Quotation analysis model construction method and device, equipment and medium

Abstract

The invention relates to the technical field of data processing, and discloses a quotation analysis model construction method and device, equipment and a medium, and the method comprises the steps: obtaining original full-amount data, and labeling the original full-amount data according to a preset structured format to obtain initial full-amount data; wherein the original total data represents complete data including historical quotation data and newly added quotation data on that day; cleaning dirty data based on the initial total data to obtain processed total data; wherein the dirty data cleaning comprises the step of carrying out outlier data screening by utilizing a trained instruction fine tuning model; determining open source data as unstructured data, and comprehensively processing the unstructured data and the processed total data to obtain target training data; and training a large language model by using the target training data to obtain a quotation analysis model. The method has the beneficial effects of standard structured text and real-time response.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method, device, equipment and medium for constructing a quotation analysis model. Background Art

[0002] The bond repo market is an important liquidity management tool in the financial market, providing financial institutions with a flexible short-term funding channel, especially in the inter-bank and non-inter-bank markets. Through structured analysis of the quote information in the repo market, it is possible to efficiently capture transaction data, support intelligent trading decisions and conduct precise quantitative analysis.

[0003] However, the use of general large models for repo market quote analysis still faces challenges such as insufficient understanding of professional fields, difficulty in format control, and long response time. At present, some solutions have been used to improve accuracy by plug-in knowledge bases or fine-tuning models, but these methods have not completely solved the problems of format control and real-time processing. In general, existing technologies still have major bottlenecks in professional adaptability, real-time performance, and format control. Summary of the invention

[0004] The present application provides a quotation parsing model construction method, device, equipment and medium, which solves the technical problems of how to solve format control and real-time processing, and achieves the technical effects of standard structured text and real-time response.

[0005] In order to achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, an embodiment of the present application provides a method for constructing a quotation parsing model, the method comprising: Obtaining original full data, and annotating the original full data according to a preset structured format to obtain initial full data; wherein the original full data represents complete data including historical quotation data and newly added quotation data of the day; Cleaning dirty data based on the initial full data to obtain processed full data; wherein the dirty data cleaning includes screening outlier data using a trained instruction fine-tuning model; Determine the open source data as unstructured data, and comprehensively process the unstructured data and the processed full data to obtain target training data; The target training data is used to train a large language model to obtain a quotation parsing model.

[0006] The method for constructing a quotation parsing model provided in this embodiment obtains the original full data and performs structured annotation to ensure the consistency and operability of the data, which is convenient for subsequent processing. Next, the dirty data is cleaned and the outlier data is screened out to effectively improve the data quality and reduce the impact of abnormal data. Then, the open source unstructured data and structured data are comprehensively processed to enhance the diversity and comprehensiveness of the data, so that the model can extract more valuable information. Finally, by training the large language model with the target training data, a model that can efficiently and accurately parse quotation data is obtained.

[0007] Optionally, the cleaning of dirty data based on the initial full data to obtain processed full data includes: Deduplication is performed on the initial full amount of data to obtain first full amount of data; Performing high similarity screening on the first full amount of data to obtain second full amount of data; Using the trained instruction fine-tuning model to filter outlier data from the second full data, to obtain third full data; A simple sample screening is performed on the third full amount of data to obtain the processed full amount of data.

[0008] This embodiment first ensures that each sample in the data set is unique through deduplication operations, thereby avoiding the interference of redundant data on subsequent analysis. Then, duplicate or highly similar samples are removed through high similarity screening, which improves the diversity of the data set and reduces the risk of overfitting certain patterns. Then, the data is fine-tuned using the trained instruction model to screen outliers and eliminate abnormal samples. These outlier data may be noise or erroneous data, which have a negative impact on the analysis results. Finally, simple sample screening further removes samples with insufficient information or abnormal formats, and optimizes data consistency. The whole process gradually cleans and simplifies the data, improves data quality, provides a more stable and reliable data foundation for subsequent modeling and analysis, and improves the efficiency and accuracy of model training while reducing computing overhead.

[0009] Optionally, the performing high-similarity screening on the first full data to obtain the second full data includes: Performing word segmentation on each data sample in the first full amount of data to obtain multiple target word segments, and storing the target word segments in a bag-of-words model; For any data sample, determine the word frequency of the target word in the bag-of-words model; According to the word frequency, converting any data sample into a word frequency vector; Based on the word frequency vector, the similarity between any two data samples is determined, and one data sample is randomly retained from the two data samples whose similarity exceeds a preset threshold to obtain the second full amount of data.

[0010] This embodiment can convert data samples into a quantifiable form by segmenting the data samples and constructing a bag-of-words model, thereby extracting key information. On this basis, the frequency of occurrence of each word is calculated and converted into a word frequency vector for subsequent analysis and processing. This process standardizes the text data and represents it as a numerical vector, eliminating the impact of grammatical and format differences on the analysis. Then, by calculating the similarity between data samples, redundant samples with high similarity can be identified and removed, thereby improving the diversity and quality of the data set. This series of operations not only helps to reduce the impact of redundant data on model training, but also improves the effectiveness of the data.

[0011] Optionally, the instruction fine-tuning model is a Qwen2-7B model; and the method for acquiring the trained instruction fine-tuning model includes: Acquire historical quotation data, and annotate the historical quotation data according to a preset structured format to obtain a quotation data set; The LoRA training method is adopted to train the Qwen2-7B model using the quotation data set, and the quotation data inconsistent with the label is removed until the preset conditions are met to obtain a trained instruction fine-tuning model.

[0012] This embodiment obtains historical quotation data and labels it, converts the original data into a structured data set, ensures the uniformity and standardization of the data, and provides high-quality input for subsequent training. Next, the Qwen2-7B model is fine-tuned using the LoRA training method, and quotation data that is inconsistent with the label is automatically removed during the training process. This training method not only improves efficiency and reduces computing resource consumption, but also cleans data through an automated screening process to ensure that the model only comes into contact with high-quality samples, thereby improving the accuracy and effectiveness of training. Ultimately, the optimized instruction fine-tuning model can accurately identify and process complex quotation data, has greater adaptability and robustness, and can provide accurate prediction and decision support for specific tasks.

[0013] Optionally, the large language model is a Qwen2.5-32B model; and the step of training the large language model using the target training data to obtain a quotation parsing model includes: The LoRA fine-tuning and RLHF training methods are adopted, and the large language model is trained using the target training data until the model converges to obtain the quotation parsing model.

[0014] This embodiment increases the low-rank matrix through LoRA fine-tuning, and only adjusts task-specific parameters, which reduces the consumption of computing resources and speeds up the training speed, while maintaining the versatility of the model in other tasks. This enables the model to adapt to the needs of quotation parsing more efficiently and improves the ability to identify key information in the quotation text. Secondly, RLHF training combines human feedback to enable the model to gradually adjust the output when generating quotation parsing, thereby improving the accuracy of the results and the degree to which they meet user expectations. The model after LoRA fine-tuning and RLHF training can not only efficiently process complex or ambiguous quotation data, but also generate structured and clear parsing results according to business needs, greatly reducing the need for manual intervention. Finally, after sufficient training and optimization, the model shows high stability and reliability in practical applications, and can quickly respond and output accurate quotation parsing.

[0015] Optionally, the Qwen2.5-32B model is specifically a Qwen2.5-32B-base model or a Qwen2.5-32B-chat model.

[0016] Optionally, the method further comprises: The obtained real-time quotation data is parsed using the quotation parsing model to obtain quotation structured data.

[0017] This embodiment obtains real-time quotation data and inputs it into the quotation parsing model for processing, thereby ensuring the real-time and accuracy of the data, and performing analysis based on the latest market quotations to improve the timeliness and accuracy of decision-making; secondly, through automated data processing, work efficiency is improved, manual intervention and errors are reduced, and the original quotation data is converted into a structured format to facilitate subsequent storage, query and analysis; in addition, the output of structured data provides a standardized data basis for subsequent analysis, prediction and decision-making.

[0018] In a second aspect, an embodiment of the present application provides a quotation parsing model construction device, the device comprising: A data acquisition unit, used to acquire original full data, and annotate the original full data according to a preset structured format to obtain initial full data; wherein the original full data represents complete data including historical quotation data and newly added quotation data of the day; A data cleaning unit, configured to clean dirty data based on the initial full data to obtain processed full data; wherein the dirty data cleaning includes screening outlier data using a trained instruction fine-tuning model; A training data determination unit, configured to determine the open source data as unstructured data, and to comprehensively process the unstructured data and the processed full data to obtain target training data; The model training unit is used to train the large language model using the target training data to obtain a quotation parsing model.

[0019] In a third aspect, an embodiment of the present application provides a computer device, including: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the quotation analysis model construction method described in the above claims by executing the computer instructions.

[0020] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which computer instructions are stored, and the computer instructions are used to enable a computer to execute the quotation parsing model construction method described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the specific implementation methods of the present application or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 A flowchart of a method for constructing a quotation analysis model provided in an embodiment of the present application; Figure 2 A flowchart of step S3 provided in an embodiment of the present application; Figure 3 A flowchart of step S33 provided in an embodiment of the present application; Figure 4 A flowchart of a method for obtaining a trained instruction fine-tuning model provided in an embodiment of the present application; Figure 5 A flowchart of step S7 provided in an embodiment of the present application; Figure 6 A flowchart of the quotation analysis provided for the embodiment of the present application; Figure 7 A block diagram of a quotation parsing model building device provided in an embodiment of the present application; Figure 8 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, 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 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 those skilled in the art without creative work are within the scope of protection of the present application.

[0024] The bond repo market is a key liquidity management tool in the financial market. It provides financial institutions with a flexible and efficient channel for short-term borrowing and lending. Banks and other financial institutions can quickly adjust their funding status through repo transactions and maintain the liquidity balance of the banking system, which is crucial to the stability of financial markets. In the repo market, time is an important factor. The concept of "time is cost" is particularly prominent. Studies have shown that differences in transaction speed directly affect the ability to seize market opportunities. Investors with faster speeds can react quickly to market price fluctuations and buy and sell in time, thereby realizing additional benefits.

[0025] The repo market can be divided into the interbank and non-interbank markets, with a large number of transactions occurring in the non-interbank market. In the non-interbank repo market, participants conclude transactions with selected counterparties on a transaction-by-transaction basis through price inquiries, and quotes are usually provided by money brokers or other intermediaries. This type of quote information is different from the standardized quotes in the secondary bond market, which are usually brief and non-standard, and contain multiple transaction elements. For example, the quote information may include expressions such as "borrow 1m 1.80 mortgage interest rate", and understanding and processing these quote information requires high-frequency and automated structured parsing.

[0026] Structural analysis of repo market quotation information is of great significance, including: Capture transaction information in a timely manner: Through efficient structured analysis, you can obtain market transaction data in the first place and grasp transaction opportunities.

[0027] Support for intelligent trading: Structured data can be directly connected to the system to support intelligent trading decisions and execution.

[0028] Data quantitative analysis: Structured data facilitates accurate quantitative analysis of downstream tasks such as fund sentiment index.

[0029] Challenges of general large models in analysis: With the continuous development of big model technology, more and more tasks can be achieved by big models. However, directly using general big models to analyze repo market quotes still faces multiple challenges: 1. Insufficient understanding of professional fields: The general large model lacks an in-depth understanding of the over-the-counter trading field, especially some proprietary terms. For example, "eryong" refers to secondary perpetual bonds, "laohu" refers to accounts that have had transactions before, and "cd" refers to large transferable certificates of deposit. These terms are difficult for the general large model to understand, thus affecting the accuracy of the analysis results.

[0030] 2. Difficulty in format control: Although general large models can handle text generation tasks, they cannot effectively control the output format, resulting in greater uncertainty in the generation of structured data. This not only increases the complexity of subsequent processing, but may also lead to unstable data quality.

[0031] 3. Long response time: Due to the huge number of parameters in the general large model, high computing resource requirements and slow processing speed, when processing large batches of quotation data, the response time often cannot meet the needs of real-time capture of transaction information (for example, each piece of data takes an average of 6 seconds to process), so it is impossible to seize trading opportunities in time.

[0032] There are already some quotation structured parsing solutions based on large models on the market: for example, the solution that uses a large model plug-in knowledge base makes up for the lack of general large models in professional field understanding by retrieving external knowledge bases. Although this solution improves accuracy and richness, it still fails to solve the problem of format control, and the use of external knowledge bases leads to longer response times. For example, there are currently fine-tuning solutions that focus on document structured parsing, which are mainly optimized based on annotated data, and have high requirements for annotations in new fields. This method is costly and has poor adaptability to exclusive fields.

[0033] In order to solve the above technical problems, according to an embodiment of the present application, a quotation parsing model construction method embodiment is provided. 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 the 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.

[0034] In this embodiment, a method for constructing a quotation analysis model is provided. Figure 1 A flowchart of a method for constructing a quotation analysis model provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the process includes the following steps: Step S1, obtaining the original full data, and annotating the original full data according to a preset structured format to obtain the initial full data; wherein the original full data represents complete data including historical quotation data and newly added quotation data of the day.

[0035] Specifically, collect complete data including historical quotation data and newly added quotation data of the day from some clients or enterprise dialog boxes. Define a preset structured format, which should be able to capture the key information in the data. The structured format can include the following fields: buying and selling direction (sell / lend), yield (%), term, quantity (billion), pledged subject, account and other requirements. Then manually annotate each piece of raw data in the original full data, and convert the raw data into a structured format, including reading each piece of quotation data, extracting key information and filling it into the predefined structured format. An example is shown in Table 1.

[0036] Table 1 Initial full data Step S3, cleaning dirty data based on the initial full data to obtain processed full data; wherein, cleaning dirty data includes using the trained instruction fine-tuning model to screen outlier data.

[0037] Specifically, dirty data usually refers to inaccurate, incomplete, invalid or inconsistent data generated due to various reasons during data collection, transmission, storage or processing. If this type of data is not cleaned up, it may have a negative impact on subsequent data analysis and model training. For example, there are missing data, duplicate data, inconsistent data or outlier data. The outlier data screening is processed by the trained instruction fine-tuning model. The trained instruction fine-tuning model has been exposed to a large amount of labeled data and has undergone specific training for outlier data, so that it can identify data anomalies based on the input data. Therefore, the trained instruction fine-tuning model is used to screen outlier data and identify abnormal data that is inconsistent with most of the data and may have problems.

[0038] Step S5, determining the open source data as unstructured data, and comprehensively processing the unstructured data and the processed full data to obtain target training data.

[0039] Specifically, since the data in the fund quotation process may be mixed with interfering content such as user conversations, voices or pictures, such as "I want to drink Coke on a rainy day, issue a credit bond...", "The overnight train is coming,..." and other mixed communication content samples will cause serious failure of fund quotations. Therefore, the impact of these unstructured data needs to be considered when training the model. In the embodiment of the present application, the processed full amount of data is structured data after formatting structure annotation, and is trained in combination with unstructured data. Unstructured data refers to data without a predefined format or structure, usually including text, images, audio, video, etc., such as articles, comments, social media posts, etc. This type of data format is flexible and complex, and it is usually necessary to use natural language processing (NLP), computer vision (CV) and other technologies when processing, such as text segmentation, semantic analysis, sentiment analysis, etc.

[0040] In this embodiment, the unstructured data mainly includes open source data sets, such as WikiQA, Alpaca_zh, ShareGPT4 and eFundGPT training corpus. The target training data is a data set that combines structured data and unstructured data and is optimized by a multi-task learning strategy. These data sets can not only improve the learning effect of the model, but also enhance the performance of the model in different tasks and fields.

[0041] Step S7, training the large language model using the target training data to obtain a quotation parsing model.

[0042] Specifically, by using the target training data to train the large language model, a quotation parsing model is obtained, which can effectively extract and parse quotation information from structured and unstructured data, and then generate structured data of fund quotations required by users.

[0043] For example, input: 5kw 7D 1.87 limited to public fund accounts, I want to drink Coke on rainy days. Output: Buy and sell direction: out, volume (100 million): 0.5, yield (%): 1.87, pledged subject: none, term: 7 days, other requirements: limited to public fund accounts.

[0044] Input: not less than 1.8% overnight: +10 mortgage rate, state-owned shares, local, +12AAA commercial gold, AAA certificate of deposit, public offering proprietary account. Output: Buy and sell direction: out, volume (billion): none, yield (%): +10bp, pledged subject: interest rate, state-owned shares, local, term: overnight, other requirements: none. Buy and sell direction: out, volume (billion): none, yield (%): +12bp, pledged subject: AAA commercial gold, AAA certificate of deposit, term: overnight, other requirements: public offering proprietary account.

[0045] The method for constructing a quotation parsing model provided in this embodiment obtains the original full data and performs structured annotation to ensure the consistency and operability of the data, which is convenient for subsequent processing. Next, the dirty data is cleaned and the outlier data is screened out to effectively improve the data quality and reduce the impact of abnormal data. Then, the open source unstructured data and structured data are comprehensively processed to enhance the diversity and comprehensiveness of the data, so that the model can extract more valuable information. Finally, by training the large language model with the target training data, a model that can efficiently and accurately parse quotation data is obtained.

[0046] Figure 2 The flowchart of step S3 provided in the embodiment of the present application may include the following steps: Step S31, deduplicate the initial full amount of data to obtain the first full amount of data.

[0047] Specifically, the original text of each quotation in the initial full data is compared. If they are completely consistent, they are marked as duplicates. Samples with incomplete original text but the same investment elements are retained to obtain the first full data. This can prevent the model learning from being affected by differences in sentence expressions and avoid data redundancy. Ensure that the deduplication operation is limited to samples with the same content, and for samples with the same investment elements but different expressions, retaining this information helps enhance the generalization ability of the model.

[0048] For example, for each quotation, extract the core investment elements (such as "7D", "1.85", "5", etc.), ignore non-business content (such as chat information), and standardize the details in the quotation. For example, "7D continues to be issued, 1.85 5" and "7D 1.85 5" can be converted into the standard format "7D 1.85 5". For quotations containing chat or irrelevant information, certain text preprocessing rules (such as removing specific keywords or dialogue parts) can be used to clean the data. Regular expressions or natural language processing (NLP) technology can also be used to identify and remove non-business information such as "I want to eat hamburgers today". After processing the quotation information and standardizing it, deduplication can be performed based on the standardized investment elements. At this time, only those samples with exactly the same investment elements are retained, and completely repeated quotation content is deleted. For example, among multiple quotations, if the investment elements are exactly the same (such as "7D 1.85 5"), only one sample is retained.

[0049] Step S33: screening the first full amount of data for high similarity to obtain the second full amount of data.

[0050] Specifically, the first full data may be redundant and contain samples with too high similarity, which may affect the accuracy and efficiency of subsequent analysis. The first full data needs to be screened for high similarity. High similarity screening is generally completed by calculating the similarity between data samples. Common similarity calculation methods are: Cosine similarity: measure the similarity by calculating the angle between samples in the vector space. Jaccard similarity: measure the proportion of shared features in two samples. Euclidean distance: calculate the distance between samples in the feature space. The smaller the distance, the more similar the samples are. During the screening process, a similarity threshold is set, and only samples with a similarity higher than this threshold will be selected. The second full data is a new data set obtained after high similarity screening. Compared with the first full data, this data set removes many samples with high similarity.

[0051] Step S35, using the trained instruction fine-tuning model to filter outlier data from the second full data to obtain third full data.

[0052] Specifically, the trained instruction fine-tuning model is used to process the second full data, i.e., the unfiltered data set, to filter out the outliers, and finally obtain the third full data, i.e., the filtered data set. The instruction fine-tuning model is a model obtained by fine-tuning on the basis of historical quotation data using technologies such as LoRA. It has learned how to identify the rules and trends in quotation data, and can distinguish between normal data and abnormal data (i.e., outliers). The trained model has the ability to effectively classify and filter data. In this process, the main task of the fine-tuning model is to automatically identify and exclude abnormal data that does not conform to most data patterns by comparing the consistency between actual data and labels. Finally, the data set after outlier data filtering is the third full data. After being processed by the instruction fine-tuning model, this data set has eliminated abnormal data that does not conform to the overall data trend and rules, and has higher quality and is more suitable for subsequent analysis or training tasks.

[0053] Step S37, performing simple sample screening on the third full amount of data to obtain processed full amount of data.

[0054] Specifically, the third full data has been screened for outliers by the instruction fine-tuning model, which means that the abnormal data has been removed, and the data quality is relatively high, which conforms to the overall rules and trends. However, even so, there may still be some extreme samples that do not meet the analysis requirements, such as samples with data points that are too short, insufficient information, or non-standard formats. These samples may affect the accuracy of subsequent analysis, so they need to be eliminated through further screening. In this process, the criterion for sample screening is mainly the length of the sample. Samples with too short lengths usually contain less information and may not be able to effectively describe the laws of the data, especially in scenarios involving time series and quantitative data. Short samples often have difficulty providing sufficient features for the model. For example, samples like "7d 1.85" may be due to the oversimplification of the data format and the relatively small amount of information, which makes it difficult to play a positive role in analysis or model training. Specifically, "7d" may represent a unit of time, meaning 7 days, but without more context, the meaning of 7 days may be difficult to effectively compare with other data points, and the information is lacking. "1.85" is a numerical value that may represent the value of a certain indicator, but without more detailed background or additional data, this value appears to be relatively isolated and lacks context. This kind of data sample is invalid in many analytical tasks because it lacks sufficient context or feature points to provide sufficient statistical significance.

[0055] The screening step includes first performing a length analysis on all samples in the third full data to identify those samples that are significantly shorter, for example, samples that contain only a few characters or data points. These samples are then marked and distinguished from other eligible samples. For these shorter samples such as "7d 1.85", they are randomly or conditionally removed according to the set screening criteria such as the length threshold. The removal process may be through setting a threshold, and all samples below the threshold will be deleted. After simple sample screening, the processed full data is obtained.

[0056] and Figure 1Compared with the embodiment shown, this embodiment first ensures that each sample in the data set is unique through a deduplication operation, thereby avoiding the interference of redundant data on subsequent analysis. Then, duplicate or highly similar samples are removed through high similarity screening, which improves the diversity of the data set and reduces the risk of overfitting certain patterns. Then, the data is fine-tuned using the trained instruction model to screen outliers and eliminate abnormal samples. These outlier data may be noise or erroneous data, which have a negative impact on the analysis results. Finally, simple sample screening further removes samples with insufficient information or abnormal formats, and optimizes the consistency of the data. The whole process gradually cleans and simplifies the data, improves the data quality, provides a more stable and reliable data foundation for subsequent modeling and analysis, and improves the efficiency and accuracy of model training while reducing computing overhead.

[0057] Figure 3 The flowchart of step S33 provided in the embodiment of the present application may include the following steps: Step S331, segment each data sample in the first full data to obtain multiple target segmentations, and store the target segmentations in a bag-of-words model.

[0058] Specifically, each data sample is segmented into words or vocabulary units. After the segmentation, the vocabulary of all data samples is aggregated to build a bag-of-words model. The bag-of-words model is used to count the number of times each word appears in the data sample, ignoring the order of the vocabulary.

[0059] Step S333: for any data sample, determine the frequency of the target word in the bag-of-words model.

[0060] Step S335, converting any data sample into a word frequency vector according to the word frequency.

[0061] Specifically, through the bag-of-words model, each data sample can be converted into a word frequency vector. Each dimension of this vector represents a word in the bag-of-words, and the value represents the frequency of the word in the sample. If a word does not appear in the document, the corresponding value is 0.

[0062] Step S337, based on the word frequency vector, determine the similarity between any two data samples, randomly retain one data sample from the two data samples whose similarity exceeds a preset threshold, and obtain the second full amount of data.

[0063] Specifically, using the vectorized data, the similarity between any two data samples can be calculated. A common similarity calculation method is cosine similarity. If the similarity between two samples exceeds a preset threshold (such as 0.8), the two samples are considered similar and one is retained. For each data sample, its similarity with all other samples is checked. If the similarity is higher than the set threshold, one sample is deleted and the other is retained.

[0064] Through the above method, we can remove those duplicate or almost identical samples according to the similarity of the text, thereby reducing the redundant information during model training. This not only helps to improve the quality of training data, but also effectively prevents the model from being overly biased towards certain duplicate samples.

[0065] and Figure 2 Compared with the embodiment shown, this embodiment can convert data samples into a quantifiable form by segmenting data samples and constructing a bag-of-words model, thereby extracting key information therein. On this basis, the frequency of occurrence of each word is calculated and converted into a word frequency vector to facilitate subsequent analysis and processing. This process standardizes the text data and represents it as a numerical vector, eliminating the impact of grammatical and format differences on the analysis. Then, by calculating the similarity between data samples, redundant samples with high similarity can be identified and removed, thereby improving the diversity and quality of the data set. This series of operations not only helps to reduce the impact of redundant data on model training, but also improves the effectiveness of the data.

[0066] Figure 4 A flowchart of a method for obtaining a trained instruction fine-tuning model provided in an embodiment of the present application, wherein the instruction fine-tuning model is a Qwen2-7B model; the process may include the following steps: Step S351, obtaining historical quotation data, and annotating the historical quotation data according to a preset structured format to obtain a quotation data set.

[0067] Step S353, adopt the LoRA training method, use the quotation data set to train the Qwen2-7B model, remove the quotation data inconsistent with the label, until the preset conditions are met, and obtain the trained instruction fine-tuning model.

[0068] Specifically, historical quotation data is collected, and the quotation data is classified and identified in a predetermined format. For details, see step S1. LoRA (Low-Rank Adaptation) is a lightweight training method that aims to efficiently adjust the parameters of the model in the form of a low-rank matrix, thereby reducing the computing resources required for training. The LoRA method can quickly adapt to new tasks without changing the weights of the original model. Using LoRA can significantly reduce the computational overhead of fine-tuning while maintaining good performance, especially for pre-trained models with a large number of parameters. The labeled quotation data set is input into the Qwen2-7B model for training. During the training process, if the results output by the model are inconsistent with the preset labels (i.e., prediction errors), these samples will be removed. In this way, the model gradually learns which data is wrong and excludes it from training. Training will continue until the model meets the preset stop condition, such as the training error reaches a certain threshold or reaches a predetermined number of training rounds. Here, the training round epoch is set to 3 and the learning rate lr=1e-5.

[0069] It should be noted that during the training process, all samples are used as training sets and test sets. This means that the model not only has to learn the data through training, but also has to verify whether its predictions are consistent with the labels at each step. This method is often called "self-supervised verification" or "self-testing". By training the model and verifying it, these outlier data can be automatically discovered and removed from the training set. Finally, through the above steps, a Qwen2-7B model fine-tuned by LoRA was obtained, which can automatically identify and remove quotation data that is inconsistent with the label. The model has been able to process historical quotation data more accurately, remove inaccurate or abnormal samples, and improve the accuracy of the overall model.

[0070] and Figure 1 Compared with the embodiment shown in the figure, this embodiment obtains historical quotation data and annotates it, converts the original data into a structured data set, ensures the uniformity and standardization of the data, and provides high-quality input for subsequent training. Next, the Qwen2-7B model is fine-tuned using the LoRA training method, and quotation data that is inconsistent with the label is automatically removed during the training process. This training method not only improves efficiency and reduces computing resource consumption, but also cleans data through an automated screening process to ensure that the model only comes into contact with high-quality samples, thereby improving the accuracy and effectiveness of training. Ultimately, the optimized instruction fine-tuning model can accurately identify and process complex quotation data, has greater adaptability and robustness, and can provide accurate prediction and decision support for specific tasks.

[0071] Figure 5The flowchart of step S7 provided in the embodiment of the present application, the large language model is the Qwen2.5-32B model; the process may include the following steps: Step S71, adopt LoRA fine-tuning and RLHF training methods, use target training data to train the large language model until the model converges to obtain a quotation parsing model.

[0072] In some preferred embodiments, the Qwen2.5-32B model is specifically a Qwen2.5-32B-base model or a Qwen2.5-32B-chat model.

[0073] Qwen2.5-32B is a large-scale pre-trained language model based on 32B (32 billion) parameters, which means it has extremely strong language understanding and generation capabilities. The Qwen2.5-32B model can be different versions of the model, including Qwen2.5-32B-base or Qwen2.5-32B-chat. Qwen2.5-32B-base model: Usually the basic version of the model, which has been pre-trained on large-scale data and can handle general language understanding and generation tasks. Qwen2.5-32B-chat model: Based on the base version, it may have been further optimized, especially for adaptive training of conversational tasks. This version may perform better in understanding and generating natural conversations.

[0074] LoRA reduces the amount of parameters that need to be updated through low-rank matrices, thereby greatly reducing the computing resources and storage overhead required for fine-tuning. Since only a small number of parameters are modified, LoRA fine-tuning usually requires less training time and computing resources, and can retain most of the capabilities of the pre-trained model, avoiding the complexity of retraining the entire model. RLHF (Reinforcement Learning with Human Feedback) is a training method that combines reinforcement learning and human feedback. The core idea of ​​this method is to guide the model to produce outputs that are more in line with human intentions by combining human evaluation feedback with reinforcement learning. RLHF usually includes two stages: Pre-training stage: First, the model will perform traditional supervised learning and train using large-scale labeled data.

[0075] Reinforcement learning phase: In this phase, the model further adjusts its behavior based on human feedback. Humans evaluate the model’s output and reward or punish the model based on the evaluation results. In this way, the model is able to better understand and generate content that meets human expectations.

[0076] Specifically, the embodiment of the present application combines LoRA and RLHF to achieve higher training efficiency and output that is more in line with human needs. LoRA is responsible for efficiently adjusting the pre-trained model so that it can adapt to specific tasks or fields, while RLHF optimizes the behavior of the model through human feedback to make it more in line with user expectations. During the above training process, the model will continuously update its parameters until it reaches a convergence state. For LoRA fine-tuning, convergence means that the parameters of the low-rank matrix have been successfully adjusted so that the model has achieved better results in specific tasks. For RLHF, convergence means that the model has learned the most appropriate behavior strategy through human feedback and can generate results that meet human expectations. After completing LoRA fine-tuning and RLHF training, the quotation parsing model is finally obtained. The model can parse the input quotation data, understand the user's intentions, and generate structured data about the quotation.

[0077] and Figure 1 Compared with the embodiment shown, this embodiment adds a low-rank matrix through LoRA fine-tuning, and only adjusts task-specific parameters, which reduces the consumption of computing resources and speeds up the training speed, while maintaining the versatility of the model in other tasks. This enables the model to adapt to the needs of quotation parsing more efficiently and improves the ability to identify key information in the quotation text. Secondly, RLHF training combines human feedback to enable the model to gradually adjust the output when generating quotation parsing, thereby improving the accuracy of the results and the degree to which they meet user expectations. The model after LoRA fine-tuning and RLHF training can not only efficiently process complex or ambiguous quotation data, but also generate structured and clear parsing results according to business needs, greatly reducing the need for manual intervention. Finally, after sufficient training and optimization, the model shows high stability and reliability in practical applications, and can quickly respond and output accurate quotation parsing.

[0078] Figure 6 A flowchart of quotation analysis provided in an embodiment of the present application, the process may include the following steps: Step S9: parse the acquired real-time quotation data using the quotation parsing model to obtain quotation structured data.

[0079] Specifically, the real-time quotation data can be obtained from the user's interactive dialogue, and the real-time quotation data is input into the quotation parsing model obtained in steps S1 to S7, and the model can directly output quotation structured data. For example, input: 5kw 7D 1.87 is limited to public fund public offering accounts, and I want to drink Coke on rainy days. Output: Buy and sell direction: out, volume (billion): 0.5, yield (%): 1.87, pledged subject: none, term: 7 days, other requirements: limited to public fund accounts.

[0080] and Figure 1 Compared with the embodiment shown, this embodiment obtains real-time quotation data and inputs it into the quotation parsing model for processing, thereby ensuring the real-time and accuracy of the data, and performing analysis based on the latest market quotations to improve the timeliness and accuracy of decision-making; secondly, through automated data processing, work efficiency is improved, manual intervention and errors are reduced, and the original quotation data is converted into a structured format to facilitate subsequent storage, query and analysis; in addition, the output of structured data provides a standardized data basis for subsequent analysis, prediction and decision-making.

[0081] In order to illustrate the technical effects achieved by the embodiments of the present invention, the Qwen, LLaMA and GLM series of open source large models in the open source community were evaluated. Qwen is a series of open source large language models launched by Alibaba Group. Qwen2 and Qwen2.5 versions were selected, covering multiple models such as base, chat, code and Math, and involving versions of different sizes from 0.5B to 72B. LLaMA is a series of open source large language models launched by Meta (formerly Facebook), of which LLaMA 3 is the latest version of the series, including two parameter versions of 8B and 70B. The LLaMA model has become one of the most popular models in the global AI community due to its excellent performance and open source strategy. GLM is a series of pre-trained models launched by Zhipu AI, of which GLM-4-9B is the latest open source version. GLM-4-9B performs well on multiple data sets, surpassing the performance of LLaMA 3-8B.

[0082] The evaluation is performed using TTF, TPOT, and latency metrics. TTFT refers to the time it takes from the user inputting a query to the model generating the first output token. TPOT refers to the time it takes for the model to generate each subsequent token. In other words, this is a measure of how fast the model processes each token in a continuous conversation or generation task. Latency refers to the time it takes from the user inputting a query to the model completing the entire response. Latency is directly affected by TTFT and TPOT.

[0083] The above basic model is directly used to evaluate the test set, as shown in Table 2, model performance comparison table.

[0084] Table 2 Model performance comparison From this, we can see that the Llama3-8B model and GLM4 performed poorly on Chinese questions and had limited capabilities. Secondly, Qwen's chat model performed better overall than the base model. In particular, the base model had a large number of circular answers and hallucination problems because it did not fine-tune the command following for the conversation context. Finally, the Qwen2.5 model has been significantly improved compared to Qwen2, showing stronger capabilities in multiple test scenarios.

[0085] Accordingly, please refer to Figure 7 A block diagram of a quotation parsing model building device provided in an embodiment of the present application, the terminal includes: The data acquisition unit 101 is used to acquire the original full data and annotate the original full data according to a preset structured format to obtain the initial full data; wherein the original full data represents the complete data including the historical quotation data and the newly added quotation data of the day; The data cleaning unit 103 is used to clean dirty data based on the initial full data to obtain processed full data; wherein the dirty data cleaning includes screening outlier data using the trained instruction fine-tuning model; The training data determination unit 105 is used to determine the open source data as unstructured data, and to comprehensively process the unstructured data and the processed full data to obtain target training data; The model training unit 107 is used to train the large language model using the target training data to obtain a quotation parsing model.

[0086] In some optional implementations, the data cleaning unit 103 includes: Deduplication is performed on the initial full amount of data to obtain the first full amount of data; Perform high similarity screening on the first full amount of data to obtain the second full amount of data; The trained instruction fine-tuning model is used to filter outlier data from the second full data to obtain the third full data; Perform simple sample screening on the third full amount of data to obtain the processed full amount of data.

[0087] In some optional implementations, performing high similarity screening on the first full amount of data to obtain the second full amount of data includes: Segment each data sample in the first full data to obtain multiple target segmented words, and store the target segmented words in a bag-of-words model; For any data sample, determine the frequency of the target word in the bag-of-words model; According to the word frequency, any data sample is converted into a word frequency vector; Based on the word frequency vector, the similarity between any two data samples is determined, and one data sample is randomly retained from the two data samples whose similarity exceeds a preset threshold to obtain the second full amount of data.

[0088] In some optional implementations, the instruction fine-tuning model is a Qwen2-7B model; the method for obtaining the trained instruction fine-tuning model includes: Obtain historical quotation data, and annotate the historical quotation data according to a preset structured format to obtain a quotation data set; The LoRA training method is adopted to train the Qwen2-7B model using the quotation dataset, and the quotation data that is inconsistent with the label is removed until the preset conditions are met to obtain a trained instruction fine-tuning model.

[0089] In some optional implementations, the large language model is a Qwen2.5-32B model; the large language model is trained using the target training data to obtain a quotation parsing model, including: Adopting LoRA fine-tuning and RLHF training methods, the large language model is trained with target training data until the model converges to obtain the quotation parsing model.

[0090] In some optional implementations, the Qwen2.5-32B model is specifically a Qwen2.5-32B-base model or a Qwen2.5-32B-chat model.

[0091] In some optional embodiments, the device further comprises: The quotation parsing unit is used to parse the acquired real-time quotation data using the quotation parsing model to obtain quotation structured data.

[0092] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0093] In this embodiment, a quotation parsing model building device is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0094] See also Figure 8 , Figure 8 A schematic diagram of the structure of a computer device provided in an embodiment of the present application is shown in FIG. Figure 8As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 8 A processor 10 is taken as an example.

[0095] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0096] The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0097] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device 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 combinations thereof.

[0098] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0099] The computer device further comprises a communication interface 30 for the computer device to communicate with other devices or a communication network.

[0100] The embodiment of the present application also provides a computer-readable storage medium. The above method according to the embodiment of the present application can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0101] The devices and units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0102] For the convenience of description, the above device is described in various units according to their functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0103] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods and devices. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0104] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices, and apparatuses according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to 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.

[0105] 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.

[0106] 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 A step that specifies a function in one or more boxes.

[0107] 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.

[0108] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0109] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

[0110] Although the embodiments of the present application have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present application, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for constructing a quotation analysis model, characterized in that: The construction method comprises: Obtaining original full data, and annotating the original full data according to a preset structured format to obtain initial full data; wherein the original full data represents complete data including historical quotation data and newly added quotation data of the day; Cleaning dirty data based on the initial full data to obtain processed full data; wherein the dirty data cleaning includes screening outlier data using a trained instruction fine-tuning model; Determine the open source data as unstructured data, and comprehensively process the unstructured data and the processed full data to obtain target training data; The target training data is used to train a large language model to obtain a quotation parsing model.

2. The method according to claim 1, characterized in that The cleaning of dirty data based on the initial full data to obtain processed full data includes: Deduplication is performed on the initial full amount of data to obtain first full amount of data; Performing high similarity screening on the first full amount of data to obtain second full amount of data; Using the trained instruction fine-tuning model to filter outlier data from the second full data, to obtain third full data; A simple sample screening is performed on the third full amount of data to obtain the processed full amount of data.

3. The method according to claim 2, characterized in that The step of screening the first full amount of data for high similarity to obtain the second full amount of data includes: Performing word segmentation on each data sample in the first full amount of data to obtain multiple target word segments, and storing the target word segments in a bag-of-words model; For any data sample, determine the word frequency of the target word in the bag-of-words model; According to the word frequency, converting any data sample into a word frequency vector; Based on the word frequency vector, the similarity between any two data samples is determined, and one data sample is randomly retained from the two data samples whose similarity exceeds a preset threshold to obtain the second full amount of data.

4. The method according to claim 1 or 2, characterized in that: The instruction fine-tuning model is a Qwen2-7B model; the method for obtaining the trained instruction fine-tuning model includes: Acquire historical quotation data, and annotate the historical quotation data according to a preset structured format to obtain a quotation data set; The LoRA training method is adopted to train the Qwen2-7B model using the quotation data set, and the quotation data inconsistent with the label is removed until the preset conditions are met to obtain a trained instruction fine-tuning model.

5. The method according to claim 1, characterized in that The large language model is a Qwen2.5-32B model; and the large language model is trained using the target training data to obtain a quotation parsing model, including: The LoRA fine-tuning and RLHF training methods are adopted, and the large language model is trained using the target training data until the model converges to obtain the quotation parsing model.

6. The method according to claim 1 or 5, characterized in that: The Qwen2.5-32B model is specifically a Qwen2.5-32B-base model or a Qwen2.5-32B-chat model.

7. The method according to claim 1, characterized in that The method further comprises: The obtained real-time quotation data is parsed using the quotation parsing model to obtain quotation structured data.

8. A quotation analysis model building device, characterized in that: The device comprises: A data acquisition unit, used to acquire original full data, and annotate the original full data according to a preset structured format to obtain initial full data; wherein the original full data represents complete data including historical quotation data and newly added quotation data of the day; A data cleaning unit, configured to clean dirty data based on the initial full data to obtain processed full data; wherein the dirty data cleaning includes screening outlier data using a trained instruction fine-tuning model; A training data determination unit, configured to determine the open source data as unstructured data, and to comprehensively process the unstructured data and the processed full data to obtain target training data; The model training unit is used to train the large language model using the target training data to obtain a quotation parsing model.

9. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the quotation analysis model construction method according to any one of claims 1 to 7 by executing the computer instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the quotation analysis model construction method according to any one of claims 1 to 7.

Images