A large model training method and system, an electronic device, and a storage medium
By screening and evaluating student and teacher models, constructing a distillation dataset, and conducting multi-stage reinforcement learning, the problems of high cost and deployment difficulty in large model training are solved, achieving efficient and economical complex logical reasoning capabilities, suitable for model training and deployment for complex tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SI-TECH INFORMATION TECH CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-29
Smart Images
Figure CN122114068A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence model collaboration technology, and in particular to a large model training method, system, electronic device and storage medium. Background Technology
[0002] With the rapid development of artificial intelligence technology, large language models have demonstrated powerful capabilities in basic tasks such as text generation and multi-turn dialogue. However, in high-end cognitive tasks such as complex logical reasoning and causal analysis, there are still problems such as untraceable reasoning process, poor interpretability, high training cost, and great difficulty in deployment.
[0003] Current mainstream large-scale model training methods mostly rely on large-scale parameter expansion and massive data feeding, resulting in the training process consuming huge amounts of computing resources. A single training session for a model with hundreds of billions of parameters often requires thousands of GPUs to run continuously for several weeks, with training costs reaching millions of dollars. At the same time, model deployment is constrained by memory and energy consumption limitations. The weight storage for a 175-parameter model alone requires more than 700GB of space, which can only be deployed by relying on expensive professional clusters, making it difficult for small and medium-sized enterprises to afford.
[0004] In terms of cultivating reasoning ability, existing methods lack a systematic transfer mechanism for thought processes. They often stimulate the model's thought chain ability through simple prompting engineering or rely on manually labeled high-quality reasoning data, which has drawbacks such as high data costs, poor scalability, and inconsistent reasoning logic. Some training schemes adopt a two-stage model of pre-training + fine-tuning, but the parameter optimization efficiency during fine-tuning is low, and no training strategy is designed for the progressive length requirements of the thought chain. This leads to the model being prone to missing steps or redundant reasoning problems in long-term reasoning tasks.
[0005] Furthermore, existing teacher model selection lacks a clear multi-dimensional evaluation system, relying solely on the accuracy of the final answer and neglecting the logic and effectiveness of the reasoning process. This results in inconsistent quality of distilled data, affecting the training effect of student models. These issues collectively constitute the main obstacles to the industrial deployment of large-scale models in complex reasoning tasks, necessitating a training method that balances training efficiency, inference performance, and deployment economy. Summary of the Invention
[0006] To address the shortcomings mentioned above, this invention provides a large model training method, system, electronic device, and storage medium.
[0007] To achieve the above objectives, the present invention provides a large model training method, comprising: Based on the domain attributes of the target task, quantitative performance requirements, and resource constraints, the student model selection is determined; based on the ability to generate the thought chain for the target task, the teacher model selection is determined. The data to be processed is input into the teacher model to obtain response data including explicit thought processes and final answers. The response data is then processed to form a distillation dataset for thought chain transfer. By mixing some distillation data and thought chain data from general technology scenarios as a cold start dataset, the student model was fine-tuned to obtain a cold start student model. The distillation dataset is input into the cold-start student model for multi-stage reinforcement learning training to obtain the target large model.
[0008] Preferably, the student model is selected based on the domain attributes, quantitative performance requirements, and resource constraints of the target task; the teacher model is selected based on the ability to generate the thought chain for the target task. The size range of the student model is limited based on the aforementioned resource constraints and inference performance. Based on the quantified performance requirements, an inference performance evaluation index is constructed, which includes at least one of Perplexity, Accuracy, and F1. If all the inference performance evaluation indicators are not lower than the corresponding preset thresholds, then the model is determined to be the student model.
[0009] The size range of the teacher model is limited based on the target task's ability to generate thought chains. Based on the domain attributes of the target task, a thinking chain evaluation index is constructed, which includes key step recall rate, logical coherence score, terminology accuracy score, failure step ratio, reasoning efficiency, and robustness. If both the thought chain evaluation index and the reasoning performance evaluation index are not lower than the corresponding preset threshold, then the model is determined to be the teacher model.
[0010] Preferably, the step of inputting the data to be processed into the teacher model to obtain response data including explicit thought processes and final answers, and processing the response data to form a distillation dataset for thought chain transfer, includes: The data to be processed is input into the teacher model to generate response data including input and output. The input is the question part of the data to be processed, and the output includes explicit thought process and final answer, wherein the explicit thought process is wrapped in preset special characters. The response data is sequentially cleaned, abnormal sample filtered, and word segmenter adapted, and then divided into a training set, a validation set, and a test set according to a preset ratio. The training set, validation set, and test set form a distillation dataset for mind chain transfer.
[0011] Preferably, the step of mixing partial distillation data and thought chain data from general technology scenarios to create a cold start dataset for fine-tuning the student model (i.e., the cold start process) to obtain the cold start student model includes: Determine the student model, which is a dedicated inference model that meets resource constraints and inference performance limitations; Based on the cold start dataset, the student model is supervised fine-tuned using low-rank adaptation techniques. The fine-tuning process includes: Hyperparameter settings, including learning rate, training period, data truncation length, low-rank decomposition dimension, and the dialogue template is deepseek_r1_distill_qwen; The context length is set by adjusting the position encoding parameters of the student model through interpolation, so that the fine-tuned student model can be adapted to long thought chain data. The vocabulary is set by parsing the tokens in the cold start dataset that are not covered by the student model tokenizer, and by modifying the special character list in the tokenizer.json file to add the new tokens to the model vocabulary.
[0012] Preferably, the distillation dataset is input into a cold-start student model for multi-stage reinforcement learning training to obtain the target large model. This includes: Samples for each difficulty range are clustered from the distillation dataset to form training subsets for multiple difficulty ranges; Based on the aforementioned training subset, a population-relative strategy optimization algorithm is employed, and training is conducted in a three-stage manner with progressive difficulty gradients. Each training stage includes: Phase 1: The response output length threshold is set to 8K, and the training objective is to optimize the logical coherence and completeness of key steps in short thought chains. Phase 2: The response output length threshold is set to 16K, and the training objective is to strengthen the step connection and terminology accuracy of medium-length thought chains. Phase 3: The response output length threshold is set to 24K, and the training objective is to improve the integrity and redundancy control of long thought chains. The phase switching condition is as follows: when the reward value increases by ≤1% for two consecutive epochs and the response length truncation rate is ≥5% in the current phase, the next phase is entered; until the reward value increases by ≤0.5% for three consecutive epochs, the output is the target large model of the enhanced mind chain; The response output length threshold refers to the maximum number of tokens allowed to be generated by the fine-tuned student model.
[0013] Preferably, the performance metrics of the target large model include a recall rate of ≥85% for key steps in the thought chain, a failure rate of ≤30%, and inference performance metrics that meet the quantitative performance requirements of the target task.
[0014] The present invention also provides a large model training system, comprising: The acquisition module is used to determine the student model selection based on the domain attributes of the target task, quantitative performance requirements, and resource constraints; and to determine the teacher model selection based on the target task's ability to generate thought chains. The input module is used to input the data to be processed into the teacher model to obtain response data including explicit thought processes and final answers, and to process the response data to form a distillation dataset for thought chain transfer. The fine-tuning module is used to mix some distillation data and thought chain data from general technical scenarios as a cold start dataset to fine-tune the student model and obtain a cold start student model. The training module is used to input the distillation dataset into the cold-start student model for multi-stage reinforcement learning training to obtain the target large model.
[0015] The present invention also provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the method described above.
[0016] The present invention also provides a storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention ensures the high-quality thinking chain generation ability of the teacher model and the efficient reasoning ability of the student model through dual screening of thinking chain evaluation and reasoning performance evaluation; phased reinforcement learning is trained progressively according to the length gradient of thinking chain, so that the recall rate of key steps of the target large model thinking chain is ≥85%, and the reasoning performance is more than 4.4% better than traditional methods. Attached Figure Description
[0018] Figure 1 This is a flowchart of the large model training method of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] Reference Figure 1 This invention provides a large model training method, comprising: Based on the domain attributes of the target task, quantitative performance requirements, and resource constraints, the student model selection is determined; based on the ability to generate the thought chain for the target task, the teacher model selection is determined. Specifically, based on resource constraints (such as single-card capacity, inference latency limit, etc.) and the reasoning ability of the thought chain, models are selected from the preset model library as student models and teacher models respectively. The preset model library includes general large-scale inference models and special models fine-tuned with private data to ensure that the candidate models meet the actual application requirements in terms of resource consumption and professional capabilities.
[0021] Constructing a thinking chain evaluation index: Based on the domain attributes of the target task, construct a thinking chain evaluation index system covering the quality and overall effectiveness of the reasoning process, specifically including: Key step recall: The proportion of core reasoning steps for the target task contained in the thought process chain generated by the candidate model; Logical coherence score: A quantifiable measure of the semantic fluency of adjacent reasoning steps in a thought process; Terminology accuracy score: The percentage of correct use of domain-specific terms in the thought process; Failure Steps Percentage (FSF): The percentage of steps in a thought process that contain logical contradictions or factual errors; Reasoning efficiency: The percentage of effective reasoning steps in the thought process that are strongly related to the final answer; Robustness: The ability of a candidate model to maintain the stability of its thought process when the input contains slightly interfering information.
[0022] Construct inference performance evaluation metrics: Based on the quantitative performance requirements of the target task, select at least one core inference performance evaluation metric, including perplexity, accuracy, F1 score, etc., to verify the reliability of the final inference results of the candidate model.
[0023] Model selection criteria: If the candidate model's thought chain evaluation index is not lower than the corresponding preset threshold (the threshold is dynamically adjusted according to task difficulty, industry standards, and business needs), then the candidate model is determined to have the ability to generate the thought chain for the target task and is identified as a teacher model; otherwise, it is eliminated. If the candidate model's reasoning performance evaluation index is not lower than the corresponding preset threshold (the threshold is dynamically adjusted according to task difficulty, industry standards, and business needs), then the candidate model is determined to have efficient reasoning ability and is identified as a student model; otherwise, it is eliminated.
[0024] The data to be processed is input into the teacher model to obtain response data including explicit thought processes and final answers. The response data is then processed to form a distilled dataset for thought chain transfer. Specifically, the data to be processed is input into the filtered teacher model, generating binary structured response data containing both input and output. The input is the question portion of the data to be processed (the module carrying the core requirements of the target task); the output includes the explicit thought process and the final answer. The explicit thought process is expressed through preset special characters (such as...). <thinking> and< / thinking> The package is broken down into continuous and traceable steps according to the reasoning logic of the target task, and the final answer is generated based on the explicit thought process.
[0025] Response data preprocessing: Perform three standardization steps on the response data sequentially: Data cleaning process: Removes duplicate data, garbled characters, blank characters, and other meaningless content to ensure data purity; Abnormal sample filtering: Remove samples with logical contradictions in the thought process or incorrect final answers to ensure data validity; Tokenizer adaptation: Unify the tokenizer encoding rules between the teacher model and the subsequent student model to achieve token alignment and avoid encoding incompatibility issues.
[0026] Dataset partitioning: The preprocessed response data is divided into training set, validation set and test set according to a preset ratio (recommended training set: validation set: test set = 8:1:1). The three sets together constitute the distillation dataset for mind chain transfer, which is used for model training, process monitoring and performance testing, respectively.
[0027] By mixing some distillation data and thought chain data from general technology scenarios, the student model is fine-tuned using a cold start dataset (i.e., the cold start process) to obtain a cold start student model. Specifically, a dedicated inference model that meets resource constraints and inference performance limits is selected as the student model. The parameter size is generally no less than 7 bytes. This parameter size range can balance training costs and deployment economy while ensuring the learning ability of the thought chain. Optional specifications include 7 bytes, 13 bytes, and 32 bytes. Using a cold-start dataset as the training set, low-rank adaptation techniques are employed for supervised fine-tuning of the student model. Only a small number of low-rank parameters are optimized without changing the model's core parameters, significantly reducing training costs. The fine-tuning process specifically includes: Hyperparameter settings: Load preset hyperparameter sets, including learning rate, training period, data truncation length, low-rank decomposition dimension, etc. The dialogue template uses deepseek_r1_distill_qwen to ensure a balance between fine-tuning effect and efficiency. Context length setting: The position encoding parameters of the student model are adjusted by interpolation so that the fine-tuned student model can adapt to long thought chain data and meet the context length requirements of complex reasoning tasks. Vocabulary settings: Parse tokens (such as domain-specific terms) in the training set that are not covered by the student model's tokenizer. Add the new tokens to the model's vocabulary by modifying the special character list in the tokenizer.json file to ensure that no semantics are lost.
[0028] Fine-tuning output: After fine-tuning, a fine-tuned student model with the basic ability to generate thought chains is obtained. This student model can output explicit thought chains in a structured format, laying the foundation for subsequent reinforcement learning.
[0029] The distillation dataset is input into the fine-tuned student model for reinforcement learning training to obtain the target large model.
[0030] From the training set of the distillation dataset, data samples covering easy, medium, and complex difficulty ranges are divided to form a multi-level reinforcement learning training set, ensuring the comprehensiveness and relevance of the training.
[0031] A group-relative strategy optimization algorithm is adopted, and the student model is trained and fine-tuned in a three-stage manner according to the difficulty gradient. The training objectives and configurations for each stage are as follows: Phase 1: The response output length threshold is set to 8K (i.e. 8192 tokens). The training objective is to optimize the logical coherence and key step integrity of short thought chains, focusing on the stability of basic reasoning paths. Phase 2: The response output length threshold is set to 16K (i.e., 16384 tokens). The training objective is to strengthen the step connection and terminology accuracy of medium-length thought chains and improve the adaptability to complex reasoning scenarios. The third stage: The response output length threshold is set to 24K (i.e. 24576 tokens). The training objective is to improve the integrity and redundancy control of long thought chains and meet the thought chain expression requirements of most target tasks.
[0032] The response output length threshold refers to the maximum number of tokens allowed to generate a response by fine-tuning the student model.
[0033] Phase switching and training convergence: The phase switching condition is: when the reward value increases by ≤1% for 2 consecutive epochs in the current phase, and the response length truncation rate is ≥5% (the proportion of response samples that are truncated due to reaching the current phase length threshold), the next phase is entered; until the reward value increases by ≤0.5% for 3 consecutive epochs, the training is judged to be converged, and the target large model with enhanced thinking chain is output. Example 1
[0034] This embodiment takes the cultural tourism knowledge question answering and route planning task as the target task. The task requires the model to have a complete inference chain. The quantitative performance requirements are: accuracy ≥ 85%, key step recall ≥ 85%, failure step ratio ≤ 30%, deployment environment is a single-card GPU server (video memory ≥ 24GB), and the resource constraint is model inference latency ≤ 2s.
[0035] This embodiment presents a large model training method, including: obtaining a teacher model and a student model: Resource constraints: Select inference models with ≤32B parameters and single-card capacity. Candidate models include the Deepseek series of distillation models. Thinking chain evaluation indicators: key step recall ≥85%, logical coherence score ≥80, terminology accuracy score ≥80, failure rate ≤30%, reasoning efficiency ≥75%, robustness ≥80%; Inference performance evaluation metrics: Accuracy ≥ 85%, F1 ≥ 0.8; Screening results: Deepseek-r1 meets the requirements of the thinking chain assessment index and is identified as a teacher model.
[0036] DeepSeek-R1-Distill-7B meets the requirements of resource constraints and inference performance evaluation metrics, and is therefore selected as the student model.
[0037] Constructing the distillation dataset: Data to be processed: includes cultural and tourism knowledge Q&A (45,000 entries) and route planning requirements (5,000 entries). The question section is a structured module consisting of user needs, attraction preferences, and time constraints. Response data generation: The teacher model output format is as follows <thinking> User Needs Analysis: Requires a 3-day family trip → Attraction Matching: Filter suitable 5A scenic spots for families → Time Conflict Detection: Exclude closed attractions → Route Generation: Plan routes based on geographical distribution< / thinking>Final Itinerary: Day 1: XX Scenic Area → Day 2: XX Amusement Park → Day 3: XX Museum; Preprocessing: Clean 300 duplicate samples, filter 200 logically contradictory samples, adapt to Qwen word segmenter, and divide into training set of 40,000 samples, validation set of 5,000 samples, and test set of 5,000 samples in an 8:1:1 ratio, and store in JSON format.
[0038] Student model fine-tuning: Student model: The DeepSeek-R1-Distill-7B inference-specific model was selected; Hyperparameter settings: learning rate 5.0e-5, training period 3 epochs, data truncation length 20000, low-rank decomposition dimension 64, dialogue template used is deepseek_r1_distill_qwen; Context length setting: The maximum context length of the model is extended from 4K to 32K using interpolation. Vocabulary settings: Add cultural and tourism-specific tokens such as intangible cultural heritage experiences and parent-child facilities to the vocabulary list; Fine-tuning results: The fine-tuned student model was obtained, with a recall rate of 78% for key steps in the thought process chain and a logical coherence score of 72, which meets the basic ability requirements.
[0039] Reinforcement learning training yields a large target model: Dedicated training set: composed of samples of easy (10,000), medium (8,000), and complex (2,000) difficulty levels selected from the training set; Three-stage training: Phase 1 (8K length): Train for 2 epochs, reduce the reward value increase from 5% to 0.8%, and the response cutoff rate to 6%, then proceed to Phase 2; Phase 2 (16K length): Train for 3 epochs, reduce the reward value increase from 3% to 0.9%, and the response cutoff rate to 5%, then proceed to Phase 3; Phase 3 (24-epoch length): Train for 4 epochs, and the reward value increases by ≤0.5% for 3 consecutive epochs, at which point the training converges; The target large model performance is as follows: recall rate of key steps in the thought chain is 87%, failure rate is 28%, accuracy is 88%, and inference latency is 1.8s, which meets the requirements of the target task.
[0040] The target large model trained by the method of this invention achieves an accuracy of 88% in cultural and tourism knowledge question answering tasks and a user satisfaction rate of 82% in route planning tasks. The training process consumes only 24 GPUs × 8 hours of computing power, which reduces the cost by 72% compared to traditional full-scale fine-tuning. When deployed, it occupies 18GB of video memory, supports single-card deployment, and has an inference latency of 1.8s, meeting the needs of real-time interaction.
[0041] This embodiment verifies the significant advantages of the method of the present invention in reducing training costs, improving inference performance, and optimizing deployment flexibility. It is applicable to various inference tasks that require complex thought processes and has broad industrial application value.
[0042] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for training large models, characterized in that, include: Based on the domain attributes of the target task, quantitative performance requirements, and resource constraints, the student model selection is determined; The teacher model selection is determined based on the teacher's ability to generate the thought chain for the target task. The data to be processed is input into the teacher model to obtain response data including explicit thought processes and final answers. The response data including explicit thought processes and final answers is processed to form a distillation dataset for thought chain transfer. The distillation data and the thought chain data under the general technology scenario are mixed together to form a cold start dataset for fine-tuning the student model, thus obtaining a cold start student model. The distillation dataset is input into the cold-start student model for multi-stage reinforcement learning training to obtain the target large model.
2. The large model training method according to claim 1, characterized in that, The process involves determining the student model selection based on the domain attributes, quantitative performance requirements, and resource constraints of the target task; and determining the teacher model selection based on the target task's ability to generate thought chains. This includes: The size range of the student model is limited based on the aforementioned resource constraints and inference performance. Based on the quantified performance requirements, an inference performance evaluation index is constructed, which includes at least one of Perplexity, Accuracy, and F1. If all the inference performance evaluation indicators are not lower than the corresponding preset threshold, then the model is determined to be the student model; The size range of the teacher model is limited based on the target task's ability to generate thought chains. Based on the domain attributes of the target task, a thinking chain evaluation index is constructed, which includes key step recall rate, logical coherence score, terminology accuracy score, failure step ratio, reasoning efficiency, and robustness. If both the thought chain evaluation index and the reasoning performance evaluation index are not lower than the corresponding preset threshold, then the model is determined to be the teacher model.
3. The large model training method according to claim 2, characterized in that, The process involves inputting the data to be processed into the teacher model to obtain response data including explicit thought processes and final answers. This response data, including explicit thought processes and final answers, is then processed to form a distillation dataset for thought chain transfer, comprising: The data to be processed is input into the teacher model to generate response data including input and output. The input is the question part of the data to be processed, and the output includes explicit thought process and final answer, wherein the explicit thought process is wrapped in preset special characters. The response data is sequentially cleaned, abnormal sample filtered, and word segmenter adapted, and then divided into a training set, a validation set, and a test set according to a preset ratio. The training set, validation set, and test set form a distillation dataset for mind chain transfer.
4. The large model training method according to claim 1, characterized in that, The process of mixing partial distillation data and thought chain data from general technology scenarios to create a cold-start dataset for fine-tuning the student model, resulting in a cold-start student model, includes: Determine the student model, which is a dedicated inference model that meets resource constraints and inference performance limitations; Based on the cold start dataset, the student model is supervised fine-tuned using low-rank adaptation techniques. The fine-tuning process includes: Hyperparameter settings, including learning rate, training period, data truncation length, low-rank decomposition dimension, and the dialogue template is deepseek_r1_distill_qwen; The context length is set by adjusting the position encoding parameters of the student model through interpolation, so that the fine-tuned student model can be adapted to long thought chain data. The vocabulary is set by parsing the tokens in the cold start dataset that are not covered by the student model tokenizer, and by modifying the special character list in the tokenizer.json file to add the new tokens to the model vocabulary.
5. The large model training method according to claim 4, characterized in that, The process involves inputting the distillation dataset into a cold-start student model for multi-stage reinforcement learning training to obtain the target large model. This includes: Samples for each difficulty range are clustered from the distillation dataset to form training subsets for multiple difficulty ranges; Based on the aforementioned training subset, a population-relative strategy optimization algorithm is employed, and training is conducted in a three-stage manner with progressive difficulty gradients. Each training stage includes: Phase 1: The response output length threshold is set to 8K, and the training objective is to optimize the logical coherence and completeness of key steps in short thought chains. Phase 2: The response output length threshold is set to 16K, and the training objective is to strengthen the step connection and terminology accuracy of medium-length thought chains. Phase 3: The response output length threshold is set to 24K, and the training objective is to improve the integrity and redundancy control of long thought chains. The phase switching condition is as follows: when the reward value increases by ≤1% for two consecutive epochs and the response length truncation rate is ≥5% in the current phase, the next phase is entered; until the reward value increases by ≤0.5% for three consecutive epochs, the output is the target large model of the enhanced mind chain; The response output length threshold refers to the maximum number of tokens allowed to be generated by the fine-tuned student model.
6. The large model training method according to claim 5, characterized in that, The performance metrics of the target large model include a recall rate of ≥85% for key steps in the thought chain, a failure rate of ≤30%, and inference performance metrics that meet the quantitative performance requirements of the target task.
7. A large model training system, characterized in that, include: The acquisition module is used to determine the student model selection based on the domain attributes of the target task, quantitative performance requirements, and resource constraints. The teacher model selection is determined based on the teacher's ability to generate the thought chain for the target task. The input module is used to input the data to be processed into the teacher model to obtain response data including explicit thought processes and final answers, and to process the response data to form a distillation dataset for thought chain transfer. The fine-tuning module is used to mix some distillation data and thought chain data from general technical scenarios as a cold start dataset to fine-tune the student model and obtain a cold start student model. The training module is used to input the distillation dataset into the cold-start student model for multi-stage reinforcement learning training to obtain the target large model.
8. An electronic device, characterized in that, It includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the method described in any one of claims 1 to 6.
9. A storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the method described in any one of claims 1 to 6.