Nonconvex optimisation plays a crucial role in science and industry.However,existing methods often encounter local optima or provide inferior solutions when solving nonconvex optimisation problems,lacking robustness i...Nonconvex optimisation plays a crucial role in science and industry.However,existing methods often encounter local optima or provide inferior solutions when solving nonconvex optimisation problems,lacking robustness in noise scenarios.To address these limitations,we aim to develop a robust,efficient and globally convergent solver for nonconvex optimisation.This is achieved by combining the efficient local exploitation ability of a parameter-learnt neural dynamics(PLND)model with the global search capability of the coevolutionary mechanism.We combine their characteristics to design a coevolutionary neural dynamics with learnable parameters(CNDLP)model.The gradient information is used to find the optimal solution more effectively,and neural dynamics models have robustness,which ensures that the influence of noise can be effectively suppressed in the calculation process.Theoretical analyses show the global convergence and robustness of the designed CNDLP model.Numerical experiments on 9 benchmark functions and a practical engineering design example are conducted with five existing meta-heuristic algorithms.Benchmarks cover diverse problems,from classical landscapes like benchmark Shubert to high-dimensional cases such as 30-dimensional Rosenbrock.Results confirm CNDLP's excellent performance in both solution quality and convergence speed under noise.展开更多
This study investigates the application of Learnable Memory Vision Transformers(LMViT)for detecting metal surface flaws,comparing their performance with traditional CNNs,specifically ResNet18 and ResNet50,as well as o...This study investigates the application of Learnable Memory Vision Transformers(LMViT)for detecting metal surface flaws,comparing their performance with traditional CNNs,specifically ResNet18 and ResNet50,as well as other transformer-based models including Token to Token ViT,ViT withoutmemory,and Parallel ViT.Leveraging awidely-used steel surface defect dataset,the research applies data augmentation and t-distributed stochastic neighbor embedding(t-SNE)to enhance feature extraction and understanding.These techniques mitigated overfitting,stabilized training,and improved generalization capabilities.The LMViT model achieved a test accuracy of 97.22%,significantly outperforming ResNet18(88.89%)and ResNet50(88.90%),aswell as the Token to TokenViT(88.46%),ViT without memory(87.18),and Parallel ViT(91.03%).Furthermore,LMViT exhibited superior training and validation performance,attaining a validation accuracy of 98.2%compared to 91.0%for ResNet 18,96.0%for ResNet50,and 89.12%,87.51%,and 91.21%for Token to Token ViT,ViT without memory,and Parallel ViT,respectively.The findings highlight the LMViT’s ability to capture long-range dependencies in images,an areawhere CNNs struggle due to their reliance on local receptive fields and hierarchical feature extraction.The additional transformer-based models also demonstrate improved performance in capturing complex features over CNNs,with LMViT excelling particularly at detecting subtle and complex defects,which is critical for maintaining product quality and operational efficiency in industrial applications.For instance,the LMViT model successfully identified fine scratches and minor surface irregularities that CNNs often misclassify.This study not only demonstrates LMViT’s potential for real-world defect detection but also underscores the promise of other transformer-based architectures like Token to Token ViT,ViT without memory,and Parallel ViT in industrial scenarios where complex spatial relationships are key.Future research may focus on enhancing LMViT’s computational efficiency for deployment in real-time quality control systems.展开更多
Artificial intelligence(AI)models have significantly impacted various areas of the atmospheric sciences,reshaping our approach to climate-related challenges.Amid this AI-driven transformation,the foundational role of ...Artificial intelligence(AI)models have significantly impacted various areas of the atmospheric sciences,reshaping our approach to climate-related challenges.Amid this AI-driven transformation,the foundational role of physics in climate science has occasionally been overlooked.Our perspective suggests that the future of climate modeling involves a synergistic partnership between AI and physics,rather than an“either/or”scenario.Scrutinizing controversies around current physical inconsistencies in large AI models,we stress the critical need for detailed dynamic diagnostics and physical constraints.Furthermore,we provide illustrative examples to guide future assessments and constraints for AI models.Regarding AI integration with numerical models,we argue that offline AI parameterization schemes may fall short of achieving global optimality,emphasizing the importance of constructing online schemes.Additionally,we highlight the significance of fostering a community culture and propose the OCR(Open,Comparable,Reproducible)principles.Through a better community culture and a deep integration of physics and AI,we contend that developing a learnable climate model,balancing AI and physics,is an achievable goal.展开更多
In recent years,various adversarial defense methods have been proposed to improve the robustness of deep neural networks.Adversarial training is one of the most potent methods to defend against adversarial attacks.How...In recent years,various adversarial defense methods have been proposed to improve the robustness of deep neural networks.Adversarial training is one of the most potent methods to defend against adversarial attacks.However,the difference in the feature space between natural and adversarial examples hinders the accuracy and robustness of the model in adversarial training.This paper proposes a learnable distribution adversarial training method,aiming to construct the same distribution for training data utilizing the Gaussian mixture model.The distribution centroid is built to classify samples and constrain the distribution of the sample features.The natural and adversarial examples are pushed to the same distribution centroid to improve the accuracy and robustness of the model.The proposed method generates adversarial examples to close the distribution gap between the natural and adversarial examples through an attack algorithm explicitly designed for adversarial training.This algorithm gradually increases the accuracy and robustness of the model by scaling perturbation.Finally,the proposed method outputs the predicted labels and the distance between the sample and the distribution centroid.The distribution characteristics of the samples can be utilized to detect adversarial cases that can potentially evade the model defense.The effectiveness of the proposed method is demonstrated through comprehensive experiments.展开更多
Adversarial distillation(AD)has emerged as a potential solution to tackle the challenging optimization problem of loss with hard labels in adversarial training.However,fixed sample-agnostic and student-egocentric atta...Adversarial distillation(AD)has emerged as a potential solution to tackle the challenging optimization problem of loss with hard labels in adversarial training.However,fixed sample-agnostic and student-egocentric attack strategies are unsuitable for distillation.Additionally,the reliability of guidance from static teachers diminishes as target models become more robust.This paper proposes an AD method called Learnable Distillation Attack Strategies and Evolvable Teachers Adversarial Distillation(LDAS&ET-AD).Firstly,a learnable distillation attack strategies generating mechanism is developed to automatically generate sample-dependent attack strategies tailored for distillation.A strategy model is introduced to produce attack strategies that enable adversarial examples(AEs)to be created in areas where the target model significantly diverges from the teachers by competing with the target model in minimizing or maximizing the AD loss.Secondly,a teacher evolution strategy is introduced to enhance the reliability and effectiveness of knowledge in improving the generalization performance of the target model.By calculating the experimentally updated target model’s validation performance on both clean samples and AEs,the impact of distillation from each training sample and AE on the target model’s generalization and robustness abilities is assessed to serve as feedback to fine-tune standard and robust teachers accordingly.Experiments evaluate the performance of LDAS&ET-AD against different adversarial attacks on the CIFAR-10 and CIFAR-100 datasets.The experimental results demonstrate that the proposed method achieves a robust precision of 45.39%and 42.63%against AutoAttack(AA)on the CIFAR-10 dataset for ResNet-18 and MobileNet-V2,respectively,marking an improvement of 2.31%and 3.49%over the baseline method.In comparison to state-of-the-art adversarial defense techniques,our method surpasses Introspective Adversarial Distillation,the top-performing method in terms of robustness under AA attack for the CIFAR-10 dataset,with enhancements of 1.40%and 1.43%for ResNet-18 and MobileNet-V2,respectively.These findings demonstrate the effectiveness of our proposed method in enhancing the robustness of deep learning networks(DNNs)against prevalent adversarial attacks when compared to other competing methods.In conclusion,LDAS&ET-AD provides reliable and informative soft labels to one of the most promising defense methods,AT,alleviating the limitations of untrusted teachers and unsuitable AEs in existing AD techniques.We hope this paper promotes the development of DNNs in real-world trust-sensitive fields and helps ensure a more secure and dependable future for artificial intelligence systems.展开更多
Deep learning-based methods have become alternatives to traditional numerical weather prediction systems,offering faster computation and the ability to utilize large historical datasets.However,the application of deep...Deep learning-based methods have become alternatives to traditional numerical weather prediction systems,offering faster computation and the ability to utilize large historical datasets.However,the application of deep learning to medium-range regional weather forecasting with limited data remains a significant challenge.In this work,three key solutions are proposed:(1)motivated by the need to improve model performance in data-scarce regional forecasting scenarios,the authors innovatively apply semantic segmentation models,to better capture spatiotemporal features and improve prediction accuracy;(2)recognizing the challenge of overfitting and the inability of traditional noise-based data augmentation methods to effectively enhance model robustness,a novel learnable Gaussian noise mechanism is introduced that allows the model to adaptively optimize perturbations for different locations,ensuring more effective learning;and(3)to address the issue of error accumulation in autoregressive prediction,as well as the challenge of learning difficulty and the lack of intermediate data utilization in one-shot prediction,the authors propose a cascade prediction approach that effectively resolves these problems while significantly improving model forecasting performance.The method achieves a competitive result in The East China Regional AI Medium Range Weather Forecasting Competition.Ablation experiments further validate the effectiveness of each component,highlighting their contributions to enhancing prediction performance.展开更多
The complexity of the battlefield environment,including its high dynamics,along with the high-dimensional spaces of state and decision-making,has brought severe challenges to unmanned combat aerial vehicles(UCAVs)in t...The complexity of the battlefield environment,including its high dynamics,along with the high-dimensional spaces of state and decision-making,has brought severe challenges to unmanned combat aerial vehicles(UCAVs)in the cooperative autonomous air combat decision-making.This paper focuses on the many-to-many air combat maneuvering decision(MMACMD)in an environment with extremely limited communication.An asynchronous hierarchical deep reinforcement learning method with learnable reward shaping(AHDRL_LRS)is proposed.First,by introducing an asynchronous hierarchical reinforcement learning framework,the large-scale MMACMD is decomposed into smaller-scale subtasks to reduce the dimensions of the decision spaces.Second,to achieve the coordinated global task allocation in the environment with extremely limited communication,the learnable reward with embedded target intention(LRETI)is proposed.Through the LRETI,the target selecting intentions generated by the high-level policy are implicitly represented as learnable parameters in the situation reward function,which is used to train the low-level flight maneuver policy.Third,to dynamically characterize the topological correlations of each unit in the UCAV swarm and enhance the transferability and scalability of the decision-making model,the flexible target intention network(FTIN)structure based on the multi-head self-attention(MHSA)model is designed for the representation of the high-level policy,which can accept input features with variable-length sequences.Moreover,a graph learning-based critic network is adopted in the low-level policy model to address the dynamic credit assignment.Finally,by comparing with the baseline methods under scenarios with various initialization from 6-vs-6 to 20-to-20 scales,the effectiveness and superiority of the proposed AHDRL_LRS are validated through the results of the simulation experiment.展开更多
Objective To develop QingNangTCM,a specialized large language model(LLM)tailored for expert-level traditional Chinese medicine(TCM)question-answering and clinical reasoning,addressing the scarcity of domain-specific c...Objective To develop QingNangTCM,a specialized large language model(LLM)tailored for expert-level traditional Chinese medicine(TCM)question-answering and clinical reasoning,addressing the scarcity of domain-specific corpora and specialized alignment.Methods We constructed QnTCM_Dataset,a corpus of 100000 entries,by integrating data from ShenNong_TCM_Dataset and SymMap v2.0,and synthesizing additional samples via retrieval-augmented generation(RAG)and persona-driven generation.The dataset comprehensively covers diagnostic inquiries,prescriptions,and herbal knowledge.Utilizing P-Tuning v2,we fine-tuned the GLM-4-9B-Chat backbone to develop QingNangTCM.A multidimensional evaluation framework,assessing accuracy,coverage,consistency,safety,professionalism,and fluency,was established using metrics such as bilingual evaluation understudy(BLEU),recall-oriented understudy for gisting evaluation(ROUGE),metric for evaluation of translation with explicit ordering(METEOR),and LLM-as-a-Judge with expert review.Qualitative analysis was conducted across four simulated clinical scenarios:symptom analysis,disease treatment,herb inquiry,and failure cases.Baseline models included GLM-4-9BChat,DeepSeek-V2,HuatuoGPT-II(7B),and GLM-4-9B-Chat(freeze-tuning).Results QingNangTCM achieved the highest scores in BLEU-1/2/3/4(0.425/0.298/0.137/0.064),ROUGE-1/2(0.368/0.157),and METEOR(0.218),demonstrating a balanced and superior normalized performance profile of 0.900 across the dimensions of accuracy,coverage,and consistency.Although its ROUGE-L score(0.299)was lower than that of HuatuoGPT-II(7B)(0.351),it significantly outperformed domain-specific models in expert-validated win rates for professionalism(86%)and safety(73%).Qualitative analysis confirmed that the model strictly adheres to the“symptom-syndrome-pathogenesis-treatment”reasoning chain,though occasional misclassifications and hallucinations persisted when dealing with rare medicinal materials and uncommon syndromes.展开更多
基金supported by the Leading Talent of the Qinghai Province Kunlun Talents Programme・High-Level Innovative and Entrepreneurial Talents(QHKLYC-GDCXCY-2024-359).
文摘Nonconvex optimisation plays a crucial role in science and industry.However,existing methods often encounter local optima or provide inferior solutions when solving nonconvex optimisation problems,lacking robustness in noise scenarios.To address these limitations,we aim to develop a robust,efficient and globally convergent solver for nonconvex optimisation.This is achieved by combining the efficient local exploitation ability of a parameter-learnt neural dynamics(PLND)model with the global search capability of the coevolutionary mechanism.We combine their characteristics to design a coevolutionary neural dynamics with learnable parameters(CNDLP)model.The gradient information is used to find the optimal solution more effectively,and neural dynamics models have robustness,which ensures that the influence of noise can be effectively suppressed in the calculation process.Theoretical analyses show the global convergence and robustness of the designed CNDLP model.Numerical experiments on 9 benchmark functions and a practical engineering design example are conducted with five existing meta-heuristic algorithms.Benchmarks cover diverse problems,from classical landscapes like benchmark Shubert to high-dimensional cases such as 30-dimensional Rosenbrock.Results confirm CNDLP's excellent performance in both solution quality and convergence speed under noise.
基金funded by Woosong University Academic Research 2024.
文摘This study investigates the application of Learnable Memory Vision Transformers(LMViT)for detecting metal surface flaws,comparing their performance with traditional CNNs,specifically ResNet18 and ResNet50,as well as other transformer-based models including Token to Token ViT,ViT withoutmemory,and Parallel ViT.Leveraging awidely-used steel surface defect dataset,the research applies data augmentation and t-distributed stochastic neighbor embedding(t-SNE)to enhance feature extraction and understanding.These techniques mitigated overfitting,stabilized training,and improved generalization capabilities.The LMViT model achieved a test accuracy of 97.22%,significantly outperforming ResNet18(88.89%)and ResNet50(88.90%),aswell as the Token to TokenViT(88.46%),ViT without memory(87.18),and Parallel ViT(91.03%).Furthermore,LMViT exhibited superior training and validation performance,attaining a validation accuracy of 98.2%compared to 91.0%for ResNet 18,96.0%for ResNet50,and 89.12%,87.51%,and 91.21%for Token to Token ViT,ViT without memory,and Parallel ViT,respectively.The findings highlight the LMViT’s ability to capture long-range dependencies in images,an areawhere CNNs struggle due to their reliance on local receptive fields and hierarchical feature extraction.The additional transformer-based models also demonstrate improved performance in capturing complex features over CNNs,with LMViT excelling particularly at detecting subtle and complex defects,which is critical for maintaining product quality and operational efficiency in industrial applications.For instance,the LMViT model successfully identified fine scratches and minor surface irregularities that CNNs often misclassify.This study not only demonstrates LMViT’s potential for real-world defect detection but also underscores the promise of other transformer-based architectures like Token to Token ViT,ViT without memory,and Parallel ViT in industrial scenarios where complex spatial relationships are key.Future research may focus on enhancing LMViT’s computational efficiency for deployment in real-time quality control systems.
基金supported by the National Natural Science Foundation of China(Grant Nos.42141019 and 42261144687)and STEP(Grant No.2019QZKK0102)supported by the Korea Environmental Industry&Technology Institute(KEITI)through the“Project for developing an observation-based GHG emissions geospatial information map”,funded by the Korea Ministry of Environment(MOE)(Grant No.RS-2023-00232066).
文摘Artificial intelligence(AI)models have significantly impacted various areas of the atmospheric sciences,reshaping our approach to climate-related challenges.Amid this AI-driven transformation,the foundational role of physics in climate science has occasionally been overlooked.Our perspective suggests that the future of climate modeling involves a synergistic partnership between AI and physics,rather than an“either/or”scenario.Scrutinizing controversies around current physical inconsistencies in large AI models,we stress the critical need for detailed dynamic diagnostics and physical constraints.Furthermore,we provide illustrative examples to guide future assessments and constraints for AI models.Regarding AI integration with numerical models,we argue that offline AI parameterization schemes may fall short of achieving global optimality,emphasizing the importance of constructing online schemes.Additionally,we highlight the significance of fostering a community culture and propose the OCR(Open,Comparable,Reproducible)principles.Through a better community culture and a deep integration of physics and AI,we contend that developing a learnable climate model,balancing AI and physics,is an achievable goal.
基金supported by the National Natural Science Foundation of China(No.U21B2003,62072250,62072250,62172435,U1804263,U20B2065,61872203,71802110,61802212)the National Key R&D Program of China(No.2021QY0700)+4 种基金the Key Laboratory of Intelligent Support Technology for Complex Environments(Nanjing University of Information Science and Technology),Ministry of Education,and the Natural Science Foundation of Jiangsu Province(No.BK20200750)Open Foundation of Henan Key Laboratory of Cyberspace Situation Awareness(No.HNTS2022002)Post Graduate Research&Practice Innvoation Program of Jiangsu Province(No.KYCX200974)Open Project Fund of Shandong Provincial Key Laboratory of Computer Network(No.SDKLCN-2022-05)the Priority Academic Program Development of Jiangsu Higher Education Institutions(PAPD)Fund and Graduate Student Scientific Research Innovation Projects of Jiangsu Province(No.KYCX231359).
文摘In recent years,various adversarial defense methods have been proposed to improve the robustness of deep neural networks.Adversarial training is one of the most potent methods to defend against adversarial attacks.However,the difference in the feature space between natural and adversarial examples hinders the accuracy and robustness of the model in adversarial training.This paper proposes a learnable distribution adversarial training method,aiming to construct the same distribution for training data utilizing the Gaussian mixture model.The distribution centroid is built to classify samples and constrain the distribution of the sample features.The natural and adversarial examples are pushed to the same distribution centroid to improve the accuracy and robustness of the model.The proposed method generates adversarial examples to close the distribution gap between the natural and adversarial examples through an attack algorithm explicitly designed for adversarial training.This algorithm gradually increases the accuracy and robustness of the model by scaling perturbation.Finally,the proposed method outputs the predicted labels and the distance between the sample and the distribution centroid.The distribution characteristics of the samples can be utilized to detect adversarial cases that can potentially evade the model defense.The effectiveness of the proposed method is demonstrated through comprehensive experiments.
基金the National Key Research and Development Program of China(2021YFB1006200)Major Science and Technology Project of Henan Province in China(221100211200).Grant was received by S.Li.
文摘Adversarial distillation(AD)has emerged as a potential solution to tackle the challenging optimization problem of loss with hard labels in adversarial training.However,fixed sample-agnostic and student-egocentric attack strategies are unsuitable for distillation.Additionally,the reliability of guidance from static teachers diminishes as target models become more robust.This paper proposes an AD method called Learnable Distillation Attack Strategies and Evolvable Teachers Adversarial Distillation(LDAS&ET-AD).Firstly,a learnable distillation attack strategies generating mechanism is developed to automatically generate sample-dependent attack strategies tailored for distillation.A strategy model is introduced to produce attack strategies that enable adversarial examples(AEs)to be created in areas where the target model significantly diverges from the teachers by competing with the target model in minimizing or maximizing the AD loss.Secondly,a teacher evolution strategy is introduced to enhance the reliability and effectiveness of knowledge in improving the generalization performance of the target model.By calculating the experimentally updated target model’s validation performance on both clean samples and AEs,the impact of distillation from each training sample and AE on the target model’s generalization and robustness abilities is assessed to serve as feedback to fine-tune standard and robust teachers accordingly.Experiments evaluate the performance of LDAS&ET-AD against different adversarial attacks on the CIFAR-10 and CIFAR-100 datasets.The experimental results demonstrate that the proposed method achieves a robust precision of 45.39%and 42.63%against AutoAttack(AA)on the CIFAR-10 dataset for ResNet-18 and MobileNet-V2,respectively,marking an improvement of 2.31%and 3.49%over the baseline method.In comparison to state-of-the-art adversarial defense techniques,our method surpasses Introspective Adversarial Distillation,the top-performing method in terms of robustness under AA attack for the CIFAR-10 dataset,with enhancements of 1.40%and 1.43%for ResNet-18 and MobileNet-V2,respectively.These findings demonstrate the effectiveness of our proposed method in enhancing the robustness of deep learning networks(DNNs)against prevalent adversarial attacks when compared to other competing methods.In conclusion,LDAS&ET-AD provides reliable and informative soft labels to one of the most promising defense methods,AT,alleviating the limitations of untrusted teachers and unsuitable AEs in existing AD techniques.We hope this paper promotes the development of DNNs in real-world trust-sensitive fields and helps ensure a more secure and dependable future for artificial intelligence systems.
基金supported by the National Natural Science Foundation of China[grant number 62376217]the Young Elite Scientists Sponsorship Program by CAST[grant number 2023QNRC001]the Joint Research Project for Meteorological Capacity Improvement[grant number 24NLTSZ003]。
文摘Deep learning-based methods have become alternatives to traditional numerical weather prediction systems,offering faster computation and the ability to utilize large historical datasets.However,the application of deep learning to medium-range regional weather forecasting with limited data remains a significant challenge.In this work,three key solutions are proposed:(1)motivated by the need to improve model performance in data-scarce regional forecasting scenarios,the authors innovatively apply semantic segmentation models,to better capture spatiotemporal features and improve prediction accuracy;(2)recognizing the challenge of overfitting and the inability of traditional noise-based data augmentation methods to effectively enhance model robustness,a novel learnable Gaussian noise mechanism is introduced that allows the model to adaptively optimize perturbations for different locations,ensuring more effective learning;and(3)to address the issue of error accumulation in autoregressive prediction,as well as the challenge of learning difficulty and the lack of intermediate data utilization in one-shot prediction,the authors propose a cascade prediction approach that effectively resolves these problems while significantly improving model forecasting performance.The method achieves a competitive result in The East China Regional AI Medium Range Weather Forecasting Competition.Ablation experiments further validate the effectiveness of each component,highlighting their contributions to enhancing prediction performance.
基金supported by the National Science Fund for Distinguished Young Scholars(Grant No.62425304)the Basic Science Center Programs of NSFC(Grant No.62088101)+1 种基金the Shanghai Municipal Science and Technology Major Project(Grant No.2021SHZDZX0100)the Shanghai Municipal of Science and Technology Project(Grant No.19511132101)。
文摘The complexity of the battlefield environment,including its high dynamics,along with the high-dimensional spaces of state and decision-making,has brought severe challenges to unmanned combat aerial vehicles(UCAVs)in the cooperative autonomous air combat decision-making.This paper focuses on the many-to-many air combat maneuvering decision(MMACMD)in an environment with extremely limited communication.An asynchronous hierarchical deep reinforcement learning method with learnable reward shaping(AHDRL_LRS)is proposed.First,by introducing an asynchronous hierarchical reinforcement learning framework,the large-scale MMACMD is decomposed into smaller-scale subtasks to reduce the dimensions of the decision spaces.Second,to achieve the coordinated global task allocation in the environment with extremely limited communication,the learnable reward with embedded target intention(LRETI)is proposed.Through the LRETI,the target selecting intentions generated by the high-level policy are implicitly represented as learnable parameters in the situation reward function,which is used to train the low-level flight maneuver policy.Third,to dynamically characterize the topological correlations of each unit in the UCAV swarm and enhance the transferability and scalability of the decision-making model,the flexible target intention network(FTIN)structure based on the multi-head self-attention(MHSA)model is designed for the representation of the high-level policy,which can accept input features with variable-length sequences.Moreover,a graph learning-based critic network is adopted in the low-level policy model to address the dynamic credit assignment.Finally,by comparing with the baseline methods under scenarios with various initialization from 6-vs-6 to 20-to-20 scales,the effectiveness and superiority of the proposed AHDRL_LRS are validated through the results of the simulation experiment.
基金Hebei Province Higher Education Scientific Research Project(QN2025367)Zhangjiakou City 2022 Municipal Science and Technology Plan Self-raised Fund Project(221105D)Hebei Province Education Science“14th Five-Year Plan”Project(2404224).
文摘Objective To develop QingNangTCM,a specialized large language model(LLM)tailored for expert-level traditional Chinese medicine(TCM)question-answering and clinical reasoning,addressing the scarcity of domain-specific corpora and specialized alignment.Methods We constructed QnTCM_Dataset,a corpus of 100000 entries,by integrating data from ShenNong_TCM_Dataset and SymMap v2.0,and synthesizing additional samples via retrieval-augmented generation(RAG)and persona-driven generation.The dataset comprehensively covers diagnostic inquiries,prescriptions,and herbal knowledge.Utilizing P-Tuning v2,we fine-tuned the GLM-4-9B-Chat backbone to develop QingNangTCM.A multidimensional evaluation framework,assessing accuracy,coverage,consistency,safety,professionalism,and fluency,was established using metrics such as bilingual evaluation understudy(BLEU),recall-oriented understudy for gisting evaluation(ROUGE),metric for evaluation of translation with explicit ordering(METEOR),and LLM-as-a-Judge with expert review.Qualitative analysis was conducted across four simulated clinical scenarios:symptom analysis,disease treatment,herb inquiry,and failure cases.Baseline models included GLM-4-9BChat,DeepSeek-V2,HuatuoGPT-II(7B),and GLM-4-9B-Chat(freeze-tuning).Results QingNangTCM achieved the highest scores in BLEU-1/2/3/4(0.425/0.298/0.137/0.064),ROUGE-1/2(0.368/0.157),and METEOR(0.218),demonstrating a balanced and superior normalized performance profile of 0.900 across the dimensions of accuracy,coverage,and consistency.Although its ROUGE-L score(0.299)was lower than that of HuatuoGPT-II(7B)(0.351),it significantly outperformed domain-specific models in expert-validated win rates for professionalism(86%)and safety(73%).Qualitative analysis confirmed that the model strictly adheres to the“symptom-syndrome-pathogenesis-treatment”reasoning chain,though occasional misclassifications and hallucinations persisted when dealing with rare medicinal materials and uncommon syndromes.
