] Figure 1 A flowchart of an information generation model training method provided in this specification;

[0046] Figure 2 A schematic diagram of the process of introducing external information provided in this manual;

[0047] Figure 3 A flowchart of an information generation method provided in this specification;

[0048] Figure 4 A schematic diagram of an information generation model training device provided in this specification;

[0049] Figure 5 A method corresponding to the Figure 1 or Figure 3 Schematic diagram of electronic equipment. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of this specification more clear, the technical solutions of this specification will be clearly and completely described below in combination with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this specification.

[0051] The technical solutions provided by the embodiments of this specification are described in detail below in conjunction with the accompanying drawings.

[0052] Figure 1 A flowchart of an information generation model training method provided in this specification includes the following steps:

[0053] S101: Obtain sample data for a basic generation model, and construct training instructions based on the sample data, wherein the sample data includes material structure data with different attribute information, and the training instructions are used to guide the basic generation model to generate material structure data with specified attributes, or guide the basic generation model to generate attribute information of specified material structure data.

[0054] In this specification, the execution subject for implementing an information generation model training method and an information generation method can be a designated device such as a server. For the sake of ease of description, the following will only take the server as an execution subject as an example to illustrate a model training method and an information generation method provided in this specification.

[0055] Among them, the server can obtain sample data for the basic generation model, the basic generation model can be a large model of the basic Transformer architecture, the sample data can include material structure data with different attribute information (such as elastic modulus, volume fraction, mechanical properties, thermal properties, etc.), and the material structure data can be the cell structure (i.e., the basic structural unit of the material) data of the material.

[0056] In practical applications, the above-mentioned training instructions may include forward instructions and reverse instructions, wherein the forward instructions are used to guide the basic generation model to generate material structure data with specified properties, that is, to generate the cell structure of the material according to the given material properties, and the reverse instructions are used to guide the basic generation model to generate attribute information of the specified material structure data, that is, to determine its material properties according to the given cell structure.

[0057] Specifically, the server may obtain a preset instruction template, and the format of the attribute template may be: "Generate a cell structure of [material attribute description]" or "Predict [material attribute description] based on the cell structure".

[0058] Afterwards, the server can provide specific training instructions corresponding to each sample data according to the above instruction objectives and the specific content of the sample data, such as forward instructions: "Generate a cell structure with an elastic modulus of 90 GPa", "Generate a cell structure with a volume fraction of 0.6", "Design a cell structure with a Poisson's ratio of 0.3"; reverse instructions: "Predict its elastic modulus based on a given cell structure", "Predict its density based on a given cell structure".

[0059] In order to facilitate large model processing, the server can encode the instructions in a unified format. For forward instructions, the encoding format can be:

[0060] {"task_type" "generation",

[0061] "material_property": {

[0062] "property_name": "elastic_modulus",

[0063] "value": "90 GPa"}}

[0064] For reverse instructions, the encoding format can be:

[0065] { "task_type": "prediction",

[0066] "material_property": {

[0067] "property_name": "thermal_conductivity"}}

[0068] S102: Inputting the training instruction and the sample data into the basic generation model to obtain a first prediction result;

[0069] S103: Determine a test result based on the first prediction result and the actual label corresponding to the sample data, adjust the training instruction based on the test result to obtain a target training instruction, and compress the model parameters of the basic generation model to obtain the information generation model to be trained.

[0070] To evaluate its initial performance on the generation task. In this way, the capabilities and potential defects of the basic model in handling actual tasks are determined. Small representative sample data are selected from the material property data set to ensure that these samples cover different material property ranges and complexities. Then, training instructions are constructed for model testing. The generated training instructions and test instructions are input into the basic generation model, and the material structure data or property information output by the model is observed as the first prediction result. The first prediction result is compared with the actual label of the sample data to obtain the test result, so as to evaluate the accuracy, generalization ability and reliability of the model.

[0071] For the forward generation task, the input training instruction is the forward instruction, the input data is the attribute information, and the output is the material structure data (material cell structure). Correspondingly, the label of the sample data is the actual material structure data corresponding to the above attribute information; for the reverse generation task, the input training instruction is the reverse instruction, the input data is the material structure data, and the output is the attribute information. Correspondingly, the label of the sample data is the actual attribute information corresponding to the above material structure data.

[0072] Among them, the evaluation indicators of the test include: accuracy: the consistency between the model prediction results and the actual values; generalization ability: the ability of the model to process unseen data samples; reliability: the consistency and stability of the model output under different conditions. After preliminary analysis, analyze the performance of the model in small sample tests and identify the difficulty of the task. For example, the model performs well in predicting certain specific attributes such as volume fraction, but performs poorly in generating cell structures with specific attributes such as Young's modulus. According to the results of the preliminary evaluation, adjust the training instructions.

[0073] The server can determine the information in the prediction result that does not meet the preset conditions based on the deviation between the first prediction result and the actual label corresponding to the sample data, and use it as adjustment reference information; then adjust the training instructions based on the adjustment reference information to obtain the target training instructions.

[0074] The server may use the information in the prediction result that does not match the input data as adjustment reference information. Of course, the server may also use the information that is not ideal or does not meet the business needs as adjustment reference information according to the actual situation.

[0075] In actual applications, the server can collect feedback information based on the test results of the basic generation model for more difficult tasks, optimize the training instructions, and then re-input the large commercial model, repeating the above process until the output is basically accurate or exceeds three cycles.

[0076] Specifically, the server can analyze the test results, identify the deficiencies of the model when processing specific instructions, and then adjust the instruction content based on the adjustment reference information, such as adding contextual information to improve the model's understanding ability. Based on the feedback information, identify the deficiencies of the model when processing specific instructions, such as insufficient description accuracy, insufficient contextual understanding, etc. Adjusting the instruction content includes: adding contextual information: adding more background information to the instruction to help the model better understand the task; modifying the property description: ensuring that the description of the material properties is more accurate and consistent; expanding the scope of the instruction: including more types of material properties and structural requirements.

[0077] At the same time, the server can compress the model parameters of the basic generation model to obtain the information generation model to be trained.

[0078] Among them, the server can perform low-rank decomposition on the weight matrix of the basic generation model, and adjust the basic generation model based on the low-rank weight matrix obtained after the low-rank decomposition to obtain the information generation model.

[0079] S104: Inputting the target training instruction and the sample data into the information generation model to determine a second prediction result through the information generation model.

[0080] S105: Determine a loss value of the information generation model according to the second prediction result and the actual label corresponding to the sample data, and train the information generation model according to the loss value.

[0081] The server may input the target training instruction and the sample data into the information generation model to determine the second prediction result through the information generation model.

[0082] Among them, the data format of the sample data can be expressed as:

[0083] "text": ["Below is a description of a three-dimensional lattice structure for porous materials. This JSON schema describes a truss lattice structure, suitable for porous materials. X is the elastic modulus and the schema is a complex geometric assembly consisting of "cad_cylinders" where "cell_size" defines the size of each cell and "period_count" represents the number of periodic repetitions in three dimensions. The "models" section contains a series of combinations of CSG (Constructive Solid Geometry) operators and shapes. The "op_union" operator is a multivariate CSG operation representing a logical "or" and is used to combine cylinders into more complex geometric shapes. Each "cylinder" in this parameter is a basic CSG model defined by its "points" - the initial 3D coordinates of the cylinder - and "radius", which represents the radius of the cylinder. The following is the JSON schema details of the truss lattice structure: {JSON schema details of truss lattice structure}"].

[0084] The server can determine the loss value of the information generation model based on the deviation between the second prediction result and the actual label corresponding to the sample data, and pre-train the information generation model according to the loss value to obtain a pre-trained information generation model.

[0085] Furthermore, the server can fine-tune the target training instructions, and re-train the pre-trained information generation model through the adjusted target training instructions (the training method is the same as the pre-training method, and this manual will not go into details). Instruction fine-tuning is to optimize the target training instructions to improve the model's understanding ability and performance. Specifically, adjust the instruction content: Add contextual information: Add more background information to the instructions to help the model better understand the task context and objectives. Modify property descriptions: Ensure that the description of material properties is more precise and consistent to avoid ambiguity. Expand the scope of instructions: Include more types of material properties and structural requirements so that the model can handle a wider range of tasks. Optimized instruction example: Original instruction: "Generate a cell structure with an elastic modulus of X", optimized instruction: "Design a cell structure with at least an elastic modulus of X, while considering the stability of the material at temperature Y, and ensuring the manufacturing feasibility of the cell structure. The instruction fine-tuning format is for example:

[0086] [{“instruction”: “Design a cell structure that has an elastic modulus of at least X, taking into account the stability of the material at temperature Y, and ensuring that the cell structure is manufacturable.”,

[0087] "input": "Cellular structure design requirements: elastic modulus is at least X, temperature stability is Y.",

[0088] "output": "The following is a cell structure design that meets the above requirements: [detailed description of the cell structure]"}].

[0089] After pre-training is completed, we fine-tune the information production model according to the above format. The purpose of fine-tuning is to improve the model's understanding ability and performance. Through fine-tuning, the model can better understand and execute complex instructions and generate cell structures that meet specific needs.

[0090] Furthermore, before pre-training the information generation model, the server can also determine whether the preset information access conditions are met based on the complexity of the target training instructions. If the information access conditions are met, the preset reasoning guidance data is obtained, and the reasoning guidance data contains step description information corresponding to several reasoning steps, which is used to guide the reasoning steps performed by the information generation model in the process of determining the prediction result. After that, the target training instructions, sample data and reasoning guidance data are input into the information generation model, so that the information generation model can execute each reasoning step under the guidance of the step description information contained in the reasoning guidance data to obtain the prediction result. For ease of understanding, this manual provides a schematic diagram of the process of introducing external information, such as Figure 2 shown.

[0091] Figure 2 A schematic diagram of the process of introducing external information provided in this manual.

[0092] Among them, the server can determine whether it is necessary to introduce reasoning guidance data based on the complexity of different target training instructions based on a preset classifier, wherein if the complexity is higher than a preset threshold, it is determined that the information access condition is met, otherwise it is not met.

[0093] In addition, the server can also determine whether the preset information access conditions are met based on the test results of the basic generation model. For each training instruction, if the deviation between the first prediction result under the training instruction and the actual label of the sample data is greater than the preset deviation, then when the target training instruction obtained after adjusting the training instruction is input into the information generation model, it is determined that the information access conditions are met.

[0094] After determining that the information access conditions are met, the server can access the inference guidance data to improve the interpretability and predictability of the model in the process of generating information.

[0095] During the forward propagation of the model information generation model, the model will not only output material information, but will also be accompanied by a series of reasoning steps, which constitute a chain of thought. The server can further build reasoning guidance information based on the chain of thought formed by these steps. For each generation step, the information generation model will add explanations, such as why a specific cell shape or size is selected. How these choices meet the given volume fraction requirements.

[0096] Take the task of forward generation with volume fraction as an example:

[0097] Specifically, the present invention introduces reasoning guidance information including the following steps in the information generation model training process, including:

[0098] Initialize the cell structure: The model first initializes a cell structure and calculates its current volume fraction VF_current.

[0099] Volume fraction adjustment: If VF_current is lower than the target volume fraction VF_target, the model will decide to add new cell elements, such as cylinders, to increase the volume fraction.

[0100] Selection and positioning: The model selects the location and size of the cylinders to ensure the minimum intersection of the newly added elements with the existing structure, reducing the overlap in the calculation.

[0101] Volume calculation: The model calculates the volume of each cylinder V_cylinder_i and adds them up to get the total volume V_sum.

[0102] Intersection estimation: The model estimates the number of intersections D between cylinders and calculates the over-calculated volume V_overcounted due to the intersections.

[0103] Actual volume calculation: The model obtains the actual total volume V_total by subtracting V_overcounted from V_sum, and calculates the current volume fraction VF_current based on it.

[0104] Iterative adjustment: The model decides whether to continue adjusting the cell structure based on the comparison results between VF_current and VF_target until VF_current approaches VF_target.

[0105] In this way, the model is not only able to generate cell structures that meet specific volume fraction requirements, but also provides a clear reasoning process, making the generated cell structures more interpretable and predictable, allowing the information generation model to understand and verify the model's design decisions.

[0106] In addition, when the information access conditions are met, the server can also determine the attribute information corresponding to the material structure data generated by the information generation model according to the target training instructions, and query the preset knowledge base for information matching the attribute information as description enhancement information, and then update the target training instructions with the description enhancement information.

[0107] Specifically, the server can design an agent that can perform a specific task, namely, the cell generation task, based on the information generation model. The purpose of the transformation is to enable the agent to use an external knowledge base to enhance its reasoning ability; then introduce retrieval-augmented generation (RAG). The introduction of RAG can help the model better understand and perform complex cell generation tasks, especially when facing properties or structures that the model has not been directly trained. Specifically, the process of transformation and RAG knowledge completion is as follows: Initialize Agent: Create a class called CellGeneratorAgent, which receives a pre-trained large model and an external knowledge base as parameters during initialization; Knowledge retrieval: Implement a method retrieve_knowledge, which receives a query parameter and retrieves relevant information from an external knowledge base; Input enhancement: Implement a method augment_model_input, which combines the retrieved information with the original input data to enhance the model input; Cell structure generation: Implement a method generate_cell_structure, which is responsible for generating a cell structure based on given material properties. The method includes the following steps:

[0108] Build a query: Build a query based on material properties to retrieve relevant knowledge.

[0109] Knowledge retrieval: Execute queries to retrieve relevant information from external knowledge bases.

[0110] Input Preparation: Prepare input data and format material properties for model processing.

[0111] Input enhancement: Combine the retrieved information with the input data to form enhanced input.

[0112] Generate structure: Use the large model to process the enhanced input and generate the cell structure.

[0113] Calling an external application programming interface (API): converting the text results generated by the model into image results with 3D effects.

[0114] Among them, the CellGeneratorAgent class is responsible for encapsulating the interaction logic between the big model and the external knowledge base. The generate_cell_structure method is the core, which first builds a query to retrieve relevant knowledge, then combines the retrieved information with the original input data, and finally uses the big model to generate the cell structure. This pseudocode is a high-level abstraction, and the specific implementation details will depend on the specific technology stack and interface of the big model and external knowledge base used.

[0115] Through the above technical scheme, the present invention can match the corresponding Agent for reasoning according to the judgment result of the task classifier when executing the cell generation task, and optimize the reasoning result according to the RAG matching result, thereby improving the training efficiency and reasoning accuracy of the material structure cell generation model.

[0116] After completing the training for the information generation model, the server can deploy the information generation model to perform the actual information generation task. For ease of understanding, this specification provides a flowchart of an information generation method applied to the above information generation model, such as Figure 3 shown.

[0117] Figure 3 A schematic diagram of a flow chart of an information generation method provided in this specification includes the following steps:

[0118] S301: receiving an information generation instruction, wherein the information generation instruction is used to instruct a pre-trained information generation model to generate material structure data with specified attribute information, or to generate attribute information according to specified material structure data, and the information generation model is trained by the above-mentioned model training method;

[0119] S302: Inputting the generation instruction into the information generation model to generate material information through the information generation model.

[0120] In practical applications, the information generation model trained by the above method involves two types of generation tasks for material information, namely forward generation tasks and reverse generation tasks. For the forward generation task, the corresponding information generation instruction is used to instruct the information generation model to generate material structure data with specified attribute information. After the server inputs the information generation instruction containing the specified attribute information into the information generation model, the information generation model can output material structure data (such as material cell structure data) with the attribute information. After that, the server can perform actual tasks based on the material structure data, such as three-dimensional modeling based on the material structure data, and then build and produce corresponding parts based on the constructed model.

[0121] For reverse generation tasks, the corresponding information generation instructions are used to instruct the information generation model to determine the attribute information of the specified material structure data. After the server inputs the material structure data into the information generation model, the information generation model can generate its corresponding attribute information. The server can then perform actual tasks based on the attribute information, such as analyzing, testing and improving the mechanical properties of components according to the attribute information of the material structure data.

[0122] From the above method, it can be seen that this solution can bring the following effects:

[0123] Improve the accuracy of cell generation tasks: This approach can significantly improve the accuracy of information generation models by using small sample testing and instruction optimization for the underlying large model. By fine-tuning the training instructions, the model can better understand and perform cell generation tasks, reducing the risk of generating incorrect structures.

[0124] Reduce training resource consumption: This solution effectively reduces the computing resources and storage space required for training by using the LoRA pre-training method instead of full pre-training. Since LoRA approximates the parameter update of the model by introducing a low-rank matrix, large models can be quickly adapted to specific tasks, which is clearly reflected in the test results.

[0125] Enhance the generalization ability of the model: Through instruction diversification and secondary optimization, this solution makes the material cell generation large model have better generalization ability. Test results show that the model can adapt to different cell generation tasks and maintain high performance even when facing new or unseen tasks.

[0126] Improve the interpretability of the model: The reasoning guidance information is referenced, and the thought chain description is added to the cell generation task, making the reasoning process of the model more transparent and interpretable. This is particularly important in practical applications because it allows users to understand how the model reaches a specific result.

[0127] Optimizing the processing of specific tasks: For the task of generating cells with a specified volume fraction, this solution uses the method of introducing reasoning guidance information to improve the performance of the model in processing such specific tasks. As shown in the figure, the accuracy of the model in volume fraction control has been significantly improved.

[0128] Enhance the knowledge utilization ability of the model: By performing Agent transformation on the trained large model and finally introducing RAG, this solution achieves knowledge completion. This enables the model to dynamically retrieve and utilize relevant information, so that a wider range of knowledge background can be considered when generating cell structures, further improving the rationality and innovation of the generated structure.

[0129] The above is one or more training methods for implementing information generation models in this specification. Based on the same idea, this specification also provides a corresponding information generation model training device, such as Figure 4 shown.

[0130] Figure 4 A schematic diagram of an information generation model training device provided in this specification, comprising:

[0131] A receiving module 401 is used to obtain sample data for a basic generation model and construct training instructions based on the sample data, wherein the sample data includes material structure data with different attribute information, and the training instructions are used to guide the basic generation model to generate material structure data with specified attributes, or guide the basic generation model to generate attribute information of specified material structure data;

[0132] A first prediction module 402, configured to input the training instruction and the sample data into the basic generation model to obtain a first prediction result;

[0133] An adjustment module 403 is used to determine a test result according to the first prediction result and the actual label corresponding to the sample data, and adjust the training instruction according to the test result to obtain a target training instruction, and compress the model parameters of the basic generation model to obtain an information generation model to be trained;

[0134] A second prediction module 404, configured to input the target training instruction and the sample data into the information generation model, so as to determine a second prediction result through the information generation model;

[0135] The training module 405 is used to determine the loss value of the information generation model according to the actual label corresponding to the second prediction result and the sample data, and train the information generation model according to the loss value.

[0136] Optionally, the adjustment module 403 is specifically used to determine, based on a deviation between the first prediction result and an actual label corresponding to the sample data, information in the prediction result that does not meet a preset condition as adjustment reference information; and adjust the training instruction based on the adjustment reference information to obtain the target training instruction.

[0137] Optionally, the adjustment module 403 is specifically used to perform low-rank decomposition on the weight matrix of the basic generation model, and adjust the basic generation model based on the weight matrix obtained after the low-rank decomposition to obtain the information generation model.

[0138] Optionally, the training module 405 is specifically used to determine the loss value of the information generation model based on the prediction result and the actual label corresponding to the sample data, and pre-train the information generation model according to the loss value to obtain a pre-trained information generation model; adjust the target training instructions of the pre-trained information generation model, and train the pre-trained information generation model with the adjusted target training instructions.

[0139] Optionally, before the target training instruction and the sample data are input into the information generation model, the second prediction module 404 is further used to determine whether a preset information access condition is met according to the complexity of the target training instruction; if so, obtain preset reasoning guidance data, wherein the reasoning guidance data includes step description information corresponding to a plurality of reasoning steps, and is used to guide the reasoning steps performed by the information generation model in the process of determining the prediction result;

[0140] The second prediction module 404 is specifically used to input the target training instructions, the sample data and the reasoning guidance data into the information generation model, so as to execute each reasoning step through the information generation model under the guidance of the step description information contained in the reasoning guidance data to obtain the prediction result.

[0141] Optionally, before the target training instruction and the sample data are input into the information generation model, the second prediction module 404 is also used to determine whether a preset information access condition is met based on the complexity of the target training instruction; if so, determine the attribute information corresponding to the material structure data generated by the information generation model based on the target training instruction, and query the preset knowledge base for information matching the attribute information as description enhancement information; and update the target training instruction with the description enhancement information.

[0142] This specification also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A training method for an information generation model is provided.

[0143] This manual also provides Figure 5 The one shown corresponds to Figure 1 A schematic diagram of the electronic device. Figure 5 As mentioned above, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1The information generation model training method described. Of course, in addition to the software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0144] For the improvement of a technology, it can be clearly distinguished whether it is a hardware improvement (for example, improvement of the circuit structure of diodes, transistors, switches, etc.) or a software improvement (improvement of the method flow). However, with the development of technology, many improvements of the method flow today can be regarded as direct improvements of the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that the improvement of a method flow cannot be implemented with a hardware entity module. For example, a programmable logic device (PLD) (such as a field programmable gate array (FPGA)) is such an integrated circuit whose logical function is determined by the user's programming of the device. Designers can "integrate" a digital system on a PLD by programming themselves, without having to ask chip manufacturers to design and make dedicated integrated circuit chips. Moreover, nowadays, instead of manually making integrated circuit chips, this kind of programming is mostly implemented by "logic compiler" software, which is similar to the software compiler used when developing and writing programs, and the original code before compilation must also be written in a specific programming language, which is called hardware description language (HDL). There is not only one kind of HDL, but many kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also know that it is only necessary to program the method flow slightly in the above-mentioned hardware description languages ​​and program it into the integrated circuit, and then it is easy to obtain the hardware circuit that implements the logic method flow.

[0145] The controller may be implemented in any suitable manner, for example, the controller may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (e.g., software or firmware) executable by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, examples of which include but are not limited to the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320, and the memory controller may also be implemented as part of the control logic of the memory. It is also known to those skilled in the art that, in addition to implementing the controller in a purely computer-readable program code manner, the controller may be implemented in the form of a logic gate, a switch, an application-specific integrated circuit, a programmable logic controller, and an embedded microcontroller by logically programming the method steps. Therefore, such a controller may be considered as a hardware component, and the devices for implementing various functions included therein may also be considered as structures within the hardware component. Or even, the devices for implementing various functions may be considered as both software modules for implementing the method and structures within the hardware component.

[0146] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0147] For the convenience of description, the above device is described in various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0148] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented 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.

[0149] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0150] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0151] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0152] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0153] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0154] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0155] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0156] It should be understood by those skilled in the art that the embodiments of this specification may be provided as methods, systems or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented 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.

[0157] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0158] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0159] The above description is only an embodiment of the present specification and is not intended to limit the present specification. For those skilled in the art, the present specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification shall be included in the scope of the claims of the present specification.

Claims

1. A method for training an information generation model, characterized in that: include: Acquire sample data for a basic generation model, and construct training instructions based on the sample data, wherein the sample data includes material structure data with different attribute information, and the training instructions are used to guide the basic generation model to generate material structure data with specified attributes, or guide the basic generation model to generate attribute information of specified material structure data, wherein the material structure data is cell structure data of the material; Inputting the training instruction and the sample data into the basic generation model to obtain a first prediction result; Determine a test result according to the first prediction result and the actual label corresponding to the sample data, and adjust the training instruction according to the test result to obtain a target training instruction, and compress the model parameters of the basic generation model to obtain an information generation model to be trained; wherein adjusting the training instruction includes: adding context information, modifying attribute description, and expanding the instruction range. Inputting the target training instruction and the sample data into the information generation model to determine a second prediction result through the information generation model; According to the actual label corresponding to the second prediction result and the sample data, the loss value of the information generation model is determined, and the information generation model is trained according to the loss value.

2. The method according to claim 1, characterized in that The training instruction is adjusted according to the test result to obtain a target training instruction, which specifically includes: According to the deviation between the first prediction result and the actual label corresponding to the sample data, determining information in the prediction result that does not meet the preset condition as adjustment reference information; The training instruction is adjusted according to the adjustment reference information to obtain the target training instruction.

3. The method according to claim 1, characterized in that The model parameters of the basic generation model are compressed to obtain the information generation model to be trained, specifically including: The weight matrix of the basic generation model is subjected to low-rank decomposition, and the basic generation model is adjusted based on the weight matrix obtained after the low-rank decomposition to obtain the information generation model.

4. The method according to claim 3, characterized in that Determining a loss value of the information generation model according to the prediction result and the actual label corresponding to the sample data, and training the information generation model according to the loss value, specifically comprising: Determining a loss value of the information generation model according to the prediction result and the actual label corresponding to the sample data, and pre-training the information generation model according to the loss value to obtain a pre-trained information generation model; The target training instruction of the pre-trained information generation model is adjusted, and the pre-trained information generation model is trained using the adjusted target training instruction.

5. The method according to claim 1, characterized in that Before inputting the target training instruction and the sample data into the information generation model, the method further includes: Determining whether a preset information access condition is met according to the complexity of the target training instruction; If yes, then obtaining preset reasoning guidance data, wherein the reasoning guidance data includes step description information corresponding to a plurality of reasoning steps, and is used to guide the reasoning steps performed by the information generation model in the process of determining the prediction result; Inputting the target training instruction and the sample data into the information generation model to determine a second prediction result through the information generation model specifically includes: The target training instruction, the sample data and the reasoning guidance data are input into the information generation model, so that each reasoning step is executed through the information generation model under the guidance of the description information of each step contained in the reasoning guidance data to obtain the prediction result.

6. The method according to claim 1, characterized in that Before inputting the target training instruction and the sample data into the information generation model, the method further includes: Determine whether the preset information access conditions are met according to the complexity of the target training instruction; If yes, then determining the attribute information corresponding to the material structure data generated by the information generation model according to the target training instruction, and searching for information matching the attribute information in a preset knowledge base as description enhancement information; The target training instruction is updated using the description enhancement information.

7. A method for generating information, characterized in that: include: Receiving an information generation instruction, wherein the information generation instruction is used to instruct a pre-trained information generation model to generate material structure data with specified attribute information, or to generate attribute information according to specified material structure data, and the information generation model is trained by the method described in any one of claims 1 to 6 above; The generation instruction is input into the information generation model to generate material information through the information generation model.

8. An information generation model training device, characterized in that: include: A receiving module, used to obtain sample data for a basic generation model, and construct training instructions based on the sample data, wherein the sample data includes material structure data with different attribute information, and the training instructions are used to guide the basic generation model to generate material structure data with specified attributes, or guide the basic generation model to generate attribute information of specified material structure data, wherein the material structure data is cell structure data of the material; A first prediction module, used for inputting the training instruction and the sample data into the basic generation model to obtain a first prediction result; An adjustment module is used to determine a test result according to the first prediction result and the actual label corresponding to the sample data, and adjust the training instruction according to the test result to obtain a target training instruction, and compress the model parameters of the basic generation model to obtain an information generation model to be trained; wherein adjusting the training instruction includes: adding context information, modifying attribute description, and expanding the instruction range. A second prediction module, used for inputting the target training instruction and the sample data into the information generation model to determine a second prediction result through the information generation model; A training module is used to determine the loss value of the information generation model according to the actual label corresponding to the second prediction result and the sample data, and train the information generation model according to the loss value.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method described in any one of claims 1 to 7 is implemented.

10. An electronic 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, the method described in any one of claims 1 to 7 is implemented.

Images