Industrial injury grade identification method and system based on large model reinforcement learning

Abstract

The invention provides an industrial injury and disability grade identification method and system based on large model reinforcement learning, and relates to the field of industrial injury and disability grade identification, and the method comprises the steps: obtaining the injury condition description of a user; searching a plurality of previous disease conditions and grade conclusions with the highest similarity with the injury condition description from an injury condition case knowledge base, and constructing an instruction prompt word transmitted to the large model in combination with a prompt prompt word engineering template; based on the instruction cue word, generating a judgment conclusion through the large model after reinforcement learning; the judgment conclusion is analyzed, and the final industrial injury and disability grade is obtained; wherein the large model carries out unsupervised learning through a pre-obtained industrial injury grade identification standard, and then carries out supervised reinforcement learning by using the training data set. According to the method, a large model technology and an industrial injury and disability grade identification scene are deeply fused, double improvement of the working efficiency and transparency of industrial injury and disability identification is realized, and identification work is powerfully promoted to be transformed and upgraded to a scientific and intelligent direction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of work injury disability grade identification, in particular to a work injury disability grade identification method and system based on large model reinforcement learning. BACKGROUND

[0002] At present, a relatively complete work injury disability grade evaluation technical system has been established around the "Labor Capacity Identification Worker Work Injury and Occupational Disease Disability Grade (GB / T 16180-2014)" standard. This system systematically stipulates the grading principles, clinical judgment standards and comprehensive evaluation methods from grade 1 to grade 10 disability, providing a scientific and authoritative technical basis for work injury identification work. However, there are still double challenges in actual operation: on the one hand, due to the large number of medical professional items covered by the standard and the complex judgment conditions, it is necessary to conduct comprehensive evaluation by professional personnel with evaluation qualifications, which puts forward higher requirements for the allocation of expert resources; on the other hand, the professional terms and medical concepts in the grading standard are difficult for non-professionals to understand, which makes it difficult for workers to accurately grasp the evaluation basis when participating in the identification process, affecting their cognition and acceptance of the identification results.

[0003] In recent years, large language model technology represented by Transformer architecture (such as GPT, Llama, DeepSeek, etc.) has made breakthrough progress in natural language understanding, multi-modal data analysis and other fields. These models have shown strong semantic analysis capabilities and knowledge reasoning levels, and have been successfully applied to medical report structured processing, legal document key information extraction and other professional scenarios; in particular, the technical progress of large models in long context analysis, reasoning and other aspects provides new technical support for work injury disability grade identification.

[0004] The open source base large model has the problem of knowledge lag after the standard is updated, which leads to the lag of the existing work injury disability grade identification technology based on large models, and the evaluation technical system of work injury disability grade also has lag, and the open source base large model has limited understanding of the provisions in the national standard, and has not been aligned with the expert evaluation method, which ultimately affects the accuracy of the large model in work injury disability grade determination. SUMMARY

[0005] In order to solve the above problems, the present application proposes a work injury disability grade identification method and system based on large model reinforcement learning, which deeply integrates large model technology with work injury disability grade identification scenarios, realizes the double improvement of work injury disability identification work efficiency and transparency, and strongly promotes the transformation and upgrading of identification work towards scientific and intelligent direction.

[0006] According to some embodiments, the present application adopts the following technical solutions: A work injury disability grade identification method based on large model reinforcement learning, comprising: obtain a description of an injury of a user; retrieve a plurality of past injury cases and grade conclusions with the highest similarity to the injury description from an injury case knowledge base, and construct instruction prompts for a large model in combination with prompt engineering templates; generate a determination conclusion based on the instruction prompts through the large model after reinforcement learning; analyze the determination conclusion to obtain a final work injury disability grade; The large model is first subjected to unsupervised learning through a pre-obtained work injury disability grade appraisal standard, and then subjected to supervised reinforcement learning through a training data set.

[0007] According to some embodiments, the present application adopts the following technical solutions: A work injury disability grade appraisal system based on large model reinforcement learning, comprising: An obtaining module configured to obtain a description of an injury of a user; A constructing module configured to retrieve a plurality of past injury cases and grade conclusions with the highest similarity to the injury description from an injury case knowledge base, and construct instruction prompts for a large model in combination with prompt engineering templates; A generating module configured to generate a determination conclusion based on the instruction prompts through the large model after reinforcement learning; An analyzing module configured to analyze the determination conclusion to obtain a final work injury disability grade; The large model is first subjected to unsupervised learning through a pre-obtained work injury disability grade appraisal standard, and then subjected to supervised reinforcement learning through a training data set.

[0008] According to some embodiments, the present application adopts the following technical solutions: A computer program product comprising a computer program, which, when executed by a processor, implements the work injury disability grade appraisal method based on large model reinforcement learning.

[0009] According to some embodiments, the present application adopts the following technical solutions: A non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implements the work injury disability grade appraisal method based on large model reinforcement learning.

[0010] According to some embodiments, the present application adopts the following technical solutions: An electronic device comprises a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the method for implementing the large model-based reinforcement learning of work injury and disability grade identification.

[0011] Compared with the prior art, the beneficial effects of the present application are: The present application constructs an intelligent auxiliary evaluation method by deeply integrating large model technology with work injury and disability grade identification scenarios. Based on the work injury and disability grade identification standard, the method combines specific injury descriptions, utilizes the capabilities of large models to realize automatic grade determination, and simultaneously generates complete reasoning logic chains and evaluation conclusions. The method can effectively optimize the identification workflow and significantly reduce the workload of manual review. Through intelligent reasoning functions, the evaluation basis is visually displayed to the laborers in a visual manner, greatly improving the understandability of the results, achieving a dual improvement in work injury and disability identification efficiency and transparency, and effectively promoting the transformation and upgrading of identification work towards scientific and intelligent direction.

[0012] In the present application, by combining continued pre-training with reinforcement learning, the national standard and expert judgment can be aligned, effectively improving the identification accuracy. In the present application, through the work injury and injury knowledge base platform, users can be provided with one-key viewing of reference cases, judgment level reasoning and conclusions, and judgment basis, reducing the human interpretation cost and operation complexity. In the present application, the large model identification conclusion and the expert identification conclusion can be compared to find abnormalities in advance and provide warning, reduce the influence of subjective factors, and provide assistance and supervision to the identification process. In the present application, the large model can be optimized and updated according to user feedback, and the latest identification standards and basis can be synchronized, so that the system can be upgraded without user awareness. BRIEF DESCRIPTION OF DRAWINGS

[0013] The drawings accompanying the specification of the present application form part of the present application and serve to provide further understanding of the present application, and the illustrative embodiments of the present application and their descriptions serve to explain the present application, and do not constitute an improper limitation of the present application.

[0014] Figure 1 It is a method flow architecture diagram of example 1.

[0015] Figure 2 It is a work injury and disability identification case data set construction flowchart of example 1.

[0016] Figure 3 It is a continued pre-training flowchart of example 1.

[0017] Figure 4 It is a work injury and injury knowledge base platform overall flowchart of example 1.

[0018] Figure 5 Optimization flowchart for positive and negative feedback model of example 1. DETAILED DESCRIPTION

[0019] The application will be further described below in conjunction with the accompanying drawings and examples.

[0020] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0021] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be understood that when the terms "comprise" and / or "comprising" are used in the specification, there is a feature, step, operation, device, component and / or combination thereof.

[0022] Term explanation: Transformers: a deep learning model with a core self-attention mechanism.

[0023] LLM: Large Language Model, a large-scale pre-training language model based on deep learning.

[0024] LSH: Locality-Sensitive Hashing, an approximate search technique for fast search of similar items in high-dimensional space.

[0025] Prompt: large model prompt, a text instruction or question input by the user, used to guide the large model to generate specific output.

[0026] Few-shot prompting: by providing a small number of examples (usually a few to tens) to the machine learning model, to help the model understand the requirements and expected output of the specific task.

[0027] RL: Reinforcement Learning, a method of policy optimization according to reward signals.

[0028] PPO: Proximal Policy Optimization, a commonly used reinforcement learning algorithm, designed to solve the instability and computational complexity of traditional policy gradient methods in the training process.

[0029] Embodiment 1 In an embodiment of the present application, a work injury disability grade identification method based on large model reinforcement learning is provided, comprising: Step S1: obtaining the injury description of the user; Step S2: retrieving a plurality of past conditions and grade conclusions with the highest similarity to the injury description from the injury case knowledge base, and combining the prompt prompt word engineering template to construct the instruction prompt word delivered to the large model; Step S3: generating a determination conclusion based on the instruction prompt word through the large model after reinforcement learning; Step S4: analyzing the determination conclusion to obtain the final work injury disability grade; Among them, the large model first performs unsupervised learning through the pre-obtained work injury disability identification standard, and then performs supervised reinforcement learning with the training data set.

[0030] As an embodiment, in view of the knowledge lag problem of the existing open source base large model after the standard is updated, a standard knowledge updating method based on continued pre-training is used to realize the rapid updating of the revised content of the national standard such as “Labor Capacity Identification, Worker Injury and Occupational Disease Disability Grade (GB / T 16180-2014) ”, and ensure that the large model always makes work injury disability grade determination based on the latest standard.

[0031] In view of the problem of insufficient accuracy of the large model in work injury disability grade determination, this embodiment proposes a hybrid training paradigm that combines pre-training and reinforcement learning. By constructing a high-quality work injury case data set with professional annotations, an efficient and convenient reward function is designed to guide the model to optimize the reasoning path during decision-making. At the same time, a positive and negative feedback mechanism is introduced, so that the model can continuously learn from expert correction opinions, align with national standards and expert determination conclusions, and gradually improve the accuracy of determination.

[0032] Meanwhile, based on the optimized large model, this embodiment also constructs a work injury condition knowledge base platform. After the user inputs the injury description or uploads the injury form file, the platform will refer to the past grade determination conclusions retrieved from the injury case knowledge base, output the reasoning process, grade determination conclusion and determination comparison table. For cases where the conclusions of artificial and large models are different, they are marked as important. Through the intelligent auxiliary function of the platform, the dependence on traditional work identification institutions and professionals can be significantly reduced, and the human and economic cost investment can be reduced.

[0033] The main process architecture of this embodiment is shown in the accompanying Figure 1 , mainly including three big steps: I. Data preprocessing and model training method, including: An injury case knowledge base is constructed by downloading injury cases from the work injury assessment data platform. This knowledge base supports updates; when new cases appear, they are directly added to the knowledge base to facilitate rapid retrieval and query capabilities. In this embodiment, the injury case is an Excel spreadsheet containing data such as injury category, injury description, assessment basis, disability level, nursing level, assessment type, and application time. Data preprocessing is required to meet the needs of subsequent training. This embodiment can automate the uploading and processing of past work injury assessment cases; specific steps are detailed in the appendix. Figure 2 ① Delete invalid data with empty injury descriptions and disability levels in the table, and replace all Arabic numerals for disability levels with Chinese numerical descriptions; ② Extract the training content for each case in the table, mainly including: injury description, assessment basis, nursing level and disability level, and verify whether the assessment basis and disability level match. ③ Use the LSH (Local Sensitive Hash) algorithm to filter out data with high similarity in all extracted cases, reduce the repetition rate of samples, and obtain a more balanced distribution of disability levels; ④ The cleaned, high-quality text dataset is passed to the deepseek-v3 API interface, and prompt words are constructed using few-shot prompting to form a dataset of work injury disability assessment cases that can be used for reinforcement learning.

[0034] The dataset is in question-answer pair format, as follows: [ { Question: "The patient's diagnosis is: lumbar sprain. Examination reveals: acute lumbar sprain, L3 / 4 disc herniation. Please analyze the final disability level based on the criteria for determining disability levels according to the 'Disability Grading Standards for Work-Related Injuries and Occupational Diseases'." “answer”: “Based on Article 4 of Level 10 and the patient’s injuries, the patient’s disability level is determined to be Level 10.” } { Question: "The patient's diagnosis is: right ulna fracture. Examination revealed: more than 4 months post-operative internal fixation for the right ulna fracture; old surgical scar on the right forearm, healing is good, and mobility is fair (internal fixation device not removed). Please analyze the final disability level based on the criteria for determining disability levels according to the 'Disability Grading Standards for Work-Related Injuries and Occupational Diseases'." Answer: "Based on Article 23 of Level 9 and the patient's injuries, the patient's disability level is determined to be level nine." ...... ] In building a large model training dataset, the time node of the data needs to be clear to ensure the model's capture of timely knowledge. For the field adaptation needs of disability rating identification, based on the Labor Capacity Identification, Worker's Industrial Injury and Occupational Disease Disability Rating (GB / T 16180-2014) national standard, the content is structured and standardized to generate formatted data that meets the requirements of unsupervised training as follows: { " text ": " In the Labor Capacity Identification, Worker's Industrial Injury and Occupational Disease Disability Rating, the case that meets the 4th clause of the 10th level is an industrial injury of the 10th level: acute trauma leading to herniation of the intervertebral disc nucleus pulposus with nerve irritation signs." } { " text ": " In the Labor Capacity Identification, Worker's Industrial Injury and Occupational Disease Disability Rating, the case that meets the 23rd clause of the 9th level is an industrial injury of the 9th level: long tubular bone fracture internal fixation or external fixation bracket postoperative." } }...... ] Unsupervised data is input into the large model for continued pre-training network (Continued Pre-training), allowing the base large model to learn the industrial injury disability rating principles, rating clauses, and the determination criteria and explanations in the appendix, increasing understanding of the national standard and reducing dependence on labeled data. The continued pre-training process is shown in the following Figure 3 : ① By organizing the specialized terms in the Labor Capacity Identification, Worker's Industrial Injury and Occupational Disease Disability Rating (GB / T 16180-2014) national standard, a new vocabulary is obtained, including medical dependence, fixation, self-care disorders, and similar professional vocabulary.

[0035] ② Add the new vocabulary to the original segmenter to form a new segmenter, making the segmentation results more accurately reflect the actual meaning of the industrial injury scene; ③ Load and segment the unsupervised industrial injury rating data obtained in step 2, load the base large model or the checkpoint model parameters obtained in the original training.

[0036] ④ Set the hyperparameters, taking DeepSeek-R1-Distill-Qwen-1.5B as the base large model, and setting some key parameters as follows: Training rounds (epoch), set to 2; Batch size per GPU (batch size), set according to GPU memory, default setting is 4; learning rate, set to 3e-4. During the training process, the learning rate is gradually reduced according to the cosine function form to the initial setting of 0.1; gradient clipping value, set to 1.0; whether to use flash attention optimization technology (use flash attention), set to False. This technology helps to improve the operation speed while saving the use of video memory, but requires additional installation of the corresponding calculation package; tensor parallel, set to 4, indicating that the size of tensor parallel is set to 4; pipeline parallel, set to 2, indicating that the size of pipeline parallel is set to 2; sequence parallel, set to 1, indicating that the size of sequence parallel is set to 1; ⑤ Traverse the training data, perform forward propagation and loss calculation, accumulate and clip the gradient, and update the parameters of the large model using backpropagation.

[0037] ⑥ Every 500 steps, save a checkpoint. After 2 rounds of data traversal, the final pre-trained large model containing work injury disability grading knowledge is completed.

[0038] Put the supervised work identification case training set constructed in step 1 into the reinforcement learning network, and perform iterative optimization of policy gradient through the following process to guide the inference strategy of the large model after continuing pre-training to update in the direction of higher reward for grade identification conclusions: ① Trajectory sampling: the large model samples each question in the training data set according to the current strategy, and updates the old strategy Generate a trajectory containing state, action, and reward; ② Reward calculation: adopt the method of discounted cumulative reward to calculate the reward of each trajectory. For a given time step t , the return is defined as the weighted sum of all future rewards starting from t :

[0039] where is the discount factor, the weight is the power of the discount factor, and the importance of the long-term reward gradually decreases; represents the reward at time step kThe immediate reward obtained, defined as follows:

[0040] where, is an indicator function that takes value 1 when token is the end-of-sequence token, and 0 otherwise; is the input prompt x and the full sequence of generated responses y corresponding to the reward; is a hyper-parameter that controls the strength of the KL divergence penalty. KL( t ) is a token-level divergence penalty function, which is calculated as:

[0041] where, is the probability of the RL model to generate token at state . is the probability of the SFT model to generate token at state .

[0042] ③ Policy Gradient Estimation: The policy gradient is calculated using the Monte Carlo method to estimate the gradient of the expected reward with respect to the policy parameters :

[0043] where, denotes the expected operator, denotes the policy function; ④ Policy Update: The policy parameters are updated by the gradient ascent method:

[0044] where, is the training learning rate. The update of the policy is limited by the clipping mechanism of PPO to prevent the extreme reward from causing excessive impact on the model update. The advantage function is as follows:

[0045] The advantages are standardized by z-score normalization:

[0046] where, and The average and standard deviation of the batch are represented by and, respectively. The influence of outliers in individual prompts is suppressed using the advantage-normalized stable gradient estimate. By dividing the training data into multiple small batches and performing parameter updates multiple times on each small batch, the large model converges faster while reducing the consumption of video memory.

[0047] II. Using the constructed work injury condition knowledge base platform to conduct work injury disability grade identification, including: The overall process of the work injury condition knowledge base platform is shown in the attached Figure 4 The platform supports user input of injury condition descriptions or uploading of injury condition table files; according to regular expressions, the user's injury condition table file is parsed to obtain one or more user injury condition descriptions, which are recorded and input to the injury condition analysis interface respectively; after obtaining the injury condition description, firstly search in the injury condition case knowledge base, return the top 3 similar past conditions and grade conclusions. Combine the prompt prompt word engineering template to construct the text passed to the large model, the standard format of the template is as follows: "[instruction] reference the patient's injury condition description and grade determination conclusion in known information, combine the "labor capacity identification, worker injury and occupational disease disability grade" determination basis to analyze the problem, and output the final disability grade and answer.

[0048] Promotion principle: For multiple injuries in the same organ or system, or for more than one organ with different parts of the body injured at the same time, the single item disability degree should be identified first. If the disability grades of several items are different, the heavier one should be graded; if two or more grades are the same, at most, one level can be promoted.

[0049] Analysis requirements: 1. For each injury, specify the specific clauses in the corresponding standard; 2. Explain the disability grade and the basis; 3. If there are multiple injuries, please explain whether they need to be processed according to the promotion principle; 4. List the key factors that may affect the rating.

[0050] Additional requirements: If there is ambiguity in the corresponding standard clauses, please explain the possibilities of different interpretations.

[0051] [Known information]{{context}} [Question]{{question}} Please answer: [answer]" Among them, for the content that is easy to appear ambiguous concepts and judgments, the [instruction] tag will be highlighted, the retrieved past conditions and grade conclusions will be automatically added to {{context}}, and the user's input question will be automatically added to {{question}.

[0052] Through the large model after reinforcement learning, the output is more standard and detailed reasoning and judgment process, with complete thinking chain. In order to facilitate the subsequent processing of the output results, by setting system information, the final output of the large model is specified as follows: "

Conclusion: According to the national standard of work injury disability grade identification, x level y item, m level n item... the patient's work injury grade is z level.

[0053] When the data in the unlabeled data set accumulates to a certain amount, check whether the proportion of positive feedback and negative feedback meets about 50%, and whether the total data quantity is greater than 1000. If the above conditions are met, the feedback data is used to build a large model reinforcement learning data set, and the data format must meet the data format requirements in steps one to four. The reinforcement learning labeled data set processed according to the user's positive and negative feedback is trained according to the method of step one to obtain the optimized model. When the model is deployed, a master model is loaded and a model copy is created, and when the user accesses, the request is evenly distributed to the computing power corresponding to the two models. After the model is optimized through the user's positive and negative feedback, the model training is completed, and the following steps are needed to switch the deployed model: ① all user requests are sent to the computing power corresponding to the master model, and the copy model is unloaded; ② Load the optimized model to the computing power device where the previous copy model is located, and take this as the main model of the optimized model; ③ All user requests are issued to the computing power corresponding to the optimized model main model, and whether the service runs stably is monitored in real time; ④ After the optimized model main model runs stably for 1h, the previous main model is unloaded, and the optimized model is loaded to the computing power device where the previous main model is located, and this is taken as the model copy of the optimized model; ⑤ The user request is balanced and issued to the computing power corresponding to the optimized model main model and model copy.

[0054] The above process can be seen in the accompanying Figure 5 The model switching does not interrupt the business of the work injury and disability knowledge base platform, and the smooth upgrade of the work injury and disability identification large model can be completed without the user's awareness.

[0055] Embodiment 2 In an embodiment of the present application, a work injury and disability grade identification system based on large model reinforcement learning is provided, comprising: An acquisition module configured to acquire a user's injury description; A construction module configured to retrieve a plurality of past conditions and grade conclusions with the highest similarity to the injury description from an injury case knowledge base, and combine a prompt prompt word engineering template to construct an instruction prompt word delivered to a large model; A generation module configured to generate a judgment conclusion based on the instruction prompt word through the large model after reinforcement learning; An analysis module configured to analyze the judgment conclusion to obtain a final work injury and disability grade; The large model first performs unsupervised learning through a pre-acquired work injury and disability identification standard, and then performs supervised reinforcement learning using a training data set.

[0056] Embodiment 3 In an embodiment of the present application, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the work injury and disability grade identification method based on large model reinforcement learning.

[0057] Embodiment 4 In an embodiment of the present application, a non-transitory computer readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the work injury and disability grade identification method based on large model reinforcement learning.

[0058] Embodiment 5 An embodiment of the present application provides an electronic device, comprising a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the method for implementing the work injury disability grade identification based on a large model reinforcement learning.

[0059] The present application is described with reference to flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams 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 produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate a device for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus for performing the functions specified in one or more flows and / or blocks.

[0060] These computer program instructions can also be loaded onto a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable data processing device to produce a computer-implemented process, so that the instructions executed by the computer or other programmable data processing device provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 one or more flows and / or blocks

[0061] Although the specific embodiments of the present application are described above with reference to the accompanying drawings, the present application is not limited to the above-described embodiments, and various modifications or changes can be made to the embodiments without departing from the scope of the present application.

Claims

1. A method for identifying the degree of disability based on large model reinforcement learning, characterized in that, The method comprises the following steps: Obtain the injury description of a user; Retrieve a plurality of past cases and grade conclusions with the highest similarity to the injury description from an injury case knowledge base, and combine a prompt prompt word engineering template to construct an instruction prompt word delivered to a large model; Generate a determination conclusion based on the instruction prompt word through the large model after reinforcement learning; Analyze the determination conclusion to obtain the final work-related injury disability grade; The large model is first subjected to unsupervised learning based on the pre-obtained work-related injury disability grade appraisal standard, and then subjected to supervised reinforcement learning using a training data set.

2. The method of claim 1, wherein the method is based on a large model reinforcement learning. The prompt prompt word engineering template comprises an instruction, known information, and a question field; In the instruction, the contents prone to ambiguous concepts and judgments are highlighted, the known information includes the retrieved past cases and grade conclusions, and the question includes the injury description.

3. The method of claim 1, wherein the method comprises: The unsupervised learning based on the pre-obtained work-related injury disability grade appraisal standard is specifically as follows: Based on the work-related injury disability grade appraisal standard, unsupervised data is constructed; The unsupervised data is input into the large model to continue pre-training the network, so that the base large model learns the work-related injury disability grade determination principle, determination clauses, and the determination criteria and explanations in the appendix.

4. The method of claim 3, wherein the method is characterized by, The construction of unsupervised data is specifically as follows: By sorting the special terms in the work-related injury disability grade appraisal standard, a new professional glossary is obtained; The new professional glossary is added to the original segmenter to form a new segmenter, so that the segmentation results can more accurately reflect the actual meaning of the work-related injury scene; The work-related injury disability grade appraisal standard is segmented using the new segmenter to obtain the unsupervised data.

5. The method of claim 1, wherein the method is based on a large model reinforcement learning. The supervised reinforcement learning using the training data set is specifically as follows: Based on the work-related injury appraisal past cases, a training data set is constructed; The training data set is transmitted to the reinforcement learning network to perform iterative optimization of the policy gradient, guiding the inference strategy of the large model after continued pre-training to update in the direction of higher rewards for grade appraisal conclusions.

6. The method of claim 5, wherein the method is characterized by, The construction of the training data set is specifically as follows: Invalid data with empty injury description and disability grade in the work-related injury appraisal past cases is deleted, and the Arabic numerals of the disability grade are all replaced with Chinese numeral descriptions; The content of each case that needs to be trained in the work-related injury appraisal past cases is extracted, mainly including injury description, appraisal basis, nursing level, and disability level, and verification of whether the appraisal basis and the disability level match; Filter the data with a similarity that meets the condition in all extracted cases to reduce the repetition rate of samples and obtain a more balanced distribution of disability levels; Based on the high-quality text data set after cleaning, a prompt word is constructed to form a work-related injury disability appraisal case data set for reinforcement learning.

7. A work injury and disability grade identification system based on large model reinforcement learning, characterized in that, The method comprises the following steps: An acquisition module is configured to obtain the injury description of a user; A construction module is configured to retrieve a plurality of past cases and grade conclusions with the highest similarity to the injury description from an injury case knowledge base, and combine a prompt prompt word engineering template to construct an instruction prompt word delivered to a large model; A generation module is configured to generate a determination conclusion based on the instruction prompt word through the large model after reinforcement learning; The analysis module is configured to analyze the determination conclusion to obtain a final work injury disability grade. The large model is first subjected to unsupervised learning by using the pre-acquired work injury disability grade appraisal standard, and then subjected to supervised reinforcement learning by using a training data set.

8. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the work injury disability grade appraisal method based on large model reinforcement learning according to any one of claims 1-6.

9. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium is configured to store computer instructions, and the computer instructions, when executed by a processor, implement the work injury disability grade appraisal method based on large model reinforcement learning according to any one of claims 1-6.

10. An electronic device, comprising: The computer program, when executed by a processor, implements the work injury disability grade appraisal method based on large model reinforcement learning according to any one of claims 1-6. The computer program, when executed by a processor, implements the work injury disability grade appraisal method based on large model reinforcement learning according to any one of claims 1-6.

Images