Large model-based information feedback file classification method, system, device and medium

Abstract

The application discloses a big model-based information feedback file classification method, system, device and medium, and belongs to the technical field of big models. The technical problem to be solved by the application is how to efficiently and quickly classify the content of feedback files, realize quick identification and accurate classification of the feedback files, and the technical scheme is as follows: information feedback file collection and data aggregation: information feedback files are obtained by a business related party, and an excel data aggregation file is formed in a set format; data desensitization: original description and category classification of the information feedback files are obtained according to the excel data aggregation file, and the original description is subjected to sensitive information replacement processing by a deepseek-V3 big model to obtain desensitized original description data; model fine-tuning: a Qwen2.5-14B model is fine-tuned in a multi-round dialogue format by a Lora mode, the Qwen2.5-14B model inputs the original description, and the Qwen2.5-14B model outputs a classification result; model deployment: model deployment and reasoning are performed by an artificial intelligence processor chip; and a classification result is generated.

Description

Technical Field

[0001] This invention relates to the field of large model technology, specifically to a method, system, device, and medium for classifying information feedback files based on large models. Background Technology

[0002] In the actual implementation of government work, due to various objective factors, the complexity of feedback documents and the diversification of scenarios are becoming increasingly serious. Traditional identification and classification methods require manual verification of each item, which can solve the problem, but its disadvantages are that it is time-consuming, labor-intensive, and inefficient.

[0003] Therefore, how to efficiently and quickly classify the content of feedback documents and achieve rapid identification and accurate classification of feedback documents is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] The technical objective of this invention is to provide a method, system, device, and medium for classifying information feedback files based on a large model, in order to solve the problem of how to efficiently and quickly classify the content of feedback files and achieve rapid identification and accurate classification of feedback files.

[0005] The technical objective of this invention is achieved as follows: a method for classifying information feedback files based on a large model, the specific method of which is as follows:

[0006] Information feedback document collection and data aggregation: Obtain information feedback documents from business stakeholders and generate an Excel data aggregation file in a set format. The content of the data aggregation Excel file includes the description of the information feedback document and the classification results by different dimensions.

[0007] Data anonymization: Based on the Excel data aggregation file, the original document description and category classification of the information feedback file are obtained. The original document description is then processed by the deepseek-V3 large model to replace sensitive information, resulting in anonymized original document description data. Among them, the category classification is used to describe the specific classification type of the information feedback file; the original document description is used to describe the specific content of the information feedback file.

[0008] Model fine-tuning: The Qwen2.5-14B model was fine-tuned using a multi-turn dialogue format via LoRa. The Qwen2.5-14B model takes an original description as input and outputs a classification result.

[0009] Model deployment: Model deployment and inference are performed using AI processor chips;

[0010] Generate classification results: Call the deployed large model to perform classification and obtain the corresponding classification results.

[0011] As a preferred option, the model fine-tuning is as follows:

[0012] Dataset preparation: Collect the dataset and divide it into training and validation sets;

[0013] Fine-tuning the Qwen2.5-14B model using the efficient fine-tuning framework LLaMA-Factory: When training the Qwen2.5-14B model using the LLaMA-Factory command-line tool, specify that training should only be performed on the CUDA device (i.e., GPU) numbered 0, and load the configuration file named train_qwen2.5-14b.yaml.

[0014] More preferably, the dataset adopts a JSON-formatted dialogue data structure to describe the dialogue flow of an information feedback file classification task; wherein, the dialogue data structure includes a JSON array of a single dialogue, and each element in the JSON array represents a dialogue record.

[0015] Ideally, when fine-tuning the Qwen2.5-14B model, the parameter configuration is as follows:

[0016] Model configuration: Configure the model name or path, maximum pixels per image, maximum number of pixels per frame, and remote code trust, and also configure LoRA parameters;

[0017] Dataset configuration: Configure data source parameters and preprocessing parameters; data source parameters include task type, data path and template format; preprocessing parameters include truncation length, maximum number of samples, cache control and parallel processing;

[0018] Output configuration: Configure the output directory for the save path, overwrite policy, and parameters related to the saved content;

[0019] Training parameter configuration: batch size per device, gradient accumulation steps, learning rate, number of training epochs, scheduler, and warm-up ratio;

[0020] Evaluation configuration: validation set partitioning and evaluation strategy; the evaluation strategy includes evaluation triggered by training steps, evaluation interval, and dynamic partitioning of the validation set.

[0021] As a preferred option, the model deployment is as follows:

[0022] Environment variable configuration: The source command loads environment variable scripts for multiple components in the AI ​​ecosystem of the AI ​​processor chip, ensuring that subsequent programs can correctly call relevant libraries and tools;

[0023] Start the AI ​​processor chip inference service: Start the AI ​​processor chip inference service through a daemon process to provide model inference capabilities.

[0024] More preferably, the environment variable configuration is as follows:

[0025] Basic toolkit: Loads environment variables for the AI ​​processor chip operator development toolkit, providing basic support for CANN (heterogeneous computing architecture);

[0026] Neural Network Operator Library: Configure the paths to the Neural Network Operator Library (NNAL) and Automatic Tuning Tool (ATB) for optimizing operator performance;

[0027] Pre-built model library: Sets the environment variables for the pre-compiled model library of the artificial intelligence processor chip, which includes pre-optimized operators or reference models;

[0028] MindIE Inference Service Framework: Loads the environment variables of the artificial intelligence processor chip, and configures the library paths and runtime parameters that the framework depends on.

[0029] More preferably, the specific steps for launching AI processor chip inference services are as follows:

[0030] Use the latest symbolic link to point to the latest version of the AI ​​processor chip service binary file, and deploy the program in the mindie-service / bin / directory;

[0031] Directly calling binary files will run in the background as a daemon process by default;

[0032] Once the service starts, it will continuously listen for inference requests and support concurrent inference for multiple models.

[0033] A large-model-based information feedback document classification system is provided to implement the large-model-based information feedback document classification method described above. The system includes:

[0034] The data collection and aggregation module is used to obtain information feedback documents from business stakeholders and generate an Excel data aggregation file in a set format. The content of the data aggregation Excel file includes a description of the information feedback document and classification results categorized by different dimensions.

[0035] The data anonymization module is used to obtain the original document description and category classification of the information feedback file based on the Excel data aggregation file, and to replace the sensitive information in the original document description using the deepseek-V3 large model to obtain the anonymized original document description data; among them, the category classification is used to describe the specific classification type of the information feedback file; the original document description is used to describe the specific content of the information feedback file.

[0036] The model fine-tuning module is used to fine-tune the Qwen2.5-14B model using a multi-turn dialogue format via LoRa. The Qwen2.5-14B model takes an original description as input and outputs a classification result.

[0037] The model deployment module is used for model deployment and inference via an artificial intelligence processor chip.

[0038] The classification result generation module is used to call the deployed large model for classification and obtain the corresponding classification results.

[0039] Qwen2.5 is an open-source large-scale model from the Tongyi Qianwen series developed by Alibaba Cloud. This series offers multiple versions and scales of open-source models, including Base and Instruct, to meet diverse computational needs. Compared to Qwen2, the new version has achieved significant improvements in several aspects, including knowledge acquisition, programming capabilities, mathematical capabilities, and instruction execution. It also inherits Qwen2's powerful long text processing capabilities, supporting context lengths of up to 128K tokens, offering a significant advantage in handling text classification tasks.

[0040] Lora is a novel fine-tuning method that introduces an additional linear layer into a pre-trained model and fine-tunes this linear layer using task-specific training data. This approach allows the model to better adapt to the specific task while reducing over-reliance on the initial model and overfitting issues.

[0041] An electronic device includes: a memory and at least one processor;

[0042] The memory contains computer programs;

[0043] The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the information feedback file classification method based on the large model as described above.

[0044] A computer-readable storage medium storing a computer program that can be executed by a processor to implement the large-model-based information feedback file classification method described above.

[0045] The information feedback file classification method, system, device, and medium based on a large model of the present invention have the following advantages:

[0046] (i) This invention improves the efficiency of information feedback file processing by data collection, data desensitization and cleaning, model fine-tuning and classification result generation.

[0047] (II) This invention integrates information and intelligent technologies to achieve rapid identification and accurate analysis of information feedback files. At the same time, compared with the same series of large models, tests have shown that under the same prompt words, the trained Qwen2.5-14B and the untrained Qwen2.5-72B have similar accuracy rates. Attached Figure Description

[0048] The invention will be further described below with reference to the accompanying drawings.

[0049] Appendix Figure 1 This is a flowchart of a large-model-based information feedback file classification method. Detailed Implementation

[0050] The following detailed description of the information feedback file classification method, system, device, and medium based on a large model of the present invention is provided with reference to the accompanying drawings and specific embodiments.

[0051] Example 1:

[0052] As attached Figure 1 As shown in the figure, this embodiment provides a method for classifying information feedback files based on a large model. The method is as follows:

[0053] S1. Information feedback document collection and data aggregation: Obtain information feedback documents from business stakeholders and form an Excel data aggregation file in a set format. The content of the data aggregation Excel file includes the description of the information feedback document and the classification results through different dimensions.

[0054] S2. Data Desensitization: Based on the Excel data aggregation file, obtain the original document description and category classification of the information feedback file, and then use the deepseek-V3 large model to replace sensitive information in the original document description to obtain the desensitized original document description data; where the category classification is used to describe the specific classification type of the information feedback file; the original document description is used to describe the specific content of the information feedback file.

[0055] S3. Model fine-tuning: The Qwen2.5-14B model is fine-tuned using a multi-turn dialogue format via LoRa. The Qwen2.5-14B model takes the original description as input and outputs the classification result.

[0056] S4. Model Deployment: Model deployment and inference are performed using artificial intelligence processor chips;

[0057] S5. Generate classification results: Call the deployed large model to perform classification and obtain the corresponding classification results.

[0058] The specific model fine-tuning in step S3 of this embodiment is as follows:

[0059] S301. Dataset preparation: Collect the dataset and divide it into training set and validation set;

[0060] S302. Fine-tuning the Qwen2.5-14B model using the efficient fine-tuning framework LLaMA-Factory: When training the Qwen2.5-14B model using the LLaMA-Factory command-line tool, specify that only the CUDA device numbered 0 (i.e., GPU) should be used for training, and load the configuration file named train_qwen2.5-14b.yaml.

[0061] In this embodiment, the dataset in step S301 adopts a JSON-formatted dialogue data structure to describe the dialogue flow of an information feedback file classification task; wherein, the dialogue data structure includes a JSON array of a single dialogue, and each element in the JSON array represents a dialogue record.

[0062] In step S302 of this embodiment, when fine-tuning the Qwen2.5-14B model, the parameters are configured as follows:

[0063] ① Model configuration: Configure the model name or path, maximum pixels per image, maximum number of pixels per frame, and remote code trust, and configure LoRA parameters;

[0064] ② Dataset configuration: Configure data source parameters and preprocessing parameters; data source parameters include task type, data path and template format; preprocessing parameters include truncation length, maximum number of samples, cache control and parallel processing;

[0065] ③ Output configuration: Configure the output directory for the save path, the overwrite policy, and the relevant parameters for the saved content;

[0066] ④ Training parameter configuration: batch size per device, gradient accumulation steps, learning rate, number of training rounds, scheduler, and warm-up ratio;

[0067] ⑤ Evaluation configuration: validation set partitioning and evaluation strategy; the evaluation strategy includes evaluation triggered by training steps, evaluation interval, and dynamic partitioning of the validation set.

[0068] The model deployment in step S4 of this embodiment is as follows:

[0069] S401, Environment Variable Configuration: Load the environment variable scripts of multiple components in the AI ​​ecosystem of the artificial intelligence processor chip through the source command to ensure that subsequent programs can correctly call the relevant libraries and tools;

[0070] S402. Start AI processor chip inference service: Start the AI ​​processor chip inference service through a daemon process to provide model inference capabilities.

[0071] The environment variable configuration in step S401 of this embodiment is as follows:

[0072] S40101, Basic Toolkit: Loads the environment variables for the AI ​​processor chip operator development toolkit, providing basic support for CANN (Heterogeneous Computing Architecture);

[0073] S40102, Neural Network Operator Library: Configure the paths to the Neural Network Operator Library (NNAL) and Automatic Tuning Tool (ATB) for optimizing operator performance;

[0074] S40103, Pre-built model library: Sets the environment variables for the pre-compiled model library of the artificial intelligence processor chip, which includes pre-optimized operators or reference models;

[0075] S40104, MindIE Inference Service Framework: Loads the environment variables of the artificial intelligence processor chip, and configures the library paths and runtime parameters that the framework depends on.

[0076] The specific steps for starting the AI ​​processor chip inference service in step S402 of this embodiment are as follows:

[0077] S40201. Use the latest symbolic link to point to the latest version of the AI ​​processor chip service binary file, and deploy the program in the mindie-service / bin / directory;

[0078] S40202: Directly calls binary files, running in the background by default in daemon mode;

[0079] S40203: After the service starts, it will continuously listen for inference requests and support concurrent inference for multiple models.

[0080] Example 2:

[0081] This embodiment provides a large-model-based information feedback file classification system, which is used to implement the large-model-based information feedback file classification method as described in Embodiment 1; the system includes:

[0082] The data collection and aggregation module is used to obtain information feedback documents from business stakeholders and generate an Excel data aggregation file in a set format. The content of the data aggregation Excel file includes a description of the information feedback document and classification results categorized by different dimensions.

[0083] The data anonymization module is used to obtain the original document description and category classification of the information feedback file based on the Excel data aggregation file, and to replace the sensitive information in the original document description using the deepseek-V3 large model to obtain the anonymized original document description data; among them, the category classification is used to describe the specific classification type of the information feedback file; the original document description is used to describe the specific content of the information feedback file.

[0084] The model fine-tuning module is used to fine-tune the Qwen2.5-14B model using a multi-turn dialogue format via LoRa. The Qwen2.5-14B model takes an original description as input and outputs a classification result.

[0085] The model deployment module is used for model deployment and inference via an artificial intelligence processor chip.

[0086] The classification result generation module is used to call the deployed large model for classification and obtain the corresponding classification results.

[0087] The working process of this system is as follows:

[0088] (1) Data Collection and Summary: The data sources for the feedback documents are relatively singular, mainly consisting of manually categorized feedback documents provided by business stakeholders. These are compiled into Excel files in a specific format, and the main content includes document descriptions and categorization results based on different dimensions.

[0089] (2) Data anonymization: This mainly includes the content and classification of information feedback documents, and mainly includes the following structural information:

[0090]

[0091]

[0092] The original descriptions in the above information were processed using a large model to replace sensitive information. The large model used was DeepSeek-V3, and the prompts are as follows:

[0093] You are a text anonymization engineer, and your task is to anonymize the input text.

[0094] The desensitization requirements are as follows:

[0095] ① For all names, phone numbers, identity information, addresses, work units, residential communities, etc., use information that actually exists in reality to replace them. For addresses, ensure that the address before and after de-identification are not in the same province. Do not use xxxx or number 123 for de-identification.

[0096] ② For the same person's name, phone number, identity information, address, and organization information, use the same result for de-identification to ensure relevance;

[0097] ③ Strictly follow the source document content format and only output the desensitized text. Do not output any additional content. (3) Model fine-tuning: Use the LoRa method to fine-tune Qwen2.5-14B. Qwen2.5-14B recommends using the multi-turn dialogue format for fine-tuning. Use the original document description as input and the classification result as output. The model fine-tuning steps are as follows:

[0098] ① Dataset preparation: Prepare training and validation sets, with the following data format:

[0099]

[0100]

[0101] ② Use the efficient fine-tuning framework LLaMA-Factory for fine-tuning. Execute the following command to fine-tune the model:

[0102] CUDA_VISIBLE_DEVICES=0llamafactory-cli train train_qwen2.5-14b.yaml

[0103] The corresponding configuration parameters are as follows:

[0104] #model

[0105] model_name_or_path:Qwen / Qwen2.5 -14B-Instruct

[0106] image_max_pixels:262144

[0107] video_max_pixels:16384

[0108] trust_remote_code:true

[0109] #method

[0110] stage:sft

[0111] do_train:true

[0112] finetuning_type:lora

[0113] lora_rank:64

[0114] lora_alpha:128

[0115] lora_dropout:0.1

[0116] lora_target:all

[0117] #dataset

[0118] dataset:pokemon_multimodal

[0119] dataset_dir:pokemon_sharegpt

[0120] template:qwen2_vl

[0121] cutoff_len:4096

[0122] max_samples:1000

[0123] overwrite_cache:true

[0124] preprocessing_num_workers:16

[0125] dataloader_num_workers:4

[0126] #output

[0127] output_dir:saves / qwen2.5-14b-lora-pokemon / sft-cli

[0128] logging_steps:10

[0129] save_steps:500

[0130] plot_loss:true

[0131] overwrite_output_dir:true

[0132] save_only_model:false

[0133] report_to:none

[0134] #train

[0135] per_device_train_batch_size:2

[0136] gradient_accumulation_steps:8

[0137] learning_rate: 2.0e-4

[0138] num_train_epochs:3.0

[0139] lr_scheduler_type:cosine

[0140] warmup_ratio: 0.1

[0141] bf16:true

[0142] ddp_timeout:180000000

[0143] resume_from_checkpoint:null

[0144] #eval

[0145] val_size:0.1

[0146] per_device_eval_batch_size:2

[0147] eval_strategy:steps

[0148] eval_steps:100.

[0149] (4) Model deployment: The Ascend 910B graphics card is used for model deployment and inference.

[0150] Import environment variables:

[0151] source / usr / local / Ascend / ascend-toolkit / set_env.sh

[0152] source / usr / local / Ascend / nnal / atb / set_env.sh

[0153] source / usr / local / Ascend / atb-models / set_env.sh

[0154] source / usr / local / Ascend / mindie / set_env.sh;

[0155] MindIE inference service is running:

[0156] / usr / local / Ascend / mindie / latest / mindie-service / bin / mindieservice_daemon;

[0157] (5) Classification result generation: Call the deployed large model for classification.

[0158] Example 3:

[0159] This embodiment also provides an electronic device, including: a memory and a processor;

[0160] The memory stores the instructions executed by the computer.

[0161] The processor executes computer execution instructions stored in the memory, causing the processor to perform the information feedback file classification method based on a large model in any embodiment of the present invention.

[0162] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0163] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0164] Example 4:

[0165] This embodiment also provides a computer-readable storage medium storing multiple instructions, which are loaded by a processor to cause the processor to execute the information feedback file classification method based on a large model according to any embodiment of the present invention. Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or apparatus may read and execute the program code stored in the storage medium.

[0166] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0167] Storage media embodiments for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RYM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0168] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0169] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for classifying information feedback files based on a large model, characterized in that, The method is as follows: Information feedback document collection and data aggregation: Obtain information feedback documents from business stakeholders and generate an Excel data aggregation file in a set format. The content of the data aggregation Excel file includes the description of the information feedback document and the classification results by different dimensions. Data anonymization: Based on the Excel data aggregation file, the original document description and category classification of the information feedback file are obtained. The original document description is then processed by the deepseek-V3 large model to replace sensitive information, resulting in anonymized original document description data. Among them, the category classification is used to describe the specific classification type of the information feedback file; the original document description is used to describe the specific content of the information feedback file. Model fine-tuning: The Qwen2.5-14B model was fine-tuned using a multi-turn dialogue format via LoRa. The Qwen2.5-14B model takes an original description as input and outputs a classification result. Model deployment: Model deployment and inference are performed using AI processor chips; Generate classification results: Call the deployed large model to perform classification and obtain the corresponding classification results; The specific steps for model fine-tuning are as follows: Dataset preparation: Collect the dataset and divide it into training and validation sets; Fine-tuning the Qwen2.5-14B model using the efficient fine-tuning framework LLaMA-Factory: When training the Qwen2.5-14B model using the LLaMA-Factory command-line tool, specify that training should only be performed on the CUDA device numbered 0, and load the configuration file named train_qwen2.5-14b.yaml; The dataset uses a JSON-formatted dialogue data structure to describe the dialogue flow of an information feedback file classification task; the dialogue data structure consists of a JSON array of a single dialogue, where each element in the JSON array represents a dialogue record. When fine-tuning the Qwen2.5-14B model, the parameter configuration is as follows: Model configuration: Configure the model name or path, maximum pixels per image, maximum number of pixels per frame, and remote code trust, and also configure LoRA parameters; Dataset configuration: Configure data source parameters and preprocessing parameters; data source parameters include task type, data path and template format; preprocessing parameters include truncation length, maximum number of samples, cache control and parallel processing; Output configuration: Configure the output directory for the save path, overwrite policy, and parameters related to the saved content; Training parameter configuration: batch size per device, gradient accumulation steps, learning rate, number of training epochs, scheduler, and warm-up ratio; Evaluation configuration: validation set partitioning and evaluation strategy; the evaluation strategy includes evaluation triggered by training steps, evaluation interval, and dynamic partitioning of the validation set; The model deployment details are as follows: Environment variable configuration: The source command loads environment variable scripts for multiple components in the AI ​​ecosystem of the AI ​​processor chip, ensuring that subsequent programs can correctly call relevant libraries and tools; Start the AI ​​processor chip inference service: Start the AI ​​processor chip inference service through a daemon process to provide model inference capabilities; The environment variable configuration is as follows: Basic toolkit: Loads the environment variables of the AI ​​processor chip operator development toolkit, providing basic CANN support; Neural network operator library: Configures the paths to the neural network operator library and automatic tuning tools, used to optimize operator performance; Pre-built model library: Sets the environment variables for the pre-compiled model library of the artificial intelligence processor chip, which includes pre-optimized operators or reference models; MindIE Inference Service Framework: Loads the environment variables of the artificial intelligence processor chip, and configures the library paths and runtime parameters that the framework depends on; The specific steps for activating the AI ​​processor chip inference service are as follows: Use the latest symbolic link to point to the latest version of the AI ​​processor chip service binary file, and deploy the program in the mindie-service / bin / directory; Directly calling binary files will run in the background as a daemon process by default; Once the service starts, it will continuously listen for inference requests and support concurrent inference for multiple models.

2. A large-scale model-based information feedback document classification system, characterized in that, This system is used to implement the information feedback file classification method based on a large model as described in claim 1; the system includes: The data collection and aggregation module is used to obtain information feedback documents from business stakeholders and generate an Excel data aggregation file in a set format. The content of the data aggregation Excel file includes a description of the information feedback document and classification results categorized by different dimensions. The data anonymization module is used to obtain the original document description and category classification of the information feedback file based on the Excel data aggregation file, and to replace the sensitive information in the original document description using the deepseek-V3 large model to obtain the anonymized original document description data; among them, the category classification is used to describe the specific classification type of the information feedback file; the original document description is used to describe the specific content of the information feedback file. The model fine-tuning module is used to fine-tune the Qwen2.5-14B model using a multi-turn dialogue format via LoRa. The Qwen2.5-14B model takes an original description as input and outputs a classification result. The model deployment module is used for model deployment and inference via an artificial intelligence processor chip. The classification result generation module is used to call the deployed large model for classification and obtain the corresponding classification results.

3. An electronic device, characterized in that, include: Memory and at least one processor; The memory contains computer programs; The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the information feedback file classification method based on a large model as described in claim 1.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement the information feedback file classification method based on a large model as described in claim 1.

Images