Data analysis method and device based on intelligent data analysis big model

Abstract

The present application provides a data analysis method and device based on a large model of intelligent data analysis. The method includes: obtaining real data from preset data elements, constructing and preprocessing into a first data set and a second data set; based on the second data set, multi-level pre-training autoregressive language model to obtain a first target model; generating a second target model through fine-tuning instructions; normalizing and adjusting after performance evaluation meets the standards to form a third target model; optimizing according to user feedback to obtain a fourth target model; inputting the data to be analyzed to obtain the analysis results. The present application ensures that the model can flexibly adapt to different scenario requirements while having generalization capabilities through multi-level pre-training and fine-tuning for specific scenarios, solving the problems of poor migration capabilities, low flexibility and scalability of data analysis models in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of data analysis, and in particular to a data analysis method based on an intelligent data analysis big model, a data analysis device based on an intelligent data analysis big model, a computer-readable storage medium, and a data analysis system. Background Art

[0002] Data analysis is an important field in the current development of science and technology, involving a wide range of application scenarios such as economics, finance, statistics, and social sciences. Traditional data analysis methods mainly rely on manually designed features and rules, which have limitations and are difficult to effectively cope with complex and changing data environments. In recent years, machine learning and deep learning technologies have been increasingly used in data analysis, which can automatically extract the inherent laws and associations of data and improve the accuracy and efficiency of analysis.

[0003] Most existing data analysis models are limited to binary classification predictions, and are unable to segment and identify users' various risk states. They are difficult to meet the complex needs of actual business scenarios, have certain accuracy and interpretability issues, and are difficult to be trusted and accepted by business personnel. In addition, the training process of traditional data analysis models relies too much on manually designed feature engineering, and lacks effective use of users' dynamic change information in time series, resulting in limited model performance. When dealing with large-scale and diverse data analysis tasks, it usually requires a lot of manpower and computing resources, and the technical threshold is high, which limits its widespread application.

[0004] Among them, data analysis models based on neural networks are an important technical path. Common models include: 1. Supervised learning models, such as logistic regression, decision trees, random forests, etc., which can be used for classification and prediction tasks. 2. Unsupervised learning models, such as K-Means clustering, are good at discovering the intrinsic structure of data. 3. Deep learning models, such as convolutional neural networks, recurrent neural networks, etc., can analyze unstructured data such as text and images. These models have shown good performance in specific fields and tasks, but the following problems are common: 1. The generalization ability and interpretability of the model are weak, and it is difficult to cope with complex and changeable actual scenarios. 2. Model training usually relies on a large amount of labeled data, and data preparation is arduous. 3. When learning new tasks, the model has poor knowledge transfer ability and needs to be retrained.

[0005] Therefore, the data analysis model in the prior art indicates poor migration capability, low flexibility and low scalability. Summary of the invention

[0006] The main purpose of this application is to provide a data analysis method based on an intelligent data analysis big model, a data analysis device based on an intelligent data analysis big model, a computer-readable storage medium and a data analysis system, so as to at least solve the problem in the prior art that the data analysis model indicates poor migration capability, low flexibility and scalability.

[0007] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a data analysis method based on an intelligent data analysis large model is provided, including: obtaining corresponding real data from multiple preset data elements to obtain a first data set, preprocessing the first data set to obtain a second data set, the real data includes data analysis and statistical reports, data analysis and statistical cases, and data analysis and statistical tutorials, and the preprocessing includes at least data enhancement, data cleaning, and data mixing; based on the second data set, multi-level pre-training of an autoregressive language model is performed to obtain a first target model, and the multi-level pre-training includes text level pre-training, sentence / paragraph level pre-training, and long text level pre-training. The method comprises the following steps: performing model instruction fine-tuning based on the first target model to obtain a second target model, wherein the model instruction fine-tuning includes question-answer pair training, instruction-response pair training and data set-analysis thinking chain-analysis action pair training; performing model performance evaluation based on the second target model, and when the evaluation score is greater than or equal to a preset value, performing normalization adjustment on the second target model to obtain a third target model, wherein the normalization adjustment is used to unify the model's response to the same question; optimizing the third target model at least based on user feedback behavior to obtain a fourth target model; inputting the data to be analyzed into the fourth target model to obtain data analysis results.

[0008] Optionally, the first data set is preprocessed to obtain a second data set, including: based on the report and the tutorial, analyzing the first case containing a causal relationship in the case, extracting dependent variables, causal paths and reasoning processes for annotation, and generating a causal link based on the dependent variables, causal paths and reasoning processes; based on the report and the tutorial, analyzing the second case containing the use of a data analysis tool in the case, extracting data analysis steps, tool types of the data analysis tools, usage steps of the data analysis tools and expected output results, and generating usage scenarios of the data analysis tools corresponding to different data types based on the data analysis steps, the tool types, the usage steps and the expected output results; analyzing the report, the tutorial and the case to obtain a causal relationship. Take the data formats in the report, the tutorial and the case to obtain multiple first data formats, construct a mapping relationship between each of the first data formats and the corresponding second data format to obtain a first mapping relationship, where the second data format is a data format when data of different data types are used in the first target model; perform data cleaning on the report, the tutorial and the case to obtain first target data, wherein the data cleaning includes deleting duplicate data, supplementing missing values, smoothing outliers, deleting data of preset types, and converting data formats according to the first mapping relationship; randomly mix the first target data according to different languages ​​to obtain second target data; and construct the second data set based on the causal link, the usage scenario, the first mapping relationship and the second target data.

[0009] Optionally, the autoregressive language model is pre-trained at multiple levels based on the second data set to obtain a first target model, including: constructing a data analysis model based on the autoregressive model, the masked language model and the GPT model, and decomposing the second data set, performing the text level pre-training with words as basic units, and obtaining a fifth target model, wherein the autoregressive model is P(x i |x1,…,x i-1 )When the previous text is input, the conditional probability of the i-th word, P(X) is the output predicted word, and the masked language model is L MLM =-∑ i∈M logP(x i |x1,…,x i-1 ,x i+1 ,…,x n ), M is a randomly selected set of mask positions, L MLM To fill in the text, the GPT model is L is the negative log-likelihood loss; based on the decomposition of the second data set, the fifth target model is pre-trained at the sentence / paragraph level with sentences and paragraphs as basic units, so as to enable the model to generate subsequent steps or summary representations based on the previous text for data analysis, and obtain the sixth target model; based on the second data set, synonym replacement and random masking are performed to obtain the third data set, and the sixth target model is pre-trained at the long text level with the third data set as the basic unit to obtain the seventh target model, wherein the training is stopped when the fluctuation of the model score based on the validation set is less than the first threshold during the model training process; based on the causal link and the usage scenario in the second data set, the sixth target model is pre-trained at the thinking chain level to obtain the first target model, wherein each step in the thinking chain generated by the first target model includes the operation type, the causal link, the tool call and the expected response.

[0010] Optionally, fine-tuning the model instructions is performed based on the first target model to obtain the second target model, including: obtaining a plurality of the question-answer pairs, the question-answer pairs including the analysis questions input by the user into the first target model and the answers generated by the first target model; obtaining a plurality of the instruction-response pairs, the instruction-response pairs including the instructions input by the user into the first target model and the execution steps generated by the first target model; obtaining a plurality of the data set-analysis thinking chain-analysis action pairs, the data set-analysis thinking chain-analysis action pairs including the data type of the input data input by the user into the first target model and the thinking chain, analysis steps and expected analysis results of the first target model for analyzing the input data; based on the question-answer pairs, the instruction-response pairs and the data set-analysis thinking chain-analysis action pairs, the first target model is trained based on the AdamW optimizer to obtain the second target model.

[0011] Optionally, performing model performance evaluation based on the second target model includes: obtaining a standard test set, inputting the standard test set into the second target model, and obtaining a first output result; determining an accuracy rate based on comparing the first output result with a benchmark output result corresponding to the standard test set, and determining the accuracy rate as a first evaluation score; obtaining a second evaluation score, and obtaining the evaluation score by performing a weighted calculation based on the first evaluation score and the second evaluation score, wherein the second evaluation score is a score input by a user.

[0012] Optionally, the second target model is normalized and adjusted to obtain a third target model, including: inputting the second data set into the second target model to obtain multiple second output results; classifying the second output results based on the sample type of the samples in the second data set to obtain multiple fourth data sets; counting the frequency of occurrence of each of the second output results in each fourth data set, determining the second output result with the highest corresponding frequency as the response of the corresponding sample type, to obtain the third target model.

[0013] Optionally, the third target model is optimized based on user feedback behavior to obtain a fourth target model, including: determining a user feedback score based on the user feedback behavior, and determining the user feedback score as direct satisfaction; determining the feedback time and interaction frequency for the user to complete the user feedback behavior based on the user feedback behavior, and determining the feedback time and the interaction frequency as indirect satisfaction; performing weighted calculation based on the direct satisfaction and the indirect satisfaction to obtain a satisfaction score; when the satisfaction score is greater than or equal to a second threshold, applying a positive reward signal to the third target model to obtain the fourth target model; when the satisfaction score is less than or equal to the third threshold, applying a negative reward signal to the third target model to obtain the fourth target model; when the satisfaction score is greater than the third threshold and less than the second threshold, determining the third target model as the fourth target model.

[0014] According to another aspect of the present application, a data analysis device based on an intelligent data analysis large model is provided, the device comprising: a first processing unit, used to obtain corresponding real data from a plurality of preset data elements to obtain a first data set, preprocess the first data set to obtain a second data set, the real data comprising data analysis and statistical reports, data analysis and statistical cases and data analysis and statistical tutorials, the preprocessing comprising at least data enhancement, data cleaning and data mixing; a first training unit, used to perform multi-level pre-training on an autoregressive language model based on the second data set to obtain a first target model, the multi-level pre-training comprising text level pre-training, sentence / paragraph level pre-training, long text level pre-training and thought chain level pre-training Pre-training; a second processing unit, used to fine-tune the model instructions based on the first target model to obtain a second target model, wherein the model instruction fine-tuning includes question-answer pair training, instruction-response pair training, and data set-analysis thinking chain-analysis action pair training; a second training unit, used to evaluate the model performance based on the second target model, and when the evaluation score is greater than or equal to a preset value, normalize the second target model to obtain a third target model, and the normalization adjustment is used to unify the model's response to the same question; a third training unit, used to optimize the third target model based at least on user feedback behavior to obtain a fourth target model; an input unit, used to input the data to be analyzed into the fourth target model to obtain data analysis results.

[0015] According to another aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute any one of the methods described.

[0016] According to another aspect of the present application, a data analysis system is provided, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs include methods for executing any one of the methods described.

[0017] Applying the technical solution of the present application, corresponding real data is obtained from multiple preset data elements to obtain a first data set, and the first data set is preprocessed to obtain a second data set; based on the second data set, the autoregressive language model is pre-trained at multiple levels to obtain a first target model; based on the first target model, the model instruction is fine-tuned to obtain a second target model; based on the second target model, the model performance is evaluated, and when the evaluation score is greater than or equal to the preset value, the second target model is normalized and adjusted to obtain a third target model, and the normalization adjustment is used to unify the model's response to the same problem; the third target model is optimized at least based on user feedback behavior to obtain a fourth target model; the data to be analyzed is input into the fourth target model to obtain data analysis results. The present application proposes a large intelligent data analysis model based on multi-level pre-training and fine-tuning for specific scenarios. Model training is based on multiple data sources to ensure the generalization of the model. When the scene is changed, unsupervised training and limited sample fine-tuning can be used to adapt to the scene, ensuring the flexibility and scalability of the system, solving the problem of poor migration ability, low flexibility and scalability of data analysis models in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A hardware structure block diagram of a mobile terminal for data analysis based on an intelligent data analysis big model provided in an embodiment of the present application is shown;

[0019] Figure 2 A schematic diagram of a process of a data analysis method based on an intelligent data analysis big model provided according to an embodiment of the present application is shown;

[0020] Figure 3 A structural block diagram of a data analysis device based on an intelligent data analysis big model provided according to an embodiment of the present application is shown.

[0021] The above drawings include the following reference numerals:

[0022] 102, processor; 104, memory; 106, transmission device; 108, input and output devices. DETAILED DESCRIPTION

[0023] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0024] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0025] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] For the convenience of description, some nouns or terms involved in the embodiments of the present application are explained below:

[0027] Large Language Model Pre-training: Large language model pre-training is the initial stage of large language model learning. It lays the foundation for the model through pre-training of large-scale corpora. In this process, the model can acquire basic language understanding and generation skills. Pre-trained language models usually adopt a "pre-training-fine-tuning" training process, first learning common language patterns in an unsupervised environment, and then fine-tuning for specific tasks in a supervised environment to improve the performance of the model in specific tasks. This process ultimately forms a model that can be deployed and applied, enabling it to demonstrate strong language processing capabilities in practical applications.

[0028] Instruction Tuning: Instruction tuning is an effective technique to improve the capabilities and controllability of large language models. It involves further training a large language model using (instructions, outputs), where instructions represent manual instructions to the model and outputs represent the expected outputs corresponding to the instructions. In this way, the model is better able to understand and execute specific instructions, thereby improving its accuracy and controllability in performing tasks. Instruction tuning enables the model to better follow human design intent and generate expected outputs when faced with complex or specific tasks.

[0029] Grouped-Query Attention (GQA): Grouped-Query Attention is a technique for large language models that improves on the self-attention mechanism. GQA combines the advantages of multi-query and multi-head attention to speed up inference time by grouping multiple queries and jointly computing their attention. Specifically, GQA groups multiple queries, with queries within each group sharing the same set of weights, and then jointly computes attention for queries within each group. This improvement not only maintains the performance of the model, but also significantly reduces the amount of computation during inference, making it more efficient in the application of large-scale language models.

[0030] Reinforcement Learning from Human Feedback (RLHF): Reinforcement Learning from Human Feedback is a reinforcement learning technique based on human preferences. It trains a "reward model" directly based on human feedback, uses the model as the reward function in reinforcement learning, and then optimizes the model's strategy through algorithms such as proximal policy optimization (PPO). This approach enables the model to better understand and follow human preferences, making it more intelligent and humane in decision-making and interaction. Through RLHF, the model can not only learn from data, but also get guidance from human feedback, further improving its performance and adaptability in complex tasks.

[0031] As introduced in the background technology, the training process of traditional data analysis models in the prior art relies too much on manually designed feature engineering, and lacks effective use of the user's dynamic change information in the time series, resulting in limited model performance. In order to solve the problems that the data analysis models in the prior art have poor indication migration ability, low flexibility and scalability, the embodiments of the present application provide a data analysis method based on an intelligent data analysis big model, a data analysis device based on an intelligent data analysis big model, a computer-readable storage medium and a data analysis system.

[0032] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0033] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 1 is a hardware structure block diagram of a mobile terminal of a data analysis method based on an intelligent data analysis large model according to an embodiment of the present invention. Figure 1 As shown, the mobile terminal may include one or more ( Figure 1Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data, wherein the mobile terminal may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the mobile terminal. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0034] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the data method based on the intelligent data analysis large model in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, the above method is implemented. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. The transmission device 106 is used to receive or send data via a network. The above-mentioned specific examples of the network may include a wireless network provided by a communication provider of the mobile terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0035] In this embodiment, a data analysis method based on an intelligent data analysis big model is provided that runs on a mobile terminal, a computer terminal or a similar computing device. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and, although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0036] Figure 2 It is a flow chart of a data analysis method based on an intelligent data analysis big model according to an embodiment of the present application.

[0037] like Figure 2 As shown, the method comprises the following steps:

[0038] Step S201, obtaining corresponding real data from multiple preset data elements to obtain a first data set, and preprocessing the first data set to obtain a second data set, wherein the real data includes a report on data analysis and statistics, a case of data analysis and statistics, and a tutorial on data analysis and statistics, and the preprocessing includes at least data enhancement, data cleaning, and data mixing;

[0039] Specifically, this step is the preparation and processing of data. In this stage, it is divided into four steps: data collection, data enhancement, data cleaning and data mixing. Data collection collects data from diversified sources such as public databases, industry reports, academic literature, news articles, forum discussions, etc., covering economics, finance, statistics, science, business, social sciences and other fields. These preset data elements include but are not limited to data analysis and statistical reports, data analysis and statistical cases, and data analysis and statistical tutorials. The preprocessing step includes at least data enhancement, data cleaning and data mixing.

[0040] Data enhancement can improve the diversity and coverage of data sets by generating synthetic data and data expansion. In order to improve the model's ability to analyze and reason about complex data, we work with data analysis experts to construct multi-dimensional data sets that include complex tables, chart explanations, causal reasoning, hypothesis testing cases, and data visualization results interpretation. These data sets cover a wide range of cases from different industries and fields, and complement the use cases of multiple data analysis technologies and tools, such as SQL, Python, R, and machine learning frameworks. The specific data enhancement technologies used include: 1) For the generation and annotation of complex causal reasoning cases, the causal links between key variables are analyzed based on real data or synthetic data, such as determining causal relationships through the correlation between economic indicators and market behaviors, and using structured causal models (SCM) and Bayesian networks to generate synthetic data, and marking causal paths and relationships with expert annotation or semi-automated tools; 2) For the creation and annotation of data analysis tool use cases, cases with specific technical processes are designed, covering tasks such as SQL queries, Python / R data cleaning and modeling, and machine learning framework predictive analysis, and the tool types, analysis steps and expected results are annotated; 3) For data quality and consistency assurance, standardized annotation and storage formats are formulated, and cross-validation and integrity checks are implemented to ensure the consistency of data sets.

[0041] Data cleaning removes noise data, fills missing values, corrects errors, etc. to improve data quality. Professional data analysts use automated tools and algorithms to remove duplicate records, fill missing values, smooth outliers, remove irrelevant information such as advertisements and irrelevant comments, and retain high-quality data analysis content. In addition, the data conversion of non-numeric data into numerical data through encoding is completed, and the data is normalized or standardized to eliminate the impact of different dimensions and ensure consistent data formats, such as date formats, units of measurement, term definitions, etc. A variety of filters are used, such as learning filters (including perplexity scorers, quality scorers, document consistency scorers, and security scorers), as well as clustering-based filters to ensure the high quality of pre-trained data. In particular, sensitive information such as personal identity information and confidential data is removed or anonymized.

[0042] Data blending integrates data from different sources to ensure that the model can handle various types of data, and builds a large pre-trained corpus consisting of English and Chinese tokens. This corpus comes from multiple sources, and Chinese and English are constructed using a random blending strategy. For different data analysis task types and difficulty levels, the data sets are grouped according to factors such as industry, data size, and analysis complexity (such as user behavior data, product catalog data, market analysis reports, data analysis tool manuals, etc.), and then blended together to form a large and diverse corpus.

[0043] It can be seen that through data cleaning and enhancement, noisy data can be removed, missing values ​​can be filled, and the authenticity and reliability of data can be improved. Data enhancement and data mixing can help improve the diversity of data sets and enable models to perform more stably in different data types and scenarios.

[0044] Step S202, performing multi-level pre-training on the autoregressive language model based on the second data set to obtain a first target model, the multi-level pre-training including text level pre-training, sentence / paragraph level pre-training, long text level pre-training and thought chain level pre-training;

[0045] Specifically, the goal of pre-training is to cultivate a general model that can understand and generate natural language in data analysis scenarios. Based on the standard Transformer architecture, this method has made a series of adjustments and adopted technologies such as Grouped-QueryAttention (GQA), SwiGLU activation function and Rotary Position Embedding (RoPE). The standard Transformer structure mainly includes components such as self-attention mechanism, multi-head attention, position encoding and feedforward neural network. Among them, the self-attention mechanism is particularly critical. By inputting the Query, Key, and Value matrices of the sequence, the mechanism can effectively capture the relationship between the elements in the sequence and provide the model with powerful context understanding capabilities.

[0046] In text-level training, the model is trained using autoregressive methods, similar to self-supervised learning tasks. For example, the Masked Language Model (MLM) enables it to predict the numerical value or text information in the data table cell based on the context, and gradually grasp the inherent laws and relevance of the data. For example, the Generative Pre-trained Transformer (GPT) enables it to master the ability to accurately try to predict the next word in a given context. The corpus data uses unlabeled mixed data.

[0047] In sentence and paragraph level training, higher-level autoregressive tasks are designed, such as next sentence prediction (NSP), to enable the model to learn to predict subsequent analysis steps or summary statements based on the conclusions of previous data analysis, such as predicting the data processing method for the next stage, inferring data trends, and writing the core ideas of data analysis reports. The corpus data uses unlabeled mixed data.

[0048] In the continuous training of long texts, by continuing to train the model on long sequence data, its ability to handle long contexts is further improved. This includes lightweight continuous pre-training of the model to adapt to contexts of more than 200,000 tokens in length. By alternating samples from different fields, a wide range of knowledge can be learned. In order to control the overfitting phenomenon, data enhancement strategies (such as synonymous replacement and random masking) are used to increase the diversity of data samples. At the same time, an early stopping mechanism is set to monitor the generalization error and training error of the model. Once the performance of the model on the validation set tends to be stable! Pre-training is stopped in advance to avoid overfitting caused by long-term training. In order to prevent the model from learning inaccurate information, information extraction tools are used to automatically screen and clean the corpus to remove inaccurate or irrelevant content. At the same time, the text similarity model is used to realize noise detection and remove repeated or irrelevant long text fragments.

[0049] In the thinking chain level training, the business logic, analysis steps, tool calls and response components of complex scenario data analysis are used for unsupervised self-learning to improve the model's logical reasoning ability. Specifically, the thinking chain is extracted and constructed from actual data analysis cases, and the key steps are identified through automated steps or manually annotated by experts, and the specific operations and reasoning logic are recorded, such as "data acquisition-data cleaning-feature engineering-model training-result interpretation". Each step record contains operation instructions, goals and expected results, so that the model can understand the process of the analysis task. The reasoning logic of each step is recorded in the format of "condition-hypothesis-inference-conclusion". For example, when performing data cleaning, the reasoning chain of "missing value detection-evaluation filling strategy-filling missing values" is recorded. For steps involving tool (such as SQL, Python, R) calls, the specific method of the call, parameter settings and expected output are recorded, and each tool call is linked to the corresponding operation step through the "operation-tool-input parameter-output response" structure, so that the model can understand how to select and use the appropriate tool in the analysis process. The format of the thinking chain is represented by a tree structure and a nested structure. Each step includes the operation type, the tool used, the reasoning logic and the expected output. The specific format is as follows: step number: such as "Step 1: Data cleaning"; operation and reasoning: such as "missing value detection → evaluation filling strategy → filling missing values"; tool call: such as "Tool: Python-Function: fillna(method='mean')"; expected response: such as "no missing values ​​in the output table". The annotation method uses a semi-automatic annotation tool to assist manual annotation, and assigns category labels (such as "operation", "tool call", "reasoning path") to each thinking chain node.

[0050] In order to ensure the accuracy and logic of the thought chain corpus, a multi-level review and verification mechanism is adopted. After the thought chain is constructed, it is reviewed by experts in the field of data analysis to verify the logic and accuracy of the reasoning path. Random sampling is performed on long thought chains to ensure the rationality and feasibility of each node in the chain. At the same time, the rationality of the thought chain is detected by combining automated tools, checking the dependency relationship between the previous and next steps, the validity of tool calls, and the consistency of context logic, and excluding unreasonable or incomplete thought chains. This series of measures effectively improves the quality of the thought chain corpus and provides a solid foundation for the accurate reasoning of the model.

[0051] It can be seen that through these levels of pre-training, the model can not only understand and generate natural language expressions for complex data analysis scenarios, but also effectively handle long contexts and have strong logical reasoning capabilities, laying a solid foundation for subsequent model instruction fine-tuning and practical applications.

[0052] Step S203, fine-tuning the model instruction based on the first target model to obtain a second target model, wherein the model instruction fine-tuning includes question-answer pair training, instruction-response pair training, and data set-analysis thinking chain-analysis action pair training;

[0053] Specifically, in the model instruction fine-tuning phase, the goal is to optimize the pre-trained model for a specific task or application scenario, focusing on the quality of the data rather than the quantity. The fine-tuning dataset is constructed through scenario simulation, simulating real-world data analysis scenarios such as data exploration, problem diagnosis, and report writing. Through user feedback and the careful construction and polishing of data analysis experts, the dataset is continuously iterated and optimized to ensure the practicality and accuracy of the model.

[0054] Furthermore, the fine-tuning dataset consists of three types of interactive data: 10,000 multi-round question-answer pairs, multi-round command-response dialogue pairs, and dataset-analysis thinking chain-analysis action. It also contains structured data formats, which enables the model to distinguish different types of information, such as system configuration, user input, and assistant response. Specifically: Multi-round question-answer pairs construct question-answer pairs related to data analysis, such as analysis questions raised by users and answers generated by models. Multi-round command-response pairs create command-response pairs. Users raise data analysis requirements, and the model provides corresponding analysis steps or codes. The data in the dataset-analysis thinking chain-analysis action can be text, numerical, image, or any other form. After the user provides the data set, the model mines information and extracts data knowledge based on the data to form a thinking chain for data analysis, and executes the steps of data analysis to gradually get responses back to the user.

[0055] Examples of data analysis instructions for financial scenarios are as follows:

[0056] Example 1: Risk data analysis:

[0057] Instruction: "Please evaluate the customer's credit risk and identify potential high-risk customers based on the customer's transaction data [data source - transaction system, data acquisition method - interface information] and credit score [data source - credit system, data acquisition method - interface information], and generate an analysis report." Response output: Process the user's transaction data, including data query, data preprocessing, statistical analysis or machine learning method writing, obtain output results, and integrate credit scores to obtain risk assessment reports, including customer credit risk levels and high-risk customer lists.

[0058] Example 2: Data mining analysis:

[0059] Input: customer transaction data, customer credit data. Output: 1. User credit analysis can be performed, the analysis idea is..., the analysis is performed..., and the result is... 2. User fraud detection can be performed, the analysis idea is..., the analysis is performed..., and the result is...

[0060] In addition, a certain proportion of general instructions are introduced to fine-tune the data, such as common sense reasoning, logical judgment, mathematical formula parsing, natural language query SQL conversion, data anomaly detection and other data analysis auxiliary tasks, to enhance the versatility and flexibility of the model in various data analysis scenarios.

[0061] Fine-tune the pre-trained model. This includes adjusting the model's hyperparameters, selecting appropriate loss functions and optimizers. Specifically, the data analysis model fine-tuning refers to the experience of general large language model experts, freezing 80% of the pre-trained model's parameters, and using multi-task instruction response fine-tuning to ensure that the model better follows instructions. The AdamW optimizer is used, with β1 set to 0.9, β2 set to 0.999, and ε set to 10. -8 . Where β1 is the momentum decay rate, β2 is the square gradient decay rate, and ε is a small value to prevent division by zero. The number of training steps is set to 300, the learning rate is 1×10^-5, the weight decay is 0.1, the maximum threshold of gradient clipping is 1.0, and the sequence length of 8192 and the batch size of 64 are used for training. The loss function targets two core tasks: discriminative and generative, of which generative tasks account for the majority. Through these steps, the model not only performs well on specific tasks, but also shows strong versatility and flexibility in a variety of data analysis scenarios.

[0062] Step S204, performing a model performance evaluation based on the second target model, and when the evaluation score is greater than or equal to a preset value, performing a normalization adjustment on the second target model to obtain a third target model, wherein the normalization adjustment is used to unify the model's response to the same problem;

[0063] Specifically, the second target model is first comprehensively evaluated for performance. This evaluation process is carried out by combining automatic evaluation and human evaluation. Automatic evaluation uses standardized tests (benchmark tests) to evaluate the performance of the model in simulated real data analysis scenarios. These tests include but are not limited to classification tasks, regression tasks, and question-answering systems, etc., and are designed to evaluate the capabilities of the model in multiple dimensions.

[0064] The advantages of automatic evaluation are its objectivity and efficiency. It calculates key performance indicators such as accuracy, recall, and F1 score by comparing with the real answer. Human evaluation uses expert judgment to evaluate not only the accuracy of the model, but also its interpretability, reliability, and practicality. In this process, the evaluation dataset can be selected from the evaluation dataset provided by the open source community or the self-built evaluation dataset to ensure the diversity and comprehensiveness of the dataset.

[0065] After the evaluation is completed, if the score of the second target model is greater than or equal to the preset value, it will be normalized. The purpose of normalization is to unify the model's response to the same question and ensure the consistency and reliability of the model output. The specific steps include: standardizing the output format of the model to eliminate ambiguity in the output; reconstructing the output using a predefined template to ensure a unified format; performing consistency testing on the model to verify whether its response to the same question is consistent; if inconsistency is found, adjusting the model parameters or algorithm until the standard is met.

[0066] The data analysis model that has been fine-tuned and evaluated is used to generate the reasoning steps and final results of the target data analysis problem to be solved multiple times. For each target problem, the frequency of generating reasoning processes and results is counted, and the reasoning process and result that appear the most times are selected as the optimal solution to the target problem. This method assumes that if an reasoning process or result is frequently selected, it is more likely to be correct or most in line with expectations.

[0067] Step S205, optimizing the third target model at least based on the user feedback behavior to obtain a fourth target model;

[0068] Specifically, by comprehensively utilizing user feedback and reinforcement learning strategies, continuous optimization of large language model services is achieved. Specifically, after using the service, the system provides a rating interface where users can rate the model's answers from 1 to 5 points to directly quantify satisfaction. At the same time, the system analyzes the user's interactive behavior, such as the length of stay, the frequency of continued interaction, and the frequency of active exit as indirect feedback signals. A reward signal is generated based on the user's satisfaction quantification results: a high satisfaction score (such as 4 or 5 points) generates a positive reward signal to strengthen the current model decision path; a low satisfaction score (such as 1 or 2 points) generates a negative reward signal to adjust the relevant model parameters to reduce the probability of unsatisfactory output. In addition, when collecting and using user feedback data, the system performs noise processing and anonymization to ensure user privacy and data security. Through these measures, not only can the performance and user experience of the large language model be continuously improved, but also user privacy can be protected during the optimization process, realizing an efficient, secure and user-friendly large language model service system.

[0069] Step S206: input the data to be analyzed into the fourth target model to obtain data analysis results.

[0070] Specifically, the data to be analyzed is input into the fourth target model after optimization and evaluation to obtain the final data analysis results. This step is the final stage of model application, ensuring that the model can provide accurate and useful results in actual data analysis tasks. The specific steps are as follows: the data to be analyzed is input into the fourth target model through an interface or data processing tool; the fourth target model analyzes the input data, generates a detailed analysis thinking chain and performs analysis actions; the model outputs the analysis results, including detailed analysis reports, data insights and suggestions, etc.; the output results are verified to ensure their accuracy and practicality, and further adjustments and optimizations are made as needed. Through the above steps, the performance of the model in data analysis tasks is comprehensively improved, and it can provide accurate, practical and efficient data analysis results to meet the diverse needs of users.

[0071] The embodiments of the present application provide a large intelligent data analysis model based on multi-level pre-training and fine-tuning for specific scenarios. It adopts a contextual learning strategy to allow the model to learn solutions to new complex problems through existing analysis processes. In combination with the pre-training-fine-tuning framework, it first uses large-scale unlabeled data for pre-training, and then fine-tunes on a small amount of instruction-response labeled data in a specific data analysis field, ensuring that the model can master general data processing skills and accurately meet the professional needs of specific fields. Model training based on multiple data sources ensures the generalization of the model. In the case of scene changes, unsupervised training and limited sample fine-tuning can be used to adapt to the scene, ensuring the flexibility and scalability of the system, and solving the problems of poor indication migration ability, low flexibility and scalability of data analysis models in the prior art.

[0072] As a possible implementation manner, preprocessing the first data set to obtain the second data set includes the following steps:

[0073] Step S301, analyzing the first case containing a causal relationship in the case based on the report and the tutorial, extracting dependent variables, causal paths and reasoning processes for annotation, and generating a causal link based on the dependent variables, causal paths and reasoning processes;

[0074] Specifically, first carefully read the reports and tutorials to identify cases involving causal relationships. Then, use natural language processing and text mining techniques to automatically extract the dependent variables, causal paths, and reasoning processes in these cases. The dependent variable is the result in the causal relationship, and the causal path is the series of events or conditions that lead to this result. The reasoning process reveals the logic of how to derive the dependent variable from the causal path. The extracted information will be structured and stored for subsequent analysis and utilization. Finally, a clear causal chain diagram is generated based on this information. This diagram can intuitively display the causal relationship and provide strong support for subsequent data analysis and model training.

[0075] Step S302, analyzing the second case involving the use of the data analysis tool in the case based on the report and the tutorial, extracting the data analysis steps, the tool type of the data analysis tool, the use steps of the data analysis tool, and the expected output results, and generating use scenarios of the data analysis tool corresponding to different data types according to the data analysis steps, tool type, use steps, and expected output results;

[0076] Specifically, the system will focus on cases involving the use of data analysis tools in reports and tutorials. Through text analysis and information extraction technology, the system will automatically identify the data analysis steps, tool types used, specific usage steps, and expected output results in these cases. This information is crucial for understanding how to use data analysis tools in practical applications. Based on this information, the system will build data analysis tool usage scenarios corresponding to different data types. These scenarios can not only help users better understand how to use the tools, but also provide rich contextual information for model training.

[0077] Step S303, analyzing the reports, tutorials and cases, extracting data formats in the reports, tutorials and cases, obtaining multiple first data formats, constructing a mapping relationship between each first data format and a corresponding second data format, obtaining a first mapping relationship, wherein the second data format is a data format when data of different data types are used in the first target model;

[0078] Specifically, the system will deeply analyze the text data in reports, tutorials, and cases, and use regular expressions, natural language processing and other technical means to accurately extract the data format information. These formats may include CSV, JSON, XML and other types. The extracted first data format will be systematically classified and organized. Then, the system will build a mapping relationship between these first data formats and the second data format. The second data format refers to the format that these data need to follow when they are actually used in the first target model. The establishment of this mapping relationship ensures that the data can be processed and analyzed correctly and efficiently in the model.

[0079] Step S304, performing data cleaning on the reports, tutorials and cases to obtain first target data, wherein the data cleaning includes deleting duplicate data, supplementing missing values, smoothing outliers, deleting data of a preset type, and converting data formats according to the first mapping relationship;

[0080] Specifically, the system will first identify and delete duplicate data entries to ensure the uniqueness of the data. Then, for missing data values, the system will use interpolation, regression and other methods to reasonably supplement to ensure the integrity of the data. In addition, the system will smooth out the abnormal points to reduce the impact of data noise on the analysis results. At the same time, according to preset rules, the system will delete certain specific types of data, such as invalid data or erroneous data. Finally, using the data format mapping relationship established in step S303, the system will convert all data into the second data format required by the model.

[0081] Step S305, randomly mixing the first target data according to different languages ​​to obtain second target data;

[0082] Specifically, the system will take into account the diversity of data and globalization needs, and randomly mix the first target data according to different languages. This means that when processing multilingual data, the model can better learn and adapt to data characteristics in different language environments. By randomly mixing data in different languages, the model can be exposed to a wider range of language expressions and data patterns during training, thereby improving its performance and accuracy in cross-language environments.

[0083] Step S306: construct a second data set based on the causal link, the usage scenario, the first mapping relationship and the second target data.

[0084] Specifically, the system will use the results of the previous steps to build a second data set. This data set will contain rich causal link information, usage scenarios of data analysis tools, precise data format mapping relationships, and carefully cleaned and diversified data. Such a data set is not only high-quality and clearly structured, but also diverse and practical, and can provide strong data support for subsequent model training.

[0085] As a possible implementation method, multi-level pre-training of the autoregressive language model is performed based on the second data set to obtain a first target model, including the following steps:

[0086] Step S401, based on the autoregressive model, the masked language model and the GPT model, a data analysis model is constructed, and the second data set is decomposed, and text level pre-training is performed with words as basic units to obtain a fifth target model, wherein the autoregressive model is P(x i |x1,…,xi-1 ) When the previous text is input, the conditional probability of the i-th word, P(X) is the output predicted word, and the masked language model is L MLM =-∑ i∈M logP(x i |x1,…,x i-1 ,x i+1 ,…,x n ), M is a randomly selected set of mask positions, L MLM To fill in the text, the GPT model is L is the negative log-likelihood loss;

[0087] Specifically, the autoregressive model predicts the next word based on the previous words. Next, the system decomposes the second dataset into word levels and performs text-level pre-training with words as the basic unit. This means that the model predicts the next word based on the previous word sequence, thereby gradually building a complete sentence. In this way, the model can learn the basic grammatical and semantic structure of the language.

[0088] Step S402, based on the decomposition of the second data set, the fifth target model is pre-trained at the sentence / paragraph level with sentences and paragraphs as basic units, so as to enable the model to generate subsequent steps or summary representations based on the previous text data analysis, thereby obtaining a sixth target model;

[0089] Specifically, the model not only considers the prediction of a single word, but also understands the contextual information at the sentence or paragraph level. Through this level of pre-training, the model can generate subsequent steps or summary representations of data analysis based on the previous text. For example, after reading a data analysis report, the model can generate subsequent analysis steps or summarize key points. This ability is particularly important for complex data analysis tasks because it allows the model to better understand and apply data.

[0090] Step S403, performing synonym replacement and random masking based on the second data set to obtain a third data set, performing long text level pre-training on the sixth target model based on the third data set as a basic unit to obtain a seventh target model, wherein the training is stopped when the fluctuation of the model score based on the validation set during the model training process is less than a first threshold;

[0091] Specifically, random masking is to improve the robustness of the model when dealing with incomplete information. Next, the system will perform long text level pre-training using the third dataset as the basic unit. This means that the model will process longer text sequences to better understand the complex relationships and structures in long texts. During the training process, the system will score the model based on the validation set and stop training when the score fluctuation is less than the first threshold. This mechanism can effectively prevent the model from overfitting and ensure its stability and accuracy in practical applications.

[0092] Step S404, pre-training the sixth target model at the thinking chain level based on the causal links and usage scenarios in the second data set to obtain the first target model, wherein each step in the thinking chain generated by the first target model includes an operation type, a causal link, a tool call, and an expected response.

[0093] Specifically, the system will use the causal links and usage scenarios in the second data set to perform thought chain level pre-training on the sixth target model to obtain the first target model. Thought chain level pre-training means that the model not only generates a single prediction, but also generates a series of logically coherent operation steps. During the training process, the system will pay special attention to the operation type of each step (such as data cleaning, analysis, visualization, etc.), causal links, tool calls (such as specific data analysis tools or APIs), and expected responses (such as generating charts or summary reports). This information can help the model generate structured and logically clear analysis steps in practical applications.

[0094] As a possible implementation method, fine-tuning the model instruction based on the first target model to obtain the second target model includes the following steps:

[0095] Step S501, obtaining a plurality of question-answer pairs, wherein the question-answer pairs include analysis questions input by a user into a first target model and answers generated by the first target model;

[0096] Specifically, the system will obtain multiple question-answer pairs by recording the user's interaction with the first target model. These question-answer pairs include the analysis questions entered by the user and the answers generated by the first target model. For example, a user may enter a question: "What are the outliers in the data set?" The model generates the corresponding answer: "There are 3 outliers in the data set, namely A, B, and C." These question-answer pairs are important data sources in the fine-tuning process. They not only contain the specific needs of the user, but also reflect the performance of the model in actual applications. By analyzing these question-answer pairs, the system can identify the advantages and disadvantages of the model and provide a basis for subsequent fine-tuning.

[0097] Step S502, obtaining a plurality of instruction-response pairs, where the instruction-response pairs include an instruction input by a user into a first target model and an execution step of generating the first target model;

[0098] Specifically, the system obtains multiple instruction-response pairs. These pairs include the instructions of the first target model input by the user and the execution steps generated by the model. For example, the user may issue an instruction: "Please clean the data and generate a visual chart." The execution steps generated by the model may include: "1. Delete duplicate data; 2. Supplement missing values; 3. Generate a bar chart. The instruction-response pairs provide clear operational guidance for the model. They help the model understand the specific needs of the user during the learning process and generate execution steps that meet the user's expectations.

[0099] Step S503, obtaining a plurality of data sets-analysis thought chains-analysis action pairs, wherein the data set-analysis thought chains-analysis action pairs include the data type of the input data input by the user into the first target model and the thought chains, analysis steps and expected analysis results of the first target model analyzing the input data;

[0100] Specifically, the system will obtain multiple data sets-analysis thinking chains-analysis action pairs. These pairs include the data type of the input data of the first target model input by the user, as well as the thinking chain, analysis steps and expected results of the model's analysis of these data. For example, the input data type may be "sales data", and the model's analysis thinking chain may include: "1. Identify sales trends; 2. Analyze seasonal fluctuations; 3. Forecast future sales." The analysis steps and expected results may include: "Generate trend charts, seasonal fluctuation charts and forecast reports." These data sets-analysis thinking chains-analysis action pairs provide the model with full process guidance from data input to analysis results, and they help the model better understand and apply the logic and steps of data analysis.

[0101] Step S504, based on question-answer pairs, instruction-response pairs, and data set-analysis thinking chain-analysis action pairs, the first target model is trained based on the AdamW optimizer to obtain the second target model.

[0102] Specifically, the system will fine-tune the first target model based on the question-answer pairs, instruction-response pairs, and data set-analysis thought chain-analysis action pairs collected above. The fine-tuning process will use the AdamW optimizer, which is an optimization algorithm that combines the Adam optimizer and weight decay, and is particularly suitable for processing large-scale deep learning models. During the fine-tuning process, the model will generate corresponding answers and execution steps based on the questions and instructions entered by the user, and will continuously adjust and optimize through the feedback mechanism. Through this iterative optimization, the model can better meet the specific needs of users and improve its analysis and execution capabilities.

[0103] As a possible implementation method, the model performance evaluation is performed based on the second target model, including the following steps:

[0104] Step S601, obtaining a standard test set, inputting the standard test set into a second target model, and obtaining a first output result;

[0105] Specifically, the system will prepare a standard test set. This test set usually contains multiple sets of data with known answers, which are used to evaluate the performance of the model. The design of the standard test set needs to cover various situations that the model may encounter in actual applications to ensure the comprehensiveness and fairness of the evaluation. Next, the system will input each test case in the standard test set into the second target model to obtain the first output result. These output results will serve as the basis for subsequent evaluation.

[0106] Step S602, comparing the first output result with the benchmark output result corresponding to the standard test set, determining the accuracy, and determining the accuracy as the first evaluation score;

[0107] Specifically, for each test case, the system compares the model output with the benchmark output in the standard test set; counts the number of correct predictions by the model; calculates the total number of predictions; and calculates the accuracy. Next, the system determines the calculated accuracy as the first evaluation score, which reflects the performance of the model on the standard test set.

[0108] Step S603, obtaining a second evaluation score, performing weighted calculation based on the first evaluation score and the second evaluation score to obtain an evaluation score, wherein the second evaluation score is a score input by the user.

[0109] Specifically, the system will obtain the score input by the user, which reflects the user's subjective evaluation of the actual performance of the model. The user score usually covers multiple dimensions, such as the accuracy of the model, response speed, and ease of operation. The specific steps of weighted calculation are as follows: Determine the weights of the first evaluation score and the second evaluation score. The weights can be adjusted according to actual needs. For example, if the user score is more important, the weight of the second evaluation score can be set higher; use the weighted average formula to calculate the final evaluation score: Evaluation score = first evaluation score * first weight + second evaluation score * second weight evaluation score. This final evaluation score takes into account the objective performance of the model and the subjective evaluation of the user, and can more comprehensively reflect the comprehensive performance of the model.

[0110] As a possible implementation method, the second target model is normalized and adjusted to obtain a third target model, including the following steps:

[0111] Step S701, inputting a second data set into a second target model to obtain a plurality of second output results;

[0112] Specifically, the second data set usually contains samples of different types and structures, which will be used in the subsequent normalization adjustment process. The specific steps are as follows: traverse each sample in the second data set; input each sample into the second target model to obtain the corresponding output result; collect all the output results to form a set, which is called the second output result set.

[0113] Step S702, classifying the second output result based on the sample type of the samples in the second data set to obtain a plurality of fourth data sets;

[0114] Specifically, the sample type may include different data structures, data fields, analysis tasks, etc. The specific steps are as follows: classify the samples in the second data set into different sample types; classify the second output result set according to the sample type to obtain multiple fourth data sets. Each fourth data set contains the second output result of the corresponding sample type.

[0115] Step S703, counting the occurrence frequency of each second output result in each fourth data set, determining the second output result with the highest corresponding frequency as the response of the corresponding sample type, and obtaining the third target model.

[0116] Specifically, each fourth data set is traversed, and the frequency of occurrence of each second output result is counted; the second output result with the highest frequency in each fourth data set is determined and used as the standard response of the corresponding sample type; these standard responses are integrated into the model to form a third target model. The third target model can output consistent and accurate results when processing samples of the same type.

[0117] It can be seen that by counting the second output results with the highest frequency in each sample type, the model can output consistent and accurate responses, which improves the consistency of the model when processing samples of the same type. Normalization adjustment makes the results of the model output more standardized, reduces the output inconsistency caused by sample differences, and improves the stability and reliability of the model.

[0118] As a possible implementation method, the third target model is optimized based on the user feedback behavior to obtain the fourth target model, including the following steps:

[0119] Step S801, determining a user feedback score based on the user feedback behavior, and determining the user feedback score as direct satisfaction;

[0120] Specifically, user feedback behavior can include user evaluation of model output results, rating of model functions, etc. The specific steps are as follows: collect user feedback behaviors, such as ratings, comments, etc.; calculate feedback ratings based on user feedback behaviors, which can be 1-5 points or other levels; determine the calculated feedback ratings as direct satisfaction, which reflects the user's direct evaluation of the model output.

[0121] Step S802, determining the feedback duration and interaction frequency of the user completing the user feedback behavior based on the user feedback behavior, and determining the feedback duration and interaction frequency as indirect satisfaction;

[0122] Specifically, feedback duration refers to the time it takes for users to complete feedback from receiving model output, and interaction frequency refers to the number of times users interact with the model. The specific steps are as follows: record the time it takes for users to complete feedback behavior and calculate the feedback duration; count the number of interactions between users and the model and calculate the interaction frequency; determine the feedback duration and interaction frequency as indirect satisfaction, which reflects the efficiency and frequency of interaction between users and the model.

[0123] Step S803, performing weighted calculation based on direct satisfaction and indirect satisfaction to obtain a satisfaction score;

[0124] Specifically, the purpose of weighted calculation is to comprehensively consider the user's direct evaluation and interaction to obtain a comprehensive satisfaction score. The specific steps are as follows: determine the weights of direct satisfaction and indirect satisfaction, and the weights can be adjusted according to actual needs; use the weighted average formula to calculate the final satisfaction score: Satisfaction score = direct satisfaction * direct weight + indirect satisfaction * indirect weight satisfaction score.

[0125] Step S804, when the satisfaction score is greater than or equal to the second threshold, applying a positive reward signal to the third target model to obtain a fourth target model;

[0126] Specifically, the positive reward signal can be a positive adjustment of the model parameters to encourage the model to continue to optimize. The specific steps are as follows: set a second threshold, which indicates that the user satisfaction is high; when the satisfaction score is greater than or equal to the second threshold, apply a positive reward signal to the third target model, such as increasing the learning rate, adjusting parameters, etc.; through the positive reward signal, the model will be further optimized to form the fourth target model.

[0127] Step S805, when the satisfaction score is less than or equal to the third threshold, applying a negative reward signal to the third target model to obtain a fourth target model;

[0128] Specifically, the negative reward signal can be a negative adjustment to the model parameters to improve the performance of the model. The specific steps are as follows: set a third threshold, which indicates that the user satisfaction is low; when the satisfaction score is less than or equal to the third threshold, apply a negative reward signal to the third target model, such as reducing the learning rate, adjusting parameters, etc.; through the negative reward signal, the model will be adjusted and improved to form a fourth target model.

[0129] Step S806: When the satisfaction score is greater than the third threshold and less than the second threshold, the third target model is determined as the fourth target model.

[0130] Specifically, a third threshold and a second threshold are set, and a situation between the two thresholds indicates medium user satisfaction. When the satisfaction score is greater than the third threshold and less than the second threshold, the third target model is not adjusted and is directly determined as the fourth target model.

[0131] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0132] The embodiment of the present application also provides a data analysis device based on a large model of intelligent data analysis. It should be noted that the data analysis device based on a large model of intelligent data analysis in the embodiment of the present application can be used to execute the data analysis method based on a large model of intelligent data analysis provided in the embodiment of the present application. The device is used to implement the above-mentioned embodiments and preferred implementation modes, and those that have been described will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware for a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived.

[0133] The following is an introduction to the data analysis device based on the intelligent data analysis big model provided in the embodiment of the present application.

[0134] Figure 3 1 is a structural block diagram of a data analysis device based on an intelligent data analysis big model according to an embodiment of the present application. Figure 3 As shown, the device includes: a first processing unit 10, a first training unit 20, a second processing unit 30, a second training unit 40, a third training unit 50 and an input unit 60.

[0135] A first processing unit 10 is used to obtain corresponding real data from a plurality of preset data elements to obtain a first data set, and preprocess the first data set to obtain a second data set, wherein the real data includes a report on data analysis and statistics, a case of data analysis and statistics, and a tutorial on data analysis and statistics, and the preprocessing includes at least data enhancement, data cleaning, and data mixing;

[0136] A first training unit 20 is used to perform multi-level pre-training on the autoregressive language model based on the second data set to obtain a first target model, wherein the multi-level pre-training includes text level pre-training, sentence / paragraph level pre-training, long text level pre-training, and thought chain level pre-training;

[0137] A second processing unit 30 is used to perform model instruction fine-tuning based on the first target model to obtain a second target model, wherein the model instruction fine-tuning includes question-answer pair training, instruction-response pair training, and data set-analysis thinking chain-analysis action pair training;

[0138] A second training unit 40 is used to perform a model performance evaluation based on the second target model. When the evaluation score is greater than or equal to a preset value, the second target model is normalized to obtain a third target model. The normalization adjustment is used to unify the model's response to the same problem.

[0139] A third training unit 50 is used to optimize the third target model based on at least the user feedback behavior to obtain a fourth target model;

[0140] The input unit 60 is used to input the data to be analyzed into the fourth target model to obtain data analysis results.

[0141] The embodiment of the present application provides a data analysis device based on an intelligent data analysis large model, the device comprising: a first processing unit, a first training unit, a second processing unit, a second training unit, a third training unit and an input unit. The device performs model training based on multiple data sources to ensure the generalization of the model, and can perform scene adaptation through unsupervised training and limited sample fine-tuning in the case of scene change, thereby ensuring the flexibility and scalability of the system, and solving the problem of poor indication migration capability, low flexibility and scalability of the data analysis model in the prior art.

[0142] As a possible implementation manner, the first processing unit includes: a first analysis module, a second analysis module, a third analysis module, a cleaning module, a mixing module and a data set construction module.

[0143] A first analysis module is used to analyze the first case containing causal relationships in the case based on the report and the tutorial, extract dependent variables, causal paths and reasoning processes for annotation, and generate causal links based on the dependent variables, causal paths and reasoning processes;

[0144] A second analysis module is used to analyze the second case containing the use of the data analysis tool in the case based on the report and the tutorial, extract the data analysis steps, the tool type of the data analysis tool, the use steps of the data analysis tool and the expected output results, and generate the use scenarios of the data analysis tool corresponding to different data types according to the data analysis steps, tool type, use steps and expected output results;

[0145] A third analysis module is used to analyze the reports, tutorials and cases, extract the data formats in the reports, tutorials and cases, obtain multiple first data formats, construct a mapping relationship between each first data format and the corresponding second data format, and obtain a first mapping relationship, where the second data format is a data format of data of different data types when used in the first target model;

[0146] A cleaning module, used to perform data cleaning on reports, tutorials and cases to obtain first target data, wherein data cleaning includes deleting duplicate data, supplementing missing values, smoothing abnormal points, deleting data of preset types, and converting data formats according to the first mapping relationship;

[0147] A mixing module, used for randomly mixing the first target data according to different languages ​​to obtain second target data;

[0148] A data set construction module is used to construct a second data set based on the causal link, the usage scenario, the first mapping relationship and the second target data.

[0149] As a possible implementation manner, the first training unit includes: a model building module, a first processing module, a second processing module and a third processing module.

[0150] A model building module is used to build a data analysis model based on an autoregressive model, a masked language model and a GPT model, and decompose the second data set to perform text-level pre-training with words as basic units to obtain a fifth target model, wherein the autoregressive model is P(x i |x1,…,x i-1 )When the previous text is input, the conditional probability of the i-th word, P(X) is the output predicted word, and the masked language model is M is a randomly selected set of mask positions, L MLM To fill in the text, the GPT model is L is the negative log-likelihood loss;

[0151] A first processing module is used to perform sentence / paragraph level pre-training on a fifth target model based on the decomposition of the second data set and taking sentences and paragraphs as basic units, so as to enable the model to generate subsequent steps or summary representations based on the previous text by performing data analysis, thereby obtaining a sixth target model;

[0152] A second processing module is used to perform synonym replacement and random masking based on the second data set to obtain a third data set, and to perform long text level pre-training on the sixth target model based on the third data set as a basic unit to obtain a seventh target model, wherein the training is stopped when the fluctuation of the model score based on the validation set during the model training process is less than a first threshold;

[0153] The third processing module is used to pre-train the sixth target model at the thinking chain level based on the causal links and usage scenarios in the second data set to obtain the first target model, wherein each step in the thinking chain generated by the first target model includes an operation type, a causal link, a tool call, and an expected response.

[0154] As a possible implementation manner, the second processing unit includes: a first acquisition module, a second acquisition module, a third acquisition module and a training module.

[0155] A first acquisition module is used to acquire multiple question-answer pairs, where the question-answer pairs include analysis questions input by the user into the first target model and answers generated by the first target model;

[0156] A second acquisition module is used to acquire a plurality of instruction-response pairs, wherein the instruction-response pair includes an instruction input by a user into the first target model and an execution step of generating the first target model;

[0157] A third acquisition module is used to acquire a plurality of data sets-analysis thought chains-analysis action pairs, wherein the data set-analysis thought chain-analysis action pairs include the data type of the input data input by the user into the first target model and the thought chain, analysis steps and expected analysis results of the first target model analyzing the input data;

[0158] The training module is used to train the first target model based on the AdamW optimizer based on question-answer pairs, instruction-response pairs and data set-analysis thinking chain-analysis action pairs to obtain the second target model.

[0159] As a possible implementation manner, the second training unit includes: a first training module, a second training module and a third training module.

[0160] A first training module is used to obtain a standard test set, input the standard test set into a second target model, and obtain a first output result;

[0161] A second training module is used to compare the first output result with a benchmark output result corresponding to the standard test set to determine an accuracy rate, and determine the accuracy rate as a first evaluation score;

[0162] The third training module is used to obtain a second evaluation score, and the evaluation score is obtained by performing weighted calculation based on the first evaluation score and the second evaluation score, where the second evaluation score is a score input by a user.

[0163] As a possible implementation manner, the second training unit also includes: a fourth training module, a fifth training module and a sixth training module.

[0164] A fourth training module, used for inputting the second data set into the second target model to obtain a plurality of second output results;

[0165] a fifth training module, configured to classify the second output result based on the sample type of the samples in the second data set to obtain a plurality of fourth data sets;

[0166] The sixth training module is used to count the occurrence frequency of each second output result in each fourth data set, determine the second output result with the highest corresponding frequency as the response of the corresponding sample type, and obtain the third target model.

[0167] As a possible implementation, the third training unit includes: a scoring module, a satisfaction module, a calculation module, a positive module, a negative module and a determination module.

[0168] A scoring module, used for determining a user feedback score based on the user feedback behavior, and determining the user feedback score as direct satisfaction;

[0169] A satisfaction module, used to determine the feedback duration and interaction frequency of the user completing the user feedback behavior based on the user feedback behavior, and determine the feedback duration and interaction frequency as indirect satisfaction;

[0170] A calculation module, used for performing weighted calculation based on direct satisfaction and indirect satisfaction to obtain a satisfaction score;

[0171] A positive module, used for applying a positive reward signal to the third target model when the satisfaction score is greater than or equal to the second threshold, so as to obtain a fourth target model;

[0172] A negative module, used for applying a negative reward signal to the third target model when the satisfaction score is less than or equal to a third threshold, so as to obtain a fourth target model;

[0173] A determination module is used to determine the third target model as the fourth target model when the satisfaction score is greater than the third threshold and less than the second threshold.

[0174] The data analysis device based on the intelligent data analysis big model includes a processor and a memory. The first processing unit, the first training unit, the second processing unit, the second training unit, the third training unit and the input unit are all stored in the memory as program units, and the processor executes the program units stored in the memory to realize the corresponding functions. The above modules are all located in the same processor; or, the above modules are located in different processors in any combination.

[0175] The processor includes a kernel, which calls the corresponding program unit from the memory. One or more kernels can be set, and the communication efficiency can be improved by adjusting the kernel parameters.

[0176] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0177] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the data analysis method based on the intelligent data analysis big model.

[0178] An embodiment of the present invention provides a data analysis system, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs include a method for executing the above-mentioned data analysis method based on the intelligent data analysis big model.

[0179] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0180] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0181] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0182] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0183] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0184] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0185] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0186] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0187] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0188] The above are only preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A data analysis method based on an intelligent data analysis big model, characterized in that: include: Acquire corresponding real data from a plurality of preset data elements to obtain a first data set, and preprocess the first data set to obtain a second data set, wherein the real data includes a data analysis and statistical report, a data analysis and statistical case, and a data analysis and statistical tutorial, and the preprocessing includes at least data enhancement, data cleaning, and data mixing; Performing multi-level pre-training on the autoregressive language model based on the second data set to obtain a first target model, wherein the multi-level pre-training includes text-level pre-training, sentence / paragraph-level pre-training, long text-level pre-training, and thought chain-level pre-training; Fine-tuning the model instruction based on the first target model to obtain a second target model, wherein the model instruction fine-tuning includes question-answer pair training, instruction-response pair training, and data set-analysis thinking chain-analysis action pair training; Performing a model performance evaluation based on the second target model, and when the evaluation score is greater than or equal to a preset value, performing a normalization adjustment on the second target model to obtain a third target model, wherein the normalization adjustment is used to unify the model's response to the same problem; Optimizing the third target model at least based on user feedback behavior to obtain a fourth target model; The data to be analyzed is input into the fourth target model to obtain data analysis results.

2. The method according to claim 1, characterized in that Preprocessing the first data set to obtain a second data set includes: Analyze the first case containing a causal relationship in the case based on the report and the tutorial, extract dependent variables, causal paths and reasoning processes for annotation, and generate a causal link based on the dependent variables, causal paths and reasoning processes; Analyze a second case in the case that includes the use of a data analysis tool based on the report and the tutorial, extract data analysis steps, a tool type of the data analysis tool, use steps of the data analysis tool, and expected output results, and generate use scenarios of the data analysis tool corresponding to different data types according to the data analysis steps, the tool type, the use steps, and the expected output results; Analyze the report, the tutorial, and the case, extract data formats in the report, the tutorial, and the case, obtain multiple first data formats, construct a mapping relationship between each of the first data formats and a corresponding second data format, and obtain a first mapping relationship, wherein the second data format is a data format when data of different data types are used in the first target model; Performing data cleaning on the report, the tutorial and the case to obtain first target data, wherein the data cleaning includes deleting duplicate data, supplementing missing values, smoothing outliers, deleting data of a preset type and converting data formats according to the first mapping relationship; Randomly mixing the first target data according to different languages ​​to obtain second target data; The second data set is constructed based on the causal link, the usage scenario, the first mapping relationship and the second target data.

3. The method according to claim 2, characterized in that Performing multi-level pre-training on the autoregressive language model based on the second data set to obtain a first target model, including: Based on the autoregressive model, the masked language model and the GPT model, a data analysis model is constructed, and the second data set is decomposed, and the text level pre-training is performed with words as the basic unit to obtain a fifth target model, wherein the autoregressive model is P(x i |x1,…,x i-1 )When the previous text is input, the conditional probability of the i-th word, P(X) is the output predicted word, and the masked language model is L MLM =-∑ i∈M logP(x i |x1,…,x i-1 ,x i+1 ,…,x n ), M is a randomly selected set of mask positions, L MLM To fill in the text, the GPT model is L is the negative log-likelihood loss; Based on the decomposition of the second data set, the sentence / paragraph level pre-training is performed on the fifth target model with sentences and paragraphs as basic units, so as to enable the model to generate subsequent steps or summary representations based on the previous text data analysis, thereby obtaining a sixth target model; Perform synonym replacement and random masking based on the second data set to obtain a third data set, perform the long text level pre-training on the sixth target model based on the third data set as a basic unit to obtain a seventh target model, wherein the training is stopped when the fluctuation of the model score based on the validation set during the model training process is less than a first threshold; The sixth target model is pre-trained at the thinking chain level based on the causal links and the usage scenarios in the second data set to obtain the first target model, wherein each step in the thinking chain generated by the first target model includes an operation type, the causal link, a tool call, and an expected response.

4. The method according to claim 1, characterized in that Fine-tuning the model instruction based on the first target model to obtain a second target model includes: Acquire a plurality of the question-answer pairs, wherein the question-answer pairs include analysis questions input by a user into the first target model and answers generated by the first target model; Acquire a plurality of the command-response pairs, wherein the command-response pairs include a command input by a user into the first target model and an execution step of generating the first target model; Acquire a plurality of data sets-analysis thought chains-analysis action pairs, wherein the data sets-analysis thought chains-analysis action pairs include the data type of the input data input by the user into the first target model and the thought chains, analysis steps and expected analysis results of the first target model analyzing the input data; Based on the question-answer pairs, the instruction-response pairs, and the data set-analysis thinking chain-analysis action pairs, the first target model is trained based on the AdamW optimizer to obtain the second target model.

5. The method according to claim 1, characterized in that Performing a model performance evaluation based on the second target model includes: Obtaining a standard test set, inputting the standard test set into the second target model, and obtaining a first output result; Determine an accuracy rate based on a comparison between the first output result and a benchmark output result corresponding to the standard test set, and determine the accuracy rate as a first evaluation score; A second evaluation score is obtained by performing a weighted calculation based on the first evaluation score and the second evaluation score, wherein the second evaluation score is a score input by a user.

6. The method according to claim 1, characterized in that The second target model is normalized and adjusted to obtain a third target model, including: Inputting the second data set into the second target model to obtain a plurality of second output results; Classifying the second output result based on sample types of samples in the second data set to obtain multiple fourth data sets; The occurrence frequency of each of the second output results in each of the fourth data sets is counted, and the second output result with the highest corresponding frequency is determined as the response of the corresponding sample type to obtain the third target model.

7. The method according to claim 1, characterized in that The third target model is optimized based on the user feedback behavior to obtain a fourth target model, including: Determine a user feedback score based on the user feedback behavior, and determine the user feedback score as direct satisfaction; Determine, based on the user feedback behavior, the feedback duration and the interaction frequency of the user completing the user feedback behavior, and determine the feedback duration and the interaction frequency as indirect satisfaction; Perform weighted calculation based on the direct satisfaction and the indirect satisfaction to obtain a satisfaction score; When the satisfaction score is greater than or equal to the second threshold, applying a positive reward signal to the third target model to obtain the fourth target model; When the satisfaction score is less than or equal to a third threshold, applying a negative reward signal to the third target model to obtain the fourth target model; When the satisfaction score is greater than the third threshold and less than the second threshold, the third target model is determined as the fourth target model.

8. A data analysis device based on an intelligent data analysis big model, characterized in that: The device comprises: A first processing unit is used to obtain corresponding real data from a plurality of preset data elements to obtain a first data set, and preprocess the first data set to obtain a second data set, wherein the real data includes a report on data analysis and statistics, a case of data analysis and statistics, and a tutorial on data analysis and statistics, and the preprocessing includes at least data enhancement, data cleaning, and data mixing; A first training unit is used to perform multi-level pre-training on the autoregressive language model based on the second data set to obtain a first target model, wherein the multi-level pre-training includes text level pre-training, sentence / paragraph level pre-training, long text level pre-training and thought chain level pre-training; A second processing unit is used to perform model instruction fine-tuning based on the first target model to obtain a second target model, wherein the model instruction fine-tuning includes question-answer pair training, instruction-response pair training, and data set-analysis thinking chain-analysis action pair training; A second training unit is used to perform a model performance evaluation based on the second target model, and when the evaluation score is greater than or equal to a preset value, perform a normalization adjustment on the second target model to obtain a third target model, wherein the normalization adjustment is used to unify the model's response to the same problem; A third training unit, configured to optimize the third target model based at least on user feedback behavior to obtain a fourth target model; The input unit is used to input the data to be analyzed into the fourth target model to obtain data analysis results.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 7.

10. A data analysis system, characterized in that: include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs include methods for executing any one of claims 1 to 7.

Images