A large model routing data synthesis method with diversified expression styles

Abstract

The application discloses a large model routing data synthesis method with diversified expression styles, constructs a user agent to simulate the language style and preference of a real user, generates query questions with diversified styles and clear semantics, inputs the query questions into a candidate language model for reasoning, and generates a benchmark answer by a large language model with superior performance; based on the comparison between the output of the candidate model and the benchmark answer, the content quality is automatically evaluated, the user satisfaction evaluation is simulated in combination with the user agent features, and multi-dimensional routing labels are comprehensively generated; finally, the question, the candidate model output and the routing label are packaged to form a routing data set, which is used for training or evaluating a large model router. The application can automatically synthesize a routing data set covering multiple expression styles, multiple question types and performance differences of multiple models, significantly improves the generalization ability and practicability of a routing model in a real open domain question and answer scene, and does not require high artificial labeling cost.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of large language model application, and particularly relates to a large model routing data synthesis method with diversified expression styles. BACKGROUND

[0002] With the wide application of large language models, how to efficiently select the model with the optimal performance and the lowest cost from the candidate model set, i.e., to realize large model routing, becomes crucial. The existing routing methods usually rely on training a dedicated routing model, and the performance of the model depends largely on the quality and diversity of the training data. In order to train such a routing model, the routing data required should contain diversified user questions as input, and the specific performance of each candidate language model under these inputs, for constructing the corresponding routing labels.

[0003] At present, the construction of routing data faces significant bottlenecks, and the existing technical solutions mainly rely on manual annotation or benchmark datasets to construct routing training data. For example, the document [E. Frick, C. Chen and J. Tennyson, et al. From prompts to rankings: large language model evaluation for prompt adaptation. arXiv preprint: 2502.14855] uses more than 2 million real user preference voting data on the Chatbot Arena platform for model output, takes each "prompt-win / loss" pair as a training sample, fine-tunes the large language model through the Bradley-Terry model, makes it output the relative win rate of each candidate model, and thus realizes the routing decision. In addition, the document [Chen Shuhao, Jiang Weisen, Lin Baiqiang, et al. RouterDC: dual contrast learning query routing framework for large language model integration. 38th Conference on Neural Information Processing Systems. 2024] based on the annotated benchmark dataset, makes all candidate models reason, takes the task accuracy as the performance label, and then trains the routing model using the contrast learning framework. However, these existing methods have obvious limitations. The routing label construction process needs to collect diversified question samples first, then reason by each candidate model, and finally rely on manual evaluation or existing annotated datasets for performance evaluation. This process highly depends on manual work, resulting in high cost, low efficiency, and difficulty in large-scale expansion.

[0004] In addition, user questions in real scenarios are extremely diverse in type and expression style, and existing training data has two major defects: first, questions are mostly from benchmark test sets, with strict and standardized styles, which cannot reflect the random and varied expression methods of users in actual applications; second, the data scale is limited and coverage is narrow, making it difficult to cover a wide range of question and answer topics and forms. These problems together result in insufficient generalization ability of the routing model in the face of real complex queries, making it difficult to accurately select the optimal model. SUMMARY

[0005] In view of the problems of insufficient training data quality and diversity in the existing large model routing method, the present application provides a large model routing data synthesis method with diverse expression styles. Through steps such as constructing a user agent, synthesizing style-diverse queries, candidate model reasoning, and routing label generation, a routing dataset covering multiple expression styles, multiple question types, and multiple model performance differences is automatically synthesized, significantly improving the generalization ability and practicality of the routing model in real open-domain question answering scenarios.

[0006] A large model routing data synthesis method with diverse expression styles includes the following steps:

[0007] (1) Obtain a seed user question dataset;

[0008] (2) Construct a user agent, which samples user features including language expression style, interest preference, and question focus from the seed user question dataset, and combines them to form a personalized user profile;

[0009] (3) Generate user query questions with diverse expression styles and clear semantic targets using LLM (Large Language Model) based on the user agent. The query questions are derived from real user question extensions or generated from language style guidance in network corpus to form diverse question variants;

[0010] (4) Input the user query question into a set of candidate language models for reasoning to obtain the output answers of each candidate language model;

[0011] (5) Input the user query question into a superior benchmark LLM, and use its output as the benchmark answer to evaluate the quality of the candidate language model output;

[0012] (6) Based on the output answers of the candidate language models and the benchmark answers, perform multi-dimensional information fusion including content quality evaluation, user satisfaction simulation, uncertainty analysis, and preference reward auxiliary scoring to obtain the routing quality label of each candidate language model;

[0013] (7) Package the user query question, the user profile constructed by the user agent, the model output answer, the benchmark answer, and the routing quality label as the routing data of the candidate language model for subsequent model training or large model routing evaluation.

[0014] Further, in step (1), real user questions are collected from public datasets such as question and answer communities, forums, social media, search engine logs, etc. to construct a seed user question dataset, ensuring that the questions in the dataset are diverse in style and widely in topic.

[0015] Further, the specific implementation of step (2) is as follows:

[0016] S21: Extract representative user structured features from the seed user question dataset through clustering and feature extraction methods, ensuring that the extracted features are both representative and diverse;

[0017] S22: Establish a structured mapping system between functional dimensions and language implementation forms using the six language function dimensions in Biber's multidimensional analysis method. Based on this system, select representative and significantly different expression style combinations through path sampling strategies to form a control path for diversified generation. During the sampling process, combine dimension distribution characteristics and style coverage indicators to ensure that the sampling results achieve balance in expression diversity and coverage breadth. Then, convert the language function dimensions involved in the path into descriptive prompt word templates that LLM can recognize, which are used to guide LLM to generate query texts with specific expression styles;

[0018] S23: Based on the extracted user structured features, use random sampling or strategic sampling methods to select several feature combinations to form structured user portraits. During the sampling process, introduce a temperature parameter to adjust diversity, while setting a coverage balance constraint to avoid excessive concentration of single type of user feature combination;

[0019] S24: Use the structured user portrait as input conditions to call LLM to generate behavior descriptions of user agents as conditional information for subsequent query question generation.

[0020] Further, the specific implementation of step (3) is: input the user portrait constructed by the user agent into the predefined descriptive prompt word template, and generate query questions that conform to its characteristic style through LLM; If the real user question is used as a seed, style migration is performed on the original question, including rewriting in multiple dimensions such as tone, length, syntax structure, and rhetorical means to ensure the diversity of expression style. In this process, by setting style conversion targets and content retention constraints, it is ensured that the semantic of the generated question does not deviate from the original intent; If network corpus is used as a reference, search engines or scraping platforms are used to obtain text content related to the question topic, extract style features of the text including syntax structure, word frequency distribution, and rhetorical patterns to form style labels, and then guide LLM to perform style migration or imitation based on these labels, thereby generating user query questions consistent with the reference style.

[0021] Further, the candidate language model set in step (4) contains a variety of language models of different sizes, different architectures, different training targets, different domain optimizations, and different providers; The benchmark LLM in step (5) selects the currently optimal large language model on the market (such as GPT-4 or Gemini Ultra).

[0022] Further, the specific implementation of step (6) is as follows:

[0023] For content quality evaluation, the evaluation covers multiple indicators such as semantic similarity, information accuracy, content completeness, question relevance, language fluency, and redundancy. The LLM performs unsupervised text quality scoring on these indicators through prompt words and fuses the total quality score of the generated model output answer.

[0024] For uncertainty analysis, the consistency of candidate language model output in multiple generations, language feature heuristic indicators (such as syntactic complexity and ambiguity word frequency), and model self-scoring confidence are comprehensively analyzed to assist in judging the credibility of candidate language model output.

[0025] For user satisfaction simulation, the LLM simulates user satisfaction feedback based on the user portrait constructed by the user agent, focusing on evaluating the performance of candidate language model output answers in problem solving, information satisfaction, easy understanding, and overall satisfaction in subjective experience dimensions. This process relies on prompt-driven generation capabilities and uses LLM to simulate the needs and preferences of different user groups.

[0026] For preference reward auxiliary scoring, the user query question, candidate language model output answer, and benchmark answer are constructed as comparative inputs to call the publicly trained PRM (Preference Reward Model), which outputs user preference ranking results and corresponding scores.

[0027] Finally, the quantitative indicators of the above four aspects are fused to generate the routing quality label of the candidate language model.

[0028] Further, for content quality evaluation, the total quality score of the model output answer is calculated by the following expression:

[0029] Score i =λ1·Sim(A (i) ,A ref )+λ2·Acc(A (i) ,A ref )+λ3·Comp(A (i) ,A ref )+λ4·Rel(A (i) ,q)+λ5·Flu(A (i) )-λ6·Red(A (i) )

[0030] Where: for any candidate language model M i , A (i) is the output answer of M i , A ref is the benchmark answer, Scorei For A (i) The overall quality score, Sim(A) (i) A ref ) is A (i) With A ref Semantic similarity, Acc(A (i) A ref ) is A (i) Compared to A ref Information accuracy, Comp(A) (i) A ref ) is A (i) Compared to A ref Content completeness, Rel(A (i) ,q) is A (i) The correlation index with the original question q, Flu(A) (i) ) is A (i) The language fluency index, Red(A) (i) ) is A (i) The redundancy is λ1 to λ6, which are the weight coefficients of each indicator.

[0031] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the above-described large-model routing data synthesis method.

[0032] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described large-model routing data synthesis method.

[0033] This invention constructs a user agent that simulates the language style and query preferences of real users, generating diverse questions that more closely resemble actual application scenarios. It utilizes a language model for automatic scoring, significantly reducing manual annotation costs and supporting the automatic synthesis of large-scale, high-quality routing data. During route label generation, this invention comprehensively considers answer quality and user satisfaction, effectively reflecting different users' multidimensional preferences for the model output, thus improving the training effect and accuracy of the routing model. Furthermore, this invention supports the systematic expansion of query styles and question types, significantly enhancing generalization ability under unseen question distributions. It can be flexibly integrated into existing model routing systems, possessing good scalability and practicality, providing strong support for building intelligent and efficient large-scale model routing solutions. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the steps of the large-model routing data synthesis method of the present invention. Detailed Implementation

[0035] In order to more specifically describe the present application, the technical solutions of the present application are described in detail below in combination with the drawings and specific embodiments.

[0036] The embodiment provides a large model routing data synthesis method with diversified expression styles, aims to solve the problems of insufficient training data quality and diversity in the existing large model routing method, and can automatically synthesize a routing data set covering multi-expression styles, multi-question types and multi-model performance differences through steps such as constructing a user agent, synthesizing a style-diverse query, candidate model reasoning and routing label generation, thereby significantly improving the generalization ability and practicality of the routing model in a real open domain question answering scene. The specific implementation steps are as shown in Figure 1

[0037] Step S11: Obtain a seed user question data set.

[0038] This step is the starting point of the entire data synthesis process, and the embodiment collects questions raised by real users from public data sets to construct an initial seed question set. The data set is the basis for expression style expansion and diversity generation in subsequent steps, and can be used as an input example or guiding paradigm for generating expression variants, and has coverage and real context. In some embodiments, these public data sets can come from question and answer communities, forums, social media, search engine logs, etc., aiming to cover the diversity of real user queries as much as possible. It is inclined to select data sets containing a wide range of topics and different language styles to ensure the richness of the initial seed set, for example, it can select questions of different types such as general knowledge question and answer, professional field consultation, daily conversation, etc., thereby laying a solid foundation for subsequent style diversification generation.

[0039] Step S12: Construct a user agent by sampling user features including language expression style, interest preference, and question focus to form a personalized user portrait.

[0040] This step is one of the cores of the present application, aiming to simulate the language expression style, interest preference and question focus of different users to generate a personalized user portrait, and to construct a corresponding user agent accordingly. Specifically,

[0041] S21: Extract representative user features from public real user question data, including but not limited to basic demographic characteristics; through clustering and feature extraction methods, ensure that the extracted features are representative and diverse, thereby effectively supporting a wide range of user distribution simulation, and laying a foundation for subsequent expression style diversification construction.

[0042] ​S22: Based on linguistic theory, further construct expression style control templates to support user agents in generating diverse query questions. Specifically, combining the six language function dimensions in Biber's multidimensional analysis method, systematically organize the 67 language style features under them, and establish a structured mapping system between function dimensions and language implementation forms. This system provides theoretical support for the design of style control prompts.

[0043] Based on this structure, a path sampling strategy is used to select representative and distinct combinations of expression styles to form diverse generation control paths. During the sampling process, dimensional distribution features and style coverage indicators are combined to ensure that the sampling results achieve a balance in terms of expression differences and coverage breadth. Subsequently, the language function dimensions involved in the sampling paths are transformed into descriptive prompt word templates that can be recognized by the large language model, which are used to guide the model to generate query text with specific expression styles.

[0044] Ultimately, these style control prompt word templates are combined with the user structured features constructed in step S21 to form complete control conditions, driving the generation stage to achieve highly controllable and highly diverse styled query generation.

[0045] S23: Based on the user feature pool constructed above, several feature combinations are selected using random sampling or strategic sampling methods to form a structured user profile. During the sampling process, temperature parameters can be introduced to adjust diversity, while coverage balance constraints are set to avoid excessive concentration of single-type user feature combinations, ensuring the richness and representativeness of style expression.

[0046] S24: Using structured user profiles as input, a large language model is invoked to generate behavioral descriptions of the user agent, which serve as conditional information for generating subsequent query questions, further enhancing the personalization and stylistic diversity of the generated content.

[0047] Step S13: Utilize a large language model to generate user query questions with diverse expression styles and clear semantic objectives based on the user agent characteristics; the generated questions can be derived from extensions of real user questions or generated by language style guidance from online corpora.

[0048] S31: Input the structured user agent into the predefined prompt word template, and generate a query question that matches its characteristic style through a large language model to achieve effective style transfer.

[0049] S32: If real user questions are used as seeds, the original questions are style-transferred, covering multiple dimensions such as tone, length, grammatical structure, and rhetoric to ensure the diversity of expression styles. In this process, by setting style conversion goals and content constraints, it is ensured that the semantics of the generated questions do not deviate from the original intent.

[0050] S33: If the network corpus is taken as the reference, the text content related to the question topic is obtained through the search engine or the crawling platform, the style features of the text are extracted, including syntax structure, word frequency distribution and rhetoric pattern, etc., and the style label is formed; then based on these labels, the language model is guided to perform style transfer or imitation, so as to generate new questions consistent with the reference style.

[0051] Step S14: input the generated query question into the candidate language model set for reasoning, and obtain the output answer of each model.

[0052] The candidate model set includes a plurality of language models with significant differences in performance and cost. This set can include models of different sizes (such as small, medium and large LLMs), models of different architectures (such as Transformer (transformer model), MoE (expert hybrid model)), models of different training targets or field optimization (such as general-purpose, code-specific, and law-specific), and models of different providers (such as OpenAI GPT (generative pre-training transformer model) series, Google Gemini (Gemini model) series, Meta LLaMA (Meta large language model) series, etc.). The selection of a variety of candidate models is to simulate the selection dilemma of the router in the real scene, so as to generate more representative routing data; record the reasoning time, computing resource consumption and other cost information of each model, so as to comprehensively consider in subsequent routing label generation.

[0053] Step S15: input the generated query question into the high-performance benchmark LLM as a reference standard output, and build a benchmark answer for evaluating the output quality of the candidate model.

[0054] The benchmark LLM usually selects the model with the best known performance, the best stability and the highest output quality in the current market as the benchmark model, for example, it can be the latest version of GPT-4 or Gemini Ultra, etc. It is crucial to select a high-quality benchmark model, because it directly determines the accuracy and reliability of the routing label. The generation process of the benchmark answer usually does not require additional training, but only relies on the reasoning ability of the model itself. In order to ensure the quality of the benchmark answer, a small amount of manual review of the output of the benchmark model or a consensus mechanism by integrating multiple high-performance models can be used to generate it.

[0055] Step S16: based on the semantic consistency, content integrity and expression fluency between the output of the candidate model and the benchmark answer, an automatic evaluation is performed to generate a quality score for each candidate model; at the same time, the "user satisfaction" of the answer is simulated and evaluated by the language model, and the final routing label is generated comprehensively.

[0056] This step is another key step of the present application, which aims to generate the final routing label according to the multi-dimensional evaluation between the output of the candidate model and the benchmark answer and the simulation satisfaction of the user agent, specifically:

[0057] S61: Based on the fitting degree with the benchmark LLM output, the answer of the candidate model is comprehensively evaluated, which covers multiple dimensions, including semantic similarity, information accuracy, information integrity, relevance, language fluency and redundancy, etc. The output answer A i of the candidate model M (i) is subjected to multi-dimensional quality evaluation with the benchmark answer A ref of the benchmark LLM output, and the total quality score is calculated:

[0058] Score i =λ1·Sim(A (i) ,A ref )+λ2·Acc(A (i) ,A ref )+λ3·Comp(A (i) ,A ref )+λ4·Rel(A (i) ,q)+λ5·Flu(A (i) )-λ6·Red(A (i) )

[0059] Wherein: Sim represents semantic similarity, Acc represents information accuracy, Comp represents content integrity, Rel represents relevance to the original question q, Flu represents language fluency, and Red represents redundancy, λ1-λ6 are weight hyperparameters of each index; the higher the score, the closer the candidate output to the benchmark answer, and the better the expression, which is an important index for generating the final routing label.

[0060] In specific implementation, the existing large language model is used to score the text quality unsupervisedly through carefully designed prompt words, without the need to train a new evaluation model. By guiding the LLM to compare the candidate answer and the benchmark answer, the semantic fit degree and expression quality of the answer can be effectively captured, ensuring that the routing label accurately reflects the performance of the candidate model in content quality. In addition, to further enhance the stability and discrimination of the evaluation, a confidence and uncertainty analysis mechanism is introduced, which comprehensively analyzes the consistency of the output of the candidate model in multiple generations, language feature heuristic indicators (such as syntactic complexity, frequency of ambiguous words, etc.) and model self-scoring confidence, to assist in judging the credibility of the model output, avoiding misleading labels caused by unstable or semantically ambiguous answers.

[0061] S62: Simulate user satisfaction evaluation combined with personalized features of user portrait. According to the role portrait constructed by the user agent, the large language model generates simulated user feedback, focusing on evaluating the performance of the answer in terms of problem solving degree, information satisfaction degree, easy understanding and overall satisfaction degree, etc. The process relies on the generation ability driven by prompt words to simulate the needs and preferences of different user groups, avoiding the high cost of artificial annotation, and capturing the comprehensive evaluation of the quality of the answer from the perspective of diversified users. To further improve the accuracy and controllability of user perspective evaluation, an auxiliary scoring mechanism based on PRM is introduced. By constructing the user query, candidate output and reference answer as a comparative input, calling the publicly trained preference scoring model, obtaining the user preference ranking result or scoring signal as a supplementary index of user satisfaction simulation, so as to more truly reflect the potential selection tendency of individualized users in multi-model answers.

[0062] Specifically, the satisfaction degree of the simulated user role agent to the candidate answer is jointly modeled from two perspectives of generative scoring and preference model, and finally the user satisfaction score is calculated:

[0063]

[0064] Wherein: is the simulated satisfaction score generated by the large language model according to the user portrait, is the scoring result based on the preference reward model, and a e [0, 1] is the fusion coefficient of the two scoring signals.

[0065] The higher the score, the more popular the candidate answer is in the current user agent perspective, which can be used as an important basis for evaluating the quality of the response of the personalized model.

[0066] S63: Fuse multi-dimensional information such as content quality evaluation, user satisfaction simulation, uncertainty analysis and preference reward score to generate the final routing quality label; the label comprehensively reflects the semantic fit degree, language quality, user satisfaction and output stability of the answer, supports the training of supervised router model, and realizes more refined multi-model selection. The whole process does not need additional model training, relies on large model inference and prompt word guidance, has good automation and expansion ability, and is suitable for large-scale construction of high-quality and diversified routing data set.

[0067] Step S17: Pack the user query question, candidate language model output answer and routing quality label to generate routing data for subsequent training or evaluation of large model router.

[0068] Each piece of routing data usually includes: (1) query question: a query question generated by the user agent in a variety of styles; (2) user agent portrait: a feature description of the user agent generating the query question; (3) candidate model output: the answer of each candidate model to the query question; (4) benchmark answer: the reference answer of the superior performance model to the query question; (5) routing label: containing the quality score, user satisfaction score, cost information, etc. of each candidate model; these data will be stored in a structured format (such as JSONL (JSON line format), CSV (comma separated data format)) for subsequent machine learning model training.

[0069] The embodiment can realize the complete process of user agent construction, query generation, model reasoning, answer evaluation and routing label generation through a modular framework design. This design makes the system have good maintainability, scalability and flexibility, and each module can be flexibly adjusted and optimized according to actual needs.

[0070] The application has obvious advantages in constructing high-quality and high-generalization routing data sets. It combines user agents that simulate real user language styles and query preferences, and uses large language models for automatic evaluation and label generation, effectively solving the problems of insufficient training data quality and diversity in existing large model routing methods. The application improves the accuracy and practicality of routing labels through multi-dimensional quality evaluation and user satisfaction simulation, and the systematic expansion of query styles and problem types ensures that the routing model still maintains strong generalization ability and robustness when facing complex scenarios that have not been seen before. The finally generated routing data set comprehensively covers multiple expression styles, multiple problem types and multiple model performance differences, making the application have strong universality and practicality in various open domain question answering scenarios.

[0071] The above description of the embodiments is to facilitate those of ordinary skill in the art to understand and apply the present application. Those skilled in the art can easily make various modifications to the above embodiments, and apply the general principles described herein to other embodiments without creative labor. Therefore, the present application is not limited to the above embodiments, and the improvements and modifications made by those skilled in the art to the present application should be within the scope of protection of the present application.

Claims

1. A method for synthesizing large-scale model routing data with diverse expressive styles, characterized in that, Includes the following steps: (1) Obtain the seed user question dataset; (2) Construct a user intelligent agent, which samples user features, including language expression style, interest preferences, and problem focus, from the seed user problem dataset and combines them to form a personalized user profile; (3) Based on the user agent, LLM is used to generate user query questions with diverse expression styles and clear semantic goals. These query questions are either extensions of real user questions or variations of questions with diverse expressions generated by language styles in the network corpus. (4) Input the user's query into the candidate language model set for reasoning, and obtain the output answer of each candidate language model; (5) Input the user query into a high-performance benchmark LLM and use its output as the benchmark answer to evaluate the output quality of the candidate language model; (6) Based on the output answers of the candidate language models and the baseline answers, a routing quality label for each candidate language model is obtained through multi-dimensional information fusion, including content quality assessment, user satisfaction simulation, uncertainty analysis, and preference reward-assisted scoring. Specifically: For content quality assessment, the assessment covers multiple indicators including semantic similarity, information accuracy, content completeness, question relevance, language fluency, and redundancy. It uses LLM to perform unsupervised text quality scoring on these indicators through prompt words and integrates them to generate a total quality score for the output answer. For uncertainty analysis, the credibility of the candidate language model output is judged by comprehensively analyzing the consistency of the candidate language model output in multiple generation, the heuristic indicators of language features, and the confidence of the model's self-score. For user satisfaction simulation, LLM is used to simulate and generate user satisfaction feedback based on the user profile built by the user agent. The evaluation focuses on the performance of the candidate language model output answer in subjective experience dimensions, including problem-solving, information satisfaction, comprehensibility, and overall satisfaction. This process relies on the cue word-driven generation capability and uses LLM to simulate the needs and preferences of different user groups. For preference reward-assisted scoring, the user query question, the candidate language model output answer, and the benchmark answer are constructed as comparison inputs, and the publicly trained PRM is called to output the user preference ranking results and corresponding scores. Finally, the quantitative indicators of the above four aspects are integrated to generate routing quality labels for candidate language models; (7) Package the user query question, the user profile constructed by the user agent, the model output answer, the baseline answer and the route quality label to generate route data as candidate language models for subsequent model training or large model route evaluation.

2. The method for synthesizing large-scale model routing data according to claim 1, characterized in that: In step (1), questions raised by real users are collected from public datasets to build a seed user question dataset, ensuring that the questions in the dataset are diverse in style and broad in topic.

3. The method for synthesizing large-scale routing data according to claim 1, characterized in that, The specific implementation method of step (2) is as follows: S21: Extract representative user structured features from the seed user problem dataset using clustering and feature extraction methods to ensure that the extracted features are both representative and reflect differences; S22: Utilizing the six language function dimensions in Biber's multidimensional analysis method, a structured mapping system between function dimensions and language implementation forms is established. Based on this system, a path sampling strategy is used to select representative and significantly different combinations of expression styles to form diverse generation control paths. During the sampling process, dimension distribution characteristics and style coverage indicators are combined to ensure that the sampling results achieve a balance in expression diversity and coverage breadth. Then, the language function dimensions involved in the path are transformed into descriptive prompt word templates that LLM can recognize, which are used to guide LLM to generate query text with specific expression styles. S23: Based on the extracted user structured features, several feature combinations are selected using random sampling or strategic sampling methods to form a structured user profile; during the sampling process, a temperature parameter is introduced to adjust the diversity, and a coverage balance constraint is set to avoid excessive concentration of a single type of user feature combination; S24: Using structured user profiles as input, call LLM to generate behavioral descriptions of the user agent, which will then be used as conditional information for generating subsequent query questions.

4. The method for synthesizing large-scale model routing data according to claim 1, characterized in that, The specific implementation of step (3) is as follows: the user profile constructed by the user agent is input into the predefined descriptive prompt word template, and the query question that conforms to its characteristic style is generated by LLM; if the real user question is used as the seed, the original question is style-transferred, covering multiple dimensions including tone, length, grammatical structure, and rhetoric, to ensure the diversity of expression style. In this process, by setting style conversion goals and content constraints, it is ensured that the semantics of the generated question does not deviate from the original intention; if the network corpus is used as a reference, the text content related to the question topic is obtained through a search engine or crawling platform, and the style features of the text, including syntactic structure, word frequency distribution and rhetoric pattern, are extracted to form style tags. Then, based on these tags, LLM is guided to perform style transfer or imitation, thereby generating user query questions that are consistent with the reference style.

5. The method for synthesizing large-scale model routing data according to claim 1, characterized in that, The candidate language model set in step (4) includes a variety of language models with different scales, architectures, training objectives, domain optimizations, and providers; the benchmark LLM in step (5) is the best-performing large language model currently available on the market.

6. The method for synthesizing large-scale routing data according to claim 1, characterized in that: For content quality assessment, the total quality score of the model's output answer is calculated using the following expression: Score i =λ1·Sim(A (i) ,A ref )+λ2·Acc(A (i) ,A ref )+λ3·Comp(A (i) ,A ref ) +λ4·Rel(A (i) ,q)+λ5·Flu(A (i) )-λ6·Red(A (i) ) Where: for any candidate language model M i A (i) For M i The output answer is A. ref As the benchmark answer, Score i For A (i) The overall quality score, Sim(A) (i) A ref ) is A (i) With A ref Semantic similarity, Acc(A (i) A ref ) is A (i) Compared to A ref Information accuracy, Comp(A) (i) A ref ) is A (i) Compared to A ref Content completeness, Rel(A (i) ,q) is A (i) The correlation index with the original question q, Flu(A) (i) ) is A (i) Language fluency index, Red(A) (i) ) is A (i) The redundancy is λ1 to λ6, which are the weight coefficients of each indicator.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor is used to execute the computer program to implement the large model routing data synthesis method as described in any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the large-model routing data synthesis method as described in any one of claims 1 to 6.

Images