Data accurate tuning method based on large language model

Abstract

The invention discloses a data accurate tuning method based on a large language model, which comprises the following steps of: S1, acquiring multi-source data and preprocessing the multi-source data to obtain standardized data with original data meta-information; s2, constructing an intelligent quality evaluation model based on a pre-trained large language model; s3, screening the standardized data obtained in the step S1 by adopting the intelligent quality evaluation model constructed in the step S2 to obtain screened preliminary qualified data; s4, performing adaptive data screening and classification on the preliminarily qualified data screened out in the step S3, and obtaining preset proportion data reserved according to the quality score of each classification; s5, performing enhancement processing on the data reserved in the step S4 according to a preset proportion, screening newly generated enhanced data by adopting the intelligent quality evaluation model constructed in the step S2, and reserving the enhanced data in the preset proportion; and S6, mixing the original data reserved in the step S4 and the enhanced data in the step S5 according to a preset proportion, and finally outputting tuning data containing the original data and the enhanced data which are mixed according to the preset proportion.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and natural language processing technology, specifically to a data precision optimization method based on a large language model. Background Technology

[0002] Large Language Models (LLMs) such as GPT, BERT, and LLaMA have made groundbreaking progress in natural language processing tasks. However, the performance of these models is highly dependent on the quality and quantity of training data. In practical applications, how to accurately optimize massive amounts of training data to improve the model's performance in specific domains or tasks has become a critical problem that urgently needs to be solved. Existing technologies mainly include: Data cleaning techniques: improving data quality through rule filtering, deduplication, and format standardization. These methods mainly rely on predefined rule sets, such as length filtering, special character ratio detection, and duplicate content identification, to perform preliminary screening and cleaning of the data. Active learning methods: selecting samples with the most information for annotation and training to reduce annotation costs. This method uses strategies such as uncertainty sampling, query committees, and expected model changes to identify the most valuable unlabeled data for manual annotation. Data augmentation techniques: expanding training data through methods such as back-translation, synonym substitution, and sentence transformation. Common augmentation strategies include dictionary-based synonym substitution, language model-based sentence rewriting, and cross-language back-translation. Course learning strategy: Training data is organized in order of increasing difficulty to improve model learning efficiency. This strategy simulates the human learning process, starting with simple samples and gradually transitioning to complex samples, thereby improving model convergence speed and final performance.

[0003] The main drawbacks of existing technical solutions are: Lack of intelligent filtering capabilities: Existing methods mainly rely on predefined rules and simple statistical indicators, failing to deeply understand the semantic quality of the data. This results in 15-20% of high-quality data being mistakenly filtered, while the omission rate of low-quality data is as high as 25%. Poor domain adaptability: Rules and thresholds are usually fixed, making it difficult to adapt to the specific needs of different domains and tasks. Performance drops by 30-40% when applied across domains. Insufficient data augmentation capabilities: Existing data augmentation methods are relatively mechanical, generating limited data diversity and easily introducing noise, with augmented data quality improving only by 10-15%. Lack of feedback optimization mechanisms: The system cannot automatically adjust the data filtering strategy based on the model training effect, requiring significant manual intervention, with manual costs accounting for over 60% of the total cost. Low processing efficiency: For TB-level large-scale data, processing time can reach several days, making it difficult to meet the needs of rapid iteration. Summary of the Invention

[0004] The purpose of this application is to provide a method for accurate data tuning based on a large language model. The specific technical solution is as follows:

[0005] A data-driven precision tuning method based on a large language model includes: S1, collecting multi-source data and preprocessing it to obtain standardized data with original data metadata; S2, constructing an intelligent quality assessment model based on a pre-trained large language model; S3, using the intelligent quality assessment model constructed in S2 to filter the standardized data obtained in S1 to obtain preliminary qualified data; S4, performing adaptive data filtering and classification on the preliminary qualified data selected in S3 to obtain a preset proportion of data to be retained for each category based on quality scores; S5, enhancing the data retained in S4 according to a preset proportion, and using the intelligent quality assessment model constructed in S2 to filter the newly generated enhanced data, retaining the enhanced data of the preset proportion; S6, mixing the original data retained in S4 and the enhanced data in S5 according to a preset proportion to ensure the balance of data distribution, avoid model overfitting to the enhanced data pattern, and finally outputting tuned data containing the original data and enhanced data mixed according to the preset proportion.

[0006] S1 includes: S1.1, configuring API, database, file system and crawler interface, using a distributed collection architecture to collect multi-source data with preset collection frequency and batch size; S1.2, standardizing the multi-source data collected in S1.1 into a unified format and retaining the original data's metadata; S1.3, performing preliminary cleaning on the multi-source data after unifying the format in S1.2, and finally obtaining standardized data with the original data's metadata.

[0007] The initial cleaning process in S1.3 includes: S1.31, removing HTML tags, special characters, and control characters; S1.32, correcting obvious encoding errors and garbled text; S1.33, filtering text with abnormal length; and S1.34, identifying and marking potentially sensitive information.

[0008] The construction of the intelligent quality assessment model in S2 includes: S2.1, using a pre-trained large language model as the base; S2.2, designing a quality assessment prompt template, with assessment dimensions including semantic coherence, information completeness, language fluency, factual accuracy, and domain relevance; S2.3, setting the temperature parameter to 0.3 to ensure assessment stability; and S2.4, setting Top-p to 0.9 to maintain a certain level of assessment flexibility.

[0009] The standardized data screening process in S3 includes: S3.1, using the intelligent quality assessment model built in S2 to score the standardized data obtained in S1 in terms of semantic coherence, information completeness, language fluency, factual accuracy, and domain relevance; S3.2, setting preset weight configurations and weighting the average scores for semantic coherence, information completeness, language fluency, factual accuracy, and domain relevance; S3.3, setting preset quality thresholds and screening out preliminary qualified data that exceeds the thresholds.

[0010] The adaptive data filtering and classification process in S4 includes: S4.1, automatically adjusting the quality threshold based on data distribution, using K-means clustering to identify quality levels and dynamically adjusting the retention ratio according to the target data volume; S4.2, using a hierarchical classification system to build a multi-label classifier, identifying multiple attributes of the data, with the classification confidence threshold set to 0.8; S4.3, using the MinHash algorithm to calculate similarity, with the similarity threshold set to 0.85, identifying near-duplicates, and further refining deduplication based on edit distance, retaining the version with the highest quality score.

[0011] The augmentation process in S5 includes: S5.1, using a pre-trained large language model as a base; S5.2, designing a data augmentation prompt, generating 3-5 similar but different variants for each original sample, with generation requirements including keeping the core semantics unchanged, changing the expression and sentence structure, appropriately adding or adjusting details, and ensuring that the generated text is natural and fluent; S5.3, using temperature sampling to increase diversity, with temperature = 0.7-0.9.

[0012] It also includes: S7, real-time monitoring of the performance of models trained using the optimized data output in S6, automatic adjustment of quality assessment weights based on usage results, optimization of data screening thresholds, adjustment of data augmentation strategies, updating of domain relevance calculation methods, and optimization of the overall strategy using reinforcement learning.

[0013] The beneficial effects of this application are as follows: By introducing a large language model and intelligent processing mechanism, the following improvements are achieved: Significantly improved data quality: Through deep semantic understanding using LLM, the accuracy of data filtering is greatly improved. Enhanced domain adaptability: The adaptive mechanism enables the system to quickly adapt to new domains, achieving a high performance retention rate when applied across domains, which is a significant improvement compared to fixed rule methods. Improved data augmentation effect: The augmented data generated by LLM achieves an average quality score of 85 points, significantly improving data diversity and model performance. Increased automation: The feedback optimization mechanism greatly reduces the degree of manual intervention, lowering system maintenance costs. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the application process.

[0015] Figure 2 This is a schematic diagram of the data precision optimization system in this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this application. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0017] like Figure 1 As shown, a data-driven precision tuning method based on a large language model includes:

[0018] S1. Collect and preprocess multi-source data to obtain standardized data with original data metadata. Specifically, this includes: S1.1. Configuring API, database, file system, and crawler interface to preset the collection frequency and batch size (default batch size: 10,000 records), implementing an incremental collection strategy to avoid duplicate collection, and using a distributed collection architecture for multi-source data collection, supporting concurrency of over 100. S1.2. Standardizing the multi-source data collected in S1.1 into a unified format while retaining the original data metadata; automatically detecting data encoding formats (UTF-8, GBK, etc.) and uniformly converting, recognizing, and parsing different formats (JSON, XML, CSV, TXT, etc.) to standardize them into a unified internal representation format while retaining the original data metadata (source, time, version, etc.). S1.3. Performing preliminary cleaning on the multi-source data after unifying the format in S1.2 to finally obtain standardized data with original data metadata. The preliminary cleaning in S1.3 includes: S1.3.1. Removing HTML tags, special symbols, and control characters. S1.3.2. Correcting obvious encoding errors and garbled characters. S1.33 Filter text of abnormal length (e.g., less than 10 characters or more than 100,000 characters). S1.34 Identify and mark potentially sensitive information.

[0019] S2. Constructing an intelligent quality assessment model based on a pre-trained large language model. Specifically, the construction of the intelligent quality assessment model includes: S2.1, using a pre-trained large language model as the foundation; S2.2, designing a quality assessment prompt template, with assessment dimensions including semantic coherence, information completeness, language fluency, factual accuracy, and domain relevance; among which, semantic coherence: checking the logical relationships of the context and identifying issues such as sentence breaks and jumps; information completeness: determining whether key information is missing; language fluency: assessing grammatical correctness and naturalness of expression; factual accuracy: cross-validating factual content through a knowledge base; domain relevance: calculating semantic similarity with the target domain. S2.3, setting the temperature parameter to 0.3 to ensure assessment stability; S2.4, setting Top-p to 0.9 to maintain a certain degree of assessment flexibility.

[0020] S3. The standardized data obtained in S1 is filtered using the intelligent quality assessment model constructed in S2 to obtain preliminary qualified data. Specifically, the filtering of standardized data includes: S3.1. The standardized data obtained in S1 is scored using the intelligent quality assessment model constructed in S2 based on semantic coherence, information completeness, language fluency, factual accuracy, and domain relevance. S3.2. Preset weight configuration and weighted average of the scores for semantic coherence, information completeness, language fluency, factual accuracy, and domain relevance; default weights: coherence 0.2, completeness 0.25, fluency 0.2, accuracy 0.25, relevance 0.1. S3.3. A preset quality threshold (default 70 points) is set to filter out preliminary qualified data that exceeds the threshold.

[0021] S4. Adaptive data filtering and classification are performed on the preliminary qualified data selected in S3 to obtain the preset proportion of data to be retained for each category based on its quality score. Specifically, adaptive data filtering and classification includes: S4.1. Automatically adjusting the quality threshold based on data distribution, using K-means clustering to identify quality levels and dynamically adjusting the retention proportion according to the target data volume. S4.2. Using a hierarchical classification system, a multi-label classifier is constructed to identify multiple attributes of the data, with a classification confidence threshold set to 0.8; categories include: task type, difficulty level, domain label, and data source, etc. S4.3. The MinHash algorithm is used to calculate similarity, with a similarity threshold set to 0.85, identifying near-duplicates and further refining deduplication based on edit distance, retaining the version with the highest quality score.

[0022] S5. The data retained in S4 is augmented according to a preset ratio. The newly generated augmented data is then filtered using the intelligent quality assessment model built in S2, retaining the preset ratio of augmented data. Specifically, the augmentation process includes: S5.1. Using a pre-trained large language model as the base. S5.2. Designing a data augmentation prompt: generating 3-5 similar but different variants for each original sample. Generation requirements include maintaining the core semantics, changing the expression and sentence structure, appropriately adding or adjusting details, and ensuring the generated text is natural and fluent. S5.3. Using temperature sampling to increase diversity, with a temperature of 0.7-0.9.

[0023] S6. Mix the original data retained in S4 and the augmented data in S5 according to a preset ratio (the default ratio of original data: augmented data = 7:3) to ensure the balance of data distribution and avoid the model overfitting the augmented data pattern. The final output is the optimized data containing the original data and augmented data mixed according to the preset ratio.

[0024] S7. Real-time monitoring of model performance trained using the optimized data output in S6. Based on the results, automatically adjust quality assessment weights, optimize data filtering thresholds, adjust data augmentation strategies, update domain relevance calculation methods, and employ reinforcement learning to optimize the overall strategy. Specifically, real-time monitoring records key metrics such as loss value, accuracy, and perplexity; sets performance baselines and warning thresholds; collects training logs and analyzes data utilization. It collects cases of model prediction errors, uses LLM to analyze error types, possible causes, and related training data issues, and generates error reports and improvement suggestions. Based on error analysis, it adjusts quality assessment weights, optimizes data filtering thresholds, adjusts data augmentation strategies, updates domain relevance calculation methods, and employs reinforcement learning to optimize the overall strategy (reward function based on model performance improvement).

[0025] Furthermore, distributed processing and system optimization.

[0026] Parallel Processing Architecture: Data processing is performed using the MapReduce paradigm, with data sharded into 1GB chunks. Parallelism is dynamically adjusted based on resources (maximum parallelism: 100). Message queues (such as Kafka) are used to coordinate tasks. Caching Mechanism: A multi-level caching system is established: memory cache, Redis cache, and disk cache. LLM inference results are cached to avoid redundant computations. The cache hit rate target is above 80%. The LRU strategy is used to manage cache space. Model Optimization: LLM is quantized (INT8 quantization) to reduce memory usage by 50%. Model parallelism and tensor parallelism techniques are used. Batch processing is optimized with dynamically adjustable batch size (16-128) to accelerate inference using TensorRT or ONNX Runtime.

[0027] like Figure 2 As shown, a data precision tuning system based on a large language model is used to perform data tuning using the methods described above. It comprises 23 core modules, divided into 6 layers:

[0028] Data input layer (modules 1-3): responsible for the access and preliminary processing of multi-source data.

[0029] Preprocessing layer (modules 4-6): performs data cleaning, encoding standardization, and metadata extraction.

[0030] LLM Evaluation Layer (Modules 7-9): Large Language Model Engine and related quality assessment and semantic analysis.

[0031] Intelligent processing layer (modules 10-12): Enables intelligent filtering, classification, and deduplication and noise reduction.

[0032] Data Augmentation and Feedback Layer (Modules 13-18): Data augmentation generation, quality control, and feedback optimization.

[0033] System support layer (modules 19-22): Provides basic support for distributed scheduling, caching, model optimization, etc.

[0034] Output layer (module 23): Outputs high-quality training data.

[0035] The modules are connected through data flow and control flow to form a complete data processing closed loop, in which the feedback system (module 16) and the LLM engine (module 7) form an optimization loop.

[0036] To make this application easier to understand, further explanation is provided below with reference to specific application methods.

[0037] Application scenario: An intelligent customer service system needs to be trained. 5 million historical customer service dialogue records have been collected. The data quality is uneven and requires precise optimization.

[0038] Specific implementation steps:

[0039] Data Acquisition and Preprocessing

[0040] Export conversation logs (JSON format) from the CRM system.

[0041] Batch processing, 10,000 records per batch, with a parallelism of 50.

[0042] The encoding is uniformly set to UTF-8, and HTML tags and special characters are removed.

[0043] De-identification: Identify and replace sensitive information such as account numbers and ID card numbers.

[0044] LLM quality assessment configuration

[0045] GPT-4 was selected as the evaluation model.

[0046] Customized assessment prompt for the financial sector:

[0047] Assess the quality of financial customer service conversations: - Accuracy of terminology usage (weight 0.3) - Completeness and resolution of responses (weight 0.3) - Standard and polite language (weight 0.2) - Compliance with regulations (weight 0.2)

[0048] The quality threshold is set to 75 points (higher than the general scenario).

[0049] Intelligent screening implementation

[0050] The data were categorized into four classes using K-means (K=4): excellent, good, average, and poor.

[0051] Retain the high-quality and good categories (approximately 3 million entries).

[0052] Deduplication: MinHash similarity threshold of 0.9, removing 150,000 duplicate entries.

[0053] Ultimately, 2.85 million high-quality conversations were retained.

[0054] Data augmentation processing

[0055] We selected the top 20% of high-quality conversations (570,000 conversations) for enhancement.

[0056] Example of generating a prompt:

[0057] Based on the following customer service dialogue, generate 3 variations with the same semantics but different expressions: Original: How can I waive my credit card annual fee? Requirement: Maintain financial professionalism and use different sentence structures.

[0058] 1.71 million augmented data entries were generated, and 1.37 million were retained after quality screening.

[0059] Implementation effect

[0060] Original data: 5 million records.

[0061] Optimized data: 2.85 million original records + 1.37 million enhanced records = 4.22 million records.

[0062] Model training results:

[0063] Intent recognition accuracy: 85.2% → 94.7% (an improvement of 9.5%).

[0064] Satisfaction rate: 72% → 88% (an increase of 16%).

[0065] Average response time: reduced by 35%.

[0066] Processing time: 4.5 hours (the original plan required 48 hours).

[0067] As can be seen from the above embodiments, the present invention can be effectively applied to data optimization tasks of different fields and scales, significantly improving data quality and processing efficiency, and providing high-quality data support for the training of large language models.

[0068] Appendix: Explanation of Technical Terms

[0069] LLM (Large Language Model): A deep learning language model with more than 1 billion parameters.

[0070] GPT (Generative Pre-trained Transformer): A series of language models developed by OpenAI.

[0071] LLaMA: An open-source large language model developed by Meta.

[0072] MinHash: Minimum hash, an algorithm for quickly estimating the similarity of sets.

[0073] K-means: K-means clustering algorithm, an unsupervised learning method.

[0074] INT8 quantization: Converts model parameters from 32-bit floating-point numbers to 8-bit integers, reducing model size.

[0075] QPS (Queries Per Second): Queries per second, a measure of system processing capacity.

[0076] P99 latency: 99% of the response time for requests, reflecting system stability.

[0077] Prompt: A prompt word used to guide the language model to generate specific outputs from input text.

[0078] MapReduce: A distributed computing programming paradigm used for large-scale data processing.

[0079] Kafka: A distributed stream processing platform used for high-throughput message passing.

[0080] Redis: An in-memory database used for high-speed caching.

[0081] Transformer: A neural network architecture based on self-attention mechanism.

[0082] Beam Search: A heuristic search algorithm.

[0083] Perplexity: A metric for measuring the quality of a language model.

[0084] Fine-tuning: Fine-tuning is the process of continuing to train a pre-trained model for a specific task.

Claims

1. A data-driven precision tuning method based on a large language model, characterized in that, include: S1. Collect multi-source data and preprocess it to obtain standardized data with original data element information; S2. Construct an intelligent quality assessment model based on a pre-trained large language model; S3. The standardized data obtained in S1 is screened using the intelligent quality assessment model constructed in S2 to obtain preliminary qualified data after screening. S4. Perform adaptive data filtering and classification on the preliminary qualified data selected in S3, and obtain the preset proportion of data to be retained for each category based on the quality score; S5. Enhance the data retained in S4 according to a preset ratio, and use the intelligent quality assessment model constructed in S2 to filter the newly generated enhanced data, retaining the enhanced data of the preset ratio. S6. Mix the original data retained in S4 and the enhanced data in S5 according to a preset ratio to ensure the balance of data distribution and avoid the model overfitting the enhanced data pattern. Finally, output the optimized data containing the original data and enhanced data mixed according to the preset ratio.

2. The data precision tuning method based on a large language model as described in claim 1, characterized in that, S1 includes: S1.1 Configure API, database, file system and crawler interface, and use a distributed collection architecture to collect multi-source data with preset collection frequency and batch size; S1.2, Standardize the multi-source data collected in S1.1 into a unified format and retain the original data's metadata; S1.

3. Perform preliminary cleaning on the multi-source data after unifying the format in S1.2 to finally obtain standardized data with original data element information.

3. The data precision tuning method based on a large language model as described in claim 2, characterized in that, The initial cleaning process in S1.3 includes: S1.

31. Remove HTML tags, special characters, and control characters; S1.32, Correct obvious encoding errors and garbled characters; S1.33, Filter text with abnormal length; S1.

34. Identify and mark potentially sensitive information.

4. The data precision tuning method based on a large language model as described in claim 2, characterized in that, The construction of the intelligent quality assessment model in S2 includes: S2.

1. Use a pre-trained large language model as the base; S2.2 Design a quality assessment prompt template, with assessment dimensions including semantic coherence, information completeness, language fluency, factual accuracy, and domain relevance; S2.3, The temperature parameter is set to 0.3 to ensure evaluation stability; S2.4 and Top-p are set to 0.9 to maintain a certain degree of evaluation flexibility.

5. The data precision tuning method based on a large language model as described in claim 4, characterized in that, The filtering of standardized data in S3 includes: S3.1 The intelligent quality assessment model constructed in S2 is used to score the standardized data obtained in S1 from the dimensions of semantic coherence, information completeness, language fluency, factual accuracy and domain relevance. S3.2 Preset weight configuration, weighted average scores for semantic coherence, information completeness, language fluency, factual accuracy, and domain relevance; S3.

3. Set a quality threshold and filter out preliminary qualified data that are above the threshold.

6. The data precision tuning method based on a large language model as described in claim 5, characterized in that, The adaptive data filtering and classification process in S4 includes: S4.1 Automatically adjust the quality threshold based on data distribution, use K-means clustering to identify quality levels and dynamically adjust the retention ratio according to the target data volume; S4.

2. A hierarchical classification system is adopted to construct a multi-label classifier to identify multiple attributes of the data, and the classification confidence threshold is set to 0.

8. S4.

3. Use the MinHash algorithm to calculate similarity, set the similarity threshold to 0.85, identify near duplicates, and further refine the deduplication based on the edit distance, retaining the version with the highest quality score.

7. The data precision tuning method based on a large language model as described in claim 6, characterized in that, The enhancement process in S5 includes: S5.1 Use a pre-trained large language model as the base; S5.2 Design a data augmentation prompt: generate 3-5 similar but different variants for each original sample. The generation requirements include keeping the core semantics unchanged, changing the expression and sentence structure, appropriately adding or adjusting details, and ensuring that the generated text is natural and fluent. S5.

3. Use temperature sampling to increase diversity, temperature = 0.7-0.

9.

8. The data precision tuning method based on a large language model as described in claim 1, characterized in that, Also includes: S7. Monitor the performance of the model trained using the optimized data output in S6 in real time, and automatically adjust the quality assessment weights, optimize the data screening threshold, adjust the data augmentation strategy, update the domain relevance calculation method, and adopt reinforcement learning to optimize the overall strategy based on the usage results.

Images