Matching method of business scene and fine tuning data set in vertical field
By constructing a four-layer technology matching framework, the problem of matching fine-tuning datasets with business scenarios in vertical fields is solved, realizing the accuracy and usability of datasets, reducing the threshold for enterprise AI applications, and ensuring the effective utilization of resources.
Patent Information
- Application Number
- CN202510974613.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-15
- Publication Date
- 2025-11-14
AI Technical Summary
In vertical industries, the lack of systematic methodologies in existing technologies leads to insufficient matching between the construction of fine-tuning datasets and business scenarios, making it difficult to implement AI projects.
By constructing a four-layer technical matching framework of "business scenario characteristics → dataset type → logical reasoning structure → question-answer pair type", suitable fine-tuning dataset types and logical reasoning structures are determined, and question-answer pair types are configured to achieve accurate matching between datasets and business scenarios.
It improves the accuracy and usability of fine-tuning datasets, lowers the threshold for enterprise AI applications, ensures that dataset construction revolves around real business needs, and avoids resource waste and trial-and-error costs.
Smart Images
Figure CN120952151A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method for matching business scenarios in a vertical field with fine-tuning datasets, applicable to the pre-construction stage of fine-tuning datasets. Background Technology
[0002] In recent years, with the rapid development of artificial intelligence (AI) technology, large language models (LLMs, such as DeepSeek and GPT) have demonstrated significant advantages in text understanding, knowledge extraction, and generation tasks. Artificial intelligence (AI) has become a core force driving transformation across various industries, and the key support behind it is high-quality datasets. These datasets are not only the foundation for AI model training, inference, and validation, but also the core driving force for AI to empower industries. The construction of high-quality datasets has become an important tool for AI to empower new industrialization, and a strategic choice for industrial enterprises to reduce costs, increase efficiency, and reshape their competitiveness.
[0003] However, within different vertical industries, the requirements for LLM vary significantly across different business scenarios. For example, industrial scenarios have extremely stringent requirements for decision reliability, even to the point of zero tolerance for model illusions; autonomous driving scenarios have extremely high requirements for real-time interaction, making it crucial to improve the response speed of LLM. Therefore, building high-quality fine-tuning datasets for vertical industries and fine-tuning LLM models requires considering the technical requirements of the business scenario for dataset construction, rather than simply building industry datasets according to uniform quality standards.
[0004] Currently, even if companies strictly adhere to existing high-quality dataset evaluation standards when constructing their own datasets, they generally fall into a systemic dilemma of "AI for AI's sake, data for data's sake" when promoting the implementation of AI projects. The reality is: "The government and industry are unclear about what kind of data is needed for high-quality industry development; industry enterprises do not know how to construct high-quality datasets; and the industry does not understand how to evaluate the quality of industry datasets (source: expert opinions from the National Bureau of Data Science and Technology website)."
[0005] Through in-depth research, the inventors discovered that the root cause of the systemic dilemma lies in the conflict between the differences in business scenarios in vertical domains and the standardization of AI training paradigms. The lack of a systematic matching methodology for fine-tuning datasets in advance leads to a lack of strong logical connection between the construction of fine-tuning datasets, the selection of logical reasoning structures, and the design of question-answer pairs and the business scenarios during the process.
[0006] Therefore, this invention focuses on the pre-construction stage of building high-quality datasets. Starting from real business scenarios, it provides enterprises with a simple, effective, and feasible method to clearly determine the following questions: What kind of dataset is needed? Are there the conditions to do so? What are the expected model results? Furthermore, it plays a normative role in the construction of high-quality datasets during and after the construction process, enabling enterprises to always stay on the right track when implementing AI applications in vertical fields. Summary of the Invention
[0007] 1. The technical problem solved by the present invention
[0008] In view of the above situation, this invention focuses on the pre-production stage of high-quality datasets in vertical fields, and constructs a four-layer technical matching relationship from "business scenario characteristics → dataset type → logical reasoning structure → question-answer pair type", forming a complete construction logic for fine-tuning datasets. Thus, it provides enterprises with a simple, effective and feasible method to achieve how to match the most suitable fine-tuned dataset for business scenarios.
[0009] 2. Technical solution of the present invention
[0010] To achieve the technical problem addressed by this invention, the present invention provides a method for matching business scenarios in a vertical domain with fine-tuned datasets, characterized by comprising:
[0011] Step S1: Determine the characteristics of the business scenario in the relevant vertical industry;
[0012] Step S2: Select the fine-tuning dataset type based on the characteristics of the business scenario;
[0013] Step S3: Match the logical reasoning structure according to the fine-tuning dataset type;
[0014] Step S4: Configure the question-answer pair type according to the logical reasoning structure.
[0015] Furthermore, the method for matching vertical domain business scenarios with fine-tuned datasets is characterized by:
[0016] In step S1, the business scenario characteristics include at least one of high technical understanding, high real-time interaction, strong logical decision-making, hybrid scenarios, and full-process optimization; the hybrid scenario refers to the simultaneous inclusion of real-time control and decision optimization.
[0017] In step S2, the fine-tuning dataset type includes derived datasets, prior datasets, and combinations of derived datasets and prior data;
[0018] In step S3, the logical reasoning structure includes a thought chain, a reasoning chain, and strong and weak mixtures of thought chains and reasoning chains;
[0019] In step S4, the question-and-answer pair types include: declarative question-and-answer pairs without thought chains and argumentative question-and-answer pairs with thought chains.
[0020] Furthermore, the method for matching vertical domain business scenarios with fine-tuned datasets is characterized by:
[0021] In step S2, the derived dataset is a semi-structured data set composed of isolated question-answer pairs generated based on vertical domain expertise; the prior dataset is a structured data set composed of inference chain units as basic elements, starting from real business scenarios; the inference chain unit is a chain structure composed of multiple question-answer pairs that are logically related around the same target object, and each question-answer pair corresponds to a process node in a business scenario.
[0022] Furthermore, the method for matching vertical domain business scenarios with fine-tuned datasets is characterized by:
[0023] In step S3, the thought chain is a logical reasoning process that describes obtaining answers, suitable for isolated argumentative question-answer pairs in the derived dataset; the reasoning chain is a logical reasoning method that organizes data through a chain structure, suitable for logically related question-answer pairs in the prior dataset; the strongly mixed index reasoning chain units all use argumentative question-answer pairs; the weakly mixed index dataset contains both declarative question-answer pairs and argumentative question-answer pairs.
[0024] Furthermore, the method for matching vertical domain business scenarios with fine-tuned datasets is characterized by:
[0025] In step S4, the declarative question-answer pairs focus on conveying existing knowledge or facts and are used to correspond to process nodes that require real-time response scenarios; the argumentative question-answer pairs use credible logical argumentation techniques to correspond to process nodes that require deep reasoning scenarios.
[0026] Optionally, the method for matching the business scenario in the vertical field with the fine-tuned dataset is characterized in that: if the business scenario in step S1 is a high-tech understanding scenario, then a derived dataset is selected in step S2, a thought chain is matched in step S3, and an argumentative question-and-answer pair is configured in step S4.
[0027] Optionally, the method for matching the business scenario in the vertical domain with the fine-tuned dataset is characterized in that: if the business scenario in step S1 is a high real-time interactive scenario, then in step S2, a prior dataset is selected, inference chains are matched in step S3, and declarative question-answer pairs are configured in step S4.
[0028] Optionally, the method for matching the business scenario in the vertical domain with the fine-tuned dataset is characterized in that: if the business scenario in step S1 is a strong logic decision-making scenario, then in step S2, a prior dataset is selected, in step S3, a strong mixture is matched, and in step S4, an argumentative question-answer pair is configured.
[0029] Optionally, the method for matching the business scenario in the vertical domain with the fine-tuned dataset is characterized in that: if the business scenario in step S1 is a mixed scenario, then in step S2, a prior dataset is selected, in step S3, weak mixing is matched, and in step S4, declarative question-answer pairs and argumentative question-answer pairs are configured.
[0030] Optionally, the method for matching the business scenario in the vertical domain with the fine-tuned dataset is characterized in that: if the business scenario in step S1 is a full-process optimization scenario, then in step S2, a combination of the derived dataset and the prior dataset is selected; in step S3, weak mixing is matched; and in step S4, declarative question-answer pairs and argumentative question-answer pairs are configured.
[0031] 3. Beneficial effects of the present invention
[0032] This invention provides a method for matching business scenarios in a vertical industry with fine-tuned datasets, and the beneficial effects it produces are as follows:
[0033] First, it addresses the systemic matching challenges in the implementation of AI in vertical industries. By constructing a four-layer technology matching framework—"business scenario characteristics → dataset type → logical reasoning structure → question-answer pair type"—it establishes a strong correlation between business needs and technical elements from the source, solving the problem of pre-decision-making and ensuring that dataset construction always revolves around real business needs.
[0034] Second, improve the accuracy and usability of fine-tuned datasets. Dynamically match technical elements according to the characteristics of different scenarios. For example, in high real-time interaction scenarios, match "prior dataset + inference chain + declarative question-answer pair". This matching mechanism significantly improves the fit between the dataset and business objectives. For example, the real-time requirements of autonomous driving can achieve millisecond-level response through structured data in the inference chain.
[0035] Third, it lowers the barrier to AI application for enterprises and controls trial and error costs. Pre-determining the construction of fine-tuned datasets can clarify what kind of datasets should be built, assess the feasibility of construction, and predict the model's performance. Therefore, enterprises can control the direction of resource investment in the early stages of data construction and avoid blind development.
[0036] In summary, with the explosive growth in demand for fine-tuning of large models in vertical industries, enterprises are experiencing a surge in demand for "scenario-customized datasets." Providing a standardized method for constructing datasets aligns with the core industry demands for cost reduction and efficiency improvement. Therefore, this invention has significant engineering practical value and broad engineering application prospects. Attached Figure Description
[0037] The present invention will now be described in further detail with reference to the accompanying drawings.
[0038] Figure 1 This is a schematic diagram illustrating the matching process between the vertical domain business scenario and the fine-tuning dataset of the present invention;
[0039] Figure 2 This is a schematic diagram of the basic framework of the inference chain unit with closed-loop feedback mechanism of the present invention. Detailed Implementation
[0040] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings.
[0041] Example 1
[0042] This embodiment combines Figure 1 The matching process between the vertical domain business scenarios and the fine-tuning dataset is explained in detail, including:
[0043] Step S1: Determine the characteristics of the business scenario in the relevant vertical industry; the characteristics of the business scenario include at least one of the following: high technical understanding, high real-time interaction, strong logical decision-making, hybrid scenarios, and full-process optimization.
[0044] High-tech understanding can be interpreted as a field characterized by high technical content and high knowledge density, making it a high-value area suitable for AI implementation, such as the industrial sector. High real-time interaction refers to the requirement that the AI system in this field must respond to external stimuli at the millisecond level, such as in the field of autonomous driving. Strong logic decision-making refers to a field that heavily relies on strong logical reasoning, where even a single mistake can cause irreparable losses, such as in healthcare and finance. Hybrid scenarios refer to scenarios that simultaneously include real-time control and decision optimization; they can be understood as a combination of high real-time performance and strong logic decision-making, but with the technical requirements weakened, such as in intelligent manufacturing, smart cities, and embodied intelligence. End-to-end optimization can be understood as having a long product development cycle and a complex production process, such as in metal materials and biomedicine.
[0045] It should be noted that the five business scenario characteristics analyzed above may not be strictly distinguishable in specific vertical fields. Sometimes it may be found that a field is suitable for several different business scenario characteristics. In this case, it is recommended to adopt the business scenario characteristics with the highest requirements. At the same time, it is not ruled out that there may be other business scenario characteristics. However, in any case, a suitable dataset construction scheme can be found through the "four-layer technology matching framework" of this invention.
[0046] Step S2: Select the fine-tuning dataset type according to the characteristics of the business scenario; the fine-tuning dataset type includes derived datasets, prior datasets, and combinations of derived datasets and prior data; the derived dataset is a semi-structured data set composed of isolated question-answer pairs generated based on vertical domain expertise; the prior dataset is a structured data set composed of inference chain units as basic elements, starting from real business scenarios; the inference chain unit is a chain structure composed of multiple question-answer pairs that are logically related around the same target object, and each question-answer pair corresponds to a process node in a business scenario.
[0047] Derivative dataset: Targeting mature industrial sectors, it uses patent documents as the core data source and introduces CoT and traceability codes into the question-answer pairs. It builds connections through industrial knowledge graphs, but the question-answer pairs are essentially isolated from each other.
[0048] Prior datasets: These are vertical domains that require logical reasoning in real business scenarios, such as industry, healthcare, finance, and autonomous driving. By introducing CoR, isolated but related question-answer pairs are connected into a chain structure. Feedback mechanisms can be introduced to generate reasoning chain units and gradually build more complex reasoning chain systems, ultimately forming a reasoning chain network.
[0049] Inference chain units are fundamental elements of the prior dataset. Figure 2 It is the basic framework of the reasoning chain unit with a closed-loop feedback mechanism. In addition to the feedback node, other nodes can be split into multiple functionally independent sub-nodes that are logically connected in sequence, or merged into composite nodes. The question-answer pair adopts the declarative or argumentative type, which should be determined according to the specific characteristics of the vertical domain. The closed-loop feedback mechanism includes: global feedback and local feedback, positive feedback and negative feedback, and pre-feedback and post-feedback.
[0050] The core idea of prior datasets is how to internalize "algorithms" into "data". The specific steps are: first, domain experts build the top-level logical framework of the prior dataset; then, question-answer pairs for specific application scenarios are injected according to the framework nodes; finally, a complete inference chain system is formed.
[0051] The process nodes are divided into real-time execution nodes and logical decision nodes according to their functions. Real-time execution nodes correspond to scenarios that require real-time response and are matched with declarative question-and-answer pairs, while logical decision nodes correspond to scenarios that require in-depth reasoning and are matched with argumentative question-and-answer pairs.
[0052] Step S3: Match logical reasoning structures according to the fine-tuning dataset type; the logical reasoning structures include chain-of-thought (CoT), chain-of-reasoning (CoR), and strong and weak hybrids of chain-of-thought and chain-of-reasoning. A technical comparison of the four logical reasoning structures is shown in Table 1; the chain-of-thought is a logical reasoning process that describes the process of obtaining answers, suitable for isolated argumentative question-and-answer pairs in the derived dataset; the chain-of-reasoning is a logical reasoning method that organizes data through a chain structure, suitable for logically related question-and-answer pairs in the prior dataset; strong hybridization means that argumentative question-and-answer pairs are used in all units of the chain-of-reasoning; weak hybridization means that the dataset contains both statement-type question-and-answer pairs without CoT and argumentative question-and-answer pairs with CoT.
[0053] Table 1. Technical Comparison of Four Logical Reasoning Structures
[0054]
[0055] Step S4: Configure the question-answer pair type according to the logical reasoning structure; the question-answer pair type includes declarative question-answer pairs and argumentative question-answer pairs; the declarative question-answer pairs focus on conveying existing knowledge or facts and are used to correspond to process nodes (real-time execution nodes) that need to perform real-time response scenarios; the argumentative question-answer pairs use credible logical argumentation techniques and are used to correspond to process nodes (logical decision nodes) that need to perform deep reasoning scenarios.
[0056] Example 2
[0057] This embodiment details its specific applications in various vertical industries.
[0058] In vertical domains, the construction of fine-tuning datasets should be adapted to the characteristics of business scenarios in order to maximize the potential of fine-tuning and optimizing LLM. Based on the types of business scenario characteristics listed in step S1 of Example 1, there are five matching types.
[0059] Type A: If the business scenario in step S1 is a high-tech understanding scenario, then in step S2, a derived dataset is selected, in step S3, a thought chain is matched, and in step S4, an argumentative question-and-answer pair is configured.
[0060] Some industrial sectors, characterized by high technological content and high knowledge density, are high-value areas suitable for AI implementation. It is particularly emphasized that the core source of derived datasets is patent literature; other sources can serve as valuable supplements. Because these sources are not based on real-world business scenarios, it is impossible to construct a Question-and-Answer (CoR) relationship. Even if based on real-world business scenarios, they are treated as isolated question-and-answer pairs. The core difference between Type A and current high-quality dataset construction lies in the different data sources. Currently, the engineering community has overlooked the technical value of patent literature in constructing fine-tuning datasets.
[0061] Type B: If the business scenario in step S1 is a high real-time interactive scenario, then in step S2, a prior dataset is selected, inference chains are matched in step S3, and declarative question-answer pairs are configured in step S4.
[0062] Autonomous driving is a typical high real-time interaction scenario, requiring AI systems to respond to external stimuli in milliseconds. Since it deals with real-time delivery scenarios, the CoR (Cooperation of Reasoning) is rigid. Therefore, we prioritize configuring declarative question-answer pairs and generating them as inference chain units to construct a prior dataset.
[0063] Type C: If the business scenario in step S1 is a strong logic decision-making scenario, then select the prior dataset in step S2, match strong hybrid in step S3, and configure argumentative question-answer pairs in step S4.
[0064] Healthcare and finance rely heavily on strong logical reasoning, where even a single mistake can lead to irreparable losses. To ensure strong logical reasoning properties, all question-answer pairs adopt argumentative question-answer pairs, CoT is introduced, and reasoning chain units are generated. CoR is introduced to construct a prior dataset.
[0065] Type D: If the business scenario characteristics in step S1 are a mixed scenario, then in step S2, a priori dataset is selected, in step S3, weak mixing is matched, and in step S4, declarative question-answer pairs and argumentative question-answer pairs are configured; the real-time execution nodes on the inference chain unit are configured only with declarative question-answer pairs, and the decision nodes are configured only with argumentative question-answer pairs.
[0066] Smart cities and embodied intelligence combine real-time control (using CoR + declarative question-answer pairs to ensure response speed) and decision optimization (using CoT + argumentative question-answer pairs to ensure reliability), which is a typical hybrid scenario. It can be understood as a combination of high real-time performance and strong logical decision-making, but with the technical requirements being weakened.
[0067] Type E: If the business scenario in step S1 is a full-process optimization scenario, then in step S2, a combination of derived dataset and prior dataset is selected, in step S3, weak mixing is matched, and in step S4, declarative question-answer pairs and argumentative question-answer pairs are configured.
[0068] Product development in metallic materials and biomedicine is a typical example of end-to-end optimization, divided into product development and production stages. It has a long product development cycle and a complex production process, involving "intelligent R&D + intelligent manufacturing", which is more like a coupling of type A and type C.
[0069] The above five types are summarized in Table 2 as the scene matching types of vertical domain fine-tuning datasets. The core principle is: derived datasets focus on technical credibility, and prior datasets strengthen the closed-loop feedback mechanism; argumentative question-answer pairs (including CoT) are adapted to deep reasoning needs, and declarative question-answer pairs (excluding CoT) meet the needs of efficient execution scenarios.
[0070] Table 2 Scene Matching Types for Vertical Domain Fine-tuning Datasets
[0071]
[0072] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Various modifications or equivalent substitutions made by those skilled in the art within the substantial scope of protection of the present invention also fall within the scope of protection of the present invention.
Claims
1. A method for matching business scenarios in a vertical domain with fine-tuned datasets, characterized in that, include: Step S1: Determine the characteristics of the business scenario in the relevant vertical industry; Step S2: Select the fine-tuning dataset type based on the characteristics of the business scenario; Step S3: Match the logical reasoning structure according to the fine-tuning dataset type; Step S4: Configure the question-answer pair type according to the logical reasoning structure.
2. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 1, characterized in that: In step S1, the business scenario characteristics include at least one of high technical understanding, high real-time interaction, strong logical decision-making, hybrid scenarios, and full-process optimization; the hybrid scenario refers to the simultaneous inclusion of real-time control and decision optimization. In step S2, the fine-tuning dataset type includes derived datasets, prior datasets, and combinations of derived datasets and prior data; In step S3, the logical reasoning structure includes a thought chain, a reasoning chain, and strong and weak mixtures of thought chains and reasoning chains; In step S4, the question-and-answer pair types include: declarative question-and-answer pairs without thought chains and argumentative question-and-answer pairs with thought chains.
3. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 2, characterized in that: In step S2, the derived dataset is a semi-structured data set composed of isolated question-answer pairs generated based on vertical domain expertise; the prior dataset is a structured data set composed of inference chain units as basic elements, starting from real business scenarios; the inference chain unit is a chain structure composed of multiple question-answer pairs that are logically related around the same target object, and each question-answer pair corresponds to a process node in a business scenario.
4. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 3, characterized in that: In step S3, the thought chain is a logical reasoning process that describes obtaining answers, suitable for isolated argumentative question-answer pairs in the derived dataset; the reasoning chain is a logical reasoning method that organizes data through a chain structure, suitable for logically related question-answer pairs in the prior dataset; the strongly mixed index reasoning chain units all use argumentative question-answer pairs; the weakly mixed index dataset contains both declarative question-answer pairs and argumentative question-answer pairs.
5. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 4, characterized in that: In step S4, the declarative question-answer pairs focus on conveying existing knowledge or facts and are used to correspond to process nodes that require real-time response scenarios; the argumentative question-answer pairs use credible logical argumentation techniques to correspond to process nodes that require deep reasoning scenarios.
6. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 5, characterized in that: If the business scenario described in step S1 is a high-tech understanding scenario, then in step S2, a derived dataset is selected, in step S3, a thought chain is matched, and in step S4, an argumentative question-and-answer pair is configured.
7. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 5, characterized in that: If the business scenario described in step S1 is a high real-time interactive scenario, then in step S2, a prior dataset is selected, inference chains are matched in step S3, and declarative question-answer pairs are configured in step S4.
8. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 5, characterized in that: If the business scenario described in step S1 is a strong logic decision-making scenario, then in step S2, a prior dataset is selected, in step S3, a strong mixture is matched, and in step S4, an argumentative question-answer pair is configured.
9. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 5, characterized in that: If the business scenario characteristics described in step S1 are a hybrid scenario, then in step S2, a prior dataset is selected, in step S3, weak hybridity is matched, and in step S4, declarative question-answer pairs and argumentative question-answer pairs are configured.
10. The method for matching vertical domain business scenarios with fine-tuned datasets according to claim 5, characterized in that: If the business scenario characteristics described in step S1 are a full-process optimization scenario, then in step S2, a combination of derived dataset and prior dataset is selected, in step S3, weak mixing is matched, and in step S4, declarative question-answer pairs and argumentative question-answer pairs are configured.