79 results about "Perplexity" patented technology

In various examples, systems and methods are disclosed relating to language model jailbreak detection using length-perplexity metrics. A system can identify a prompt for a language model—such as an LLM, VLM, etc.—and generate a perplexity score for the prompt. The system can determine, based at least on the perplexity score and a length of the prompt, that the prompt is indicative of a jailbreak attempt for the large language model. The system can restrict the prompt from input to the large language model—or block an output generated based on the prompt from being shared—responsive to determining that the prompt is indicative of the jailbreak attempt.

The embodiments of the present application relate to the technical field of artificial intelligence. Provided are a prompt anomaly detection method and apparatus, and a device and a storage medium. The method comprises: acquiring a prompt to be subjected to detection and the perplexity of said prompt; if the perplexity exceeds a second threshold value and does not exceed a first threshold value, using an adversarial perturbation mode to process said prompt, in order to acquire a perturbed prompt; on the basis of prediction by a large language model, obtaining response semantics of said prompt and perturbed response semantics corresponding to the perturbed prompt; and if the response semantics and the perturbed response semantics are inconsistent, said prompt being anomalous. Two threshold values are set for perplexity detection, and after perturbation processing is performed on a prompt to be subjected to detection, whether the prompt is anomalous can be determined again by means of the comparison between response semantics and perturbed response semantics, so that the accuracy of prompt anomaly detection can be effectively improved, thereby enhancing the security of large language models.

The invention relates to the technical field of retrieval enhancement generation, and discloses an reflection rollback type retrieval enhancement generation method and device, equipment and a storage medium. The method comprises the steps of generating a rewritten query statement based on a question input by a user and an answer candidate set; executing retrieval based on the rewritten query statement, and adding a retrieval result into an answer candidate set; the state value of the latest retrieval result is calculated through a state evaluation function fusing the confusion degree and the confidence degree; judging whether the state value of the latest retrieval result is lower than the state value of the previous retrieval result of the answer candidate set, and if so, triggering a rollback mechanism; and summarizing each retrieval result in the answer candidate set through a large language model, and outputting a final answer. According to the reflection rollback type retrieval enhancement generation method provided by the invention, a self-reflection and rollback mechanism is introduced, so that the problem of performance reduction caused by low-quality external retrieval information in a large language model is effectively solved.

A system and method for automatically evaluating computer generated content may include: calculating a plurality of metrics for an input text, where the plurality of metrics may include one or more perplexity scores describing a prediction of the input text by a large language model (LLM); determining, based on one or more of the calculated metrics, whether to accept or reject the input text; and performing an exchange of data between remotely connected computer devices based on the determining to accept or reject the text. In some embodiments, calculating of metrics and determining whether to accept or reject the input text may be performed without relying on any information received subsequent to the initial receiving of the input text. Some embodiments may perform automated computerized actions such as, e.g., deploy or discard an update to the LLM based on the determining whether to accept or reject the input text.

The invention discloses an LLMs pre-training dataset optimization method, which comprises the following steps of: selecting internal data of a dataset, and performing hidden Markov model confusion degree calculation on a text segment by segment by adopting a sliding window; and based on a hidden Markov model confusion degree calculation result, obtaining a comprehensive confusion degree score of the whole text through confusion degree weighted average calculation, and using the comprehensive confusion degree score to screen sentences with semantic chaos in the data set. According to the method, a sliding window technology and a weighted average strategy are introduced on the basis of traditional puzzle calculation, so that a comprehensive quality evaluation index can be provided for the whole text, the language model adaptation degree of the text can be evaluated more accurately, and compared with single puzzle calculation, the quality condition of the text can be reflected more comprehensively; and meanwhile, sentences with disordered semantics in the data set can be effectively screened out, a network low-quality corpus text obtained for free is converted into a high-quality and valuable corpus, the LLMs pre-training cost is effectively saved, and the model capability is improved.

The invention relates to the technical field of information, and discloses an instruction data screening method and device, a medium and a product. The method comprises the following steps: determining a first confusion degree and a second confusion degree according to a pre-trained lightweight general language model and a target sample set; the first confusion degree is used for representing the prediction confusion degree under the instruction condition; the second confusion degree is used for representing the prediction confusion degree under the instruction-free condition; determining an information gain score according to the first confusion degree and the second confusion degree; the information gain score is used for quantifying the contribution degree of the instruction content to reduction of the response content generation difficulty; determining a training data set according to the information gain score and the target sample set; the training data set is used for improving the generalization ability and instruction following precision of the large language model in the instruction fine tuning stage. The technical problems of high deployment threshold, high cost, supervision dependence, low training efficiency and poor model generalization ability can be at least solved.

The invention discloses an earth and rockfill dam illness feature mining method and system based on historical text data, and the method comprises the steps: collecting earth and rockfill dam historical illness data, constructing an earth and rockfill dam illness diagnosis text corpus set, carrying out the structured preprocessing of the corpus set, and obtaining an illness feature and danger removal measure text corpus set; generating a structured word sequence subset through a word segmentation tool; processing the structured word sequence subset by adopting an LDA topic model, determining an optimal topic number through a confusion degree curve, and outputting a final topic and a corresponding topic word; constructing a visual network graph based on the subject term co-occurrence frequency; and carrying out centrality analysis on nodes in the visual network diagram, identifying key nodes in the visual network diagram, quantitatively analyzing association rules of the danger characteristics and danger removing measures, and completing feature mining of the earth and rockfill dam danger. The problems that traditional manual diagnosis is high in subjectivity and low in utilization rate of historical engineering data are solved, and intelligent auxiliary decision making of the earth and rockfill dam danger is achieved.

The invention provides an AI text recognition method and device based on ensemble learning and advanced semantic statistical feature analysis, and the method comprises the steps: 1, respectively sending a to-be-recognized text into a Bert detector and a high-order natural language statistical feature detector for recognition, the high-order natural language statistical feature detector comprises a word logarithm probability detector, a word ranking logarithm detector, an Entropy detector and a confusion degree detector; and 2, performing election on detection results output by the Bert detector and the high-order natural language statistical feature detector by using an election module to obtain an AI text recognition result. According to the method, an integrated learning strategy is adopted, and a pre-training language model subjected to fine tuning is combined with high-order natural language statistical characteristics, so that when the model detects a large language model to generate a text, the strong expression ability of the pre-training language model can be fully utilized, and a deep rule of the text can be captured through the high-order statistical characteristics; and the detection accuracy is improved.

According to the method and system for constructing the corpus based on the webpage data set, the large-scale and high-quality corpus can be constructed based on the massive webpage data set through a series of operations of mass filtering, weight removal, hatred speech detection and confusion degree calculation; the steps of quality filtering and data deduplication greatly reduce repeated and low-quality contents in a corpus, and improve the generalization ability of the model and the reliability of output; in addition, high-standard and multi-level quality of the corpus is further ensured through hatred speech detection and puzzle calculation, potential negative effects are reduced, and the model is more accurate and efficient when processing various natural language tasks.

The invention discloses a method and a system for screening high-quality instruction following data aiming at a super-long text. The method comprises the following steps of: calculating a confusion score of the text under a homologous model; splitting the text into a plurality of paragraphs with equal length, and calculating an importance score corresponding to each paragraph; extracting lexical elements in the text paragraphs, constructing an average attention weight by using a multi-head attention mechanism, and calculating an attention weight score; calculating a context-aware attention alignment difficulty score of the text; calculating a final score of the text according to the confusion degree score and the attention alignment difficulty score corresponding to each given text; and sorting the final scores from large to small, and selecting texts corresponding to the scores which are sorted from the top 10% to 20% as screened high-quality instruction following data. The high-quality instruction following data in the data set is obtained through final scoring and is used for large-scale language model training, and the trained large-scale language model obtains excellent performance on tasks in a long text scene.

The invention belongs to the technical field of natural language processing, and particularly relates to a power grid marketing data labeling method, platform and equipment based on a confusion-driven large language model. The method comprises the following steps: S1, collecting related data according to specific requirements in the field of power grid marketing, and preprocessing the data to form a required sample data set; s2, questions related to power grid marketing business are collected, and the questions are answered by the large language model to construct an initial question-answer pair; s3, confusion detection is carried out on the initial question and answer pairs in combination with the sample data set, and the initial question and answer pairs with accuracy risks are screened out; and S4, performing manual examination and correction on the initial question and answer pairs with the accuracy risk by a technician to form annotation data, collecting a large amount of data subjected to manual examination and annotation, and forming a knowledge base for performing re-supervision training on a commercial large language model. According to the method, the problem that the data labeling task load of a worker is too large due to model illusion of a large power grid marketing model is solved.

The embodiment of the invention provides a model parameter compression method and device for a large language model, equipment and a storage medium. The method comprises the steps of obtaining verification data and inputting the verification data into a large language model to obtain an input activation tensor received by each network layer; obtaining the reasoning confusion degree of the large language model for reasoning the verification data, and taking the minimization of the reasoning confusion degree as an optimization target to carry out iterative cutting decision to obtain a cutting decision vector; performing outlier clipping processing on the input activation tensor of the network layer based on the clipping decision vector to obtain a target activation tensor; for each network layer, calculating an importance score of a model parameter based on the target activation tensor, and pruning the initial model parameter tensor in combination with the importance score to obtain an intermediate model parameter tensor; and performing quantization processing on the intermediate model parameter tensor of the network layer to obtain a target large language model. Therefore, the storage and calculation complexity of the large language model can be reduced while the performance of the large language model is maintained.

The invention discloses an optimization method and device for a natural language processing model, and relates to the technical field of information security, and the method comprises the steps: reducing the text confusion degree through employing an iterative beam search mode, and guaranteeing the text quality at the same time, i.e., carrying out the quantitative evaluation and screening of readability and usefulness in a processing process, and improving the processing efficiency. And performing model fine tuning by using the optimized and transformed mixed data to reduce the influence of harmful texts in the third-party data on the model security alignment, thereby improving the security performance and the alignment degree of the large model, and solving the problem that the security alignment of the large language model cannot be effectively coped with in the related technology, and the security alignment efficiency is improved. The technical problem that possible harmful information damage in the third-party data is difficult to fundamentally reduce is solved, and the technical effects that the safety performance and the alignment degree of the large language model are improved, it is ensured that the large language model can give safe and reliable response when facing the data safety problem, and harmful information is prevented from being provided are achieved.

State of art techniques using moderate sized Language Models (LMs) for text classification need fine-tuning or in-context learning. A method and system providing a two-step classification using moderate-sized (#params≤2.7B) causal LM (Gen AI) is disclosed. Firstly, for a text instance to be classified, a set of perplexity and log-likelihood based features are obtained from an LM. Further, a light-weight classifier is trained in the second step to predict the final label. The system enables a new way of exploiting the available labelled instances, in addition to the existing ways like fine-tuning LMs or in-context learning. It neither needs any parameter updates in LMs like fine-tuning nor it is restricted by the number of training examples to be provided in the prompt like in-context learning. The key advantages of the disclosed system are explainability through most suitable key phrases and its applicability in resource poor environment.

The embodiment of the invention provides a model training method and device, electronic equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the steps of obtaining preset question and answer training data, wherein the question and answer training data comprises input question data and a model output result corresponding to the input question data; reasoning trajectory generation is carried out based on each piece of input question data and the corresponding model output result, and a plurality of candidate reasoning trajectories are obtained; performing confusion degree evaluation based on the plurality of candidate reasoning trajectories to obtain a confusion degree score corresponding to each candidate reasoning trajectory; a target reasoning trajectory is determined from the plurality of candidate reasoning trajectories based on the confusion score to guide model training. According to the embodiment of the invention, stable model training can be carried out on a large language model used for business scenes such as intelligent investment advisers or insurance claims and the like in a low-cost and high-efficiency manner, the logic stability of the large language model during business processing is improved, and business services with high professionality and depth are provided.

The invention belongs to the technical field of computer information security, particularly discloses a content security protection method and device based on semantic consistency, and aims to solve the problem that high-level cue word injection attacks are difficult to effectively recognize in the prior art. Carrying out real-time interception on dominant illegal texts through a content filtering module; quantizing generation rationality differences of prompt words between attack and normal language models by using a word vector confusion degree calculation module; evaluating the logic coherence of the sentence structure of the prompt word through a statement semantic consistency judgment module; integrating the multi-dimensional feature data and carrying out risk classification by adopting a machine learning algorithm; and routing the suspected attack request to a value fine tuning model for processing according to a judgment result. According to the technical scheme, high-precision recognition and response to complex cue word injection attacks are achieved, and the content safety protection capacity and compliance guarantee level of a large model in an interaction scene are remarkably improved.

The embodiment of the invention discloses a large model real-time safety protection method and system based on dynamic response. The method comprises the following steps: acquiring text data input to a target large model by a user; determining the confusion degree, gradient and semantic feature vector of the text data under the target large model based on the standby large model; splicing the confusion degree, the gradient and the semantic feature vector into a detection feature vector; based on a multi-classifier integrated detector comprising a deep learning classifier, a rule detector and an abnormal behavior detector, performing adversarial detection on the detection feature vector to obtain a plurality of adversarial detection data; carrying out weighted fusion processing on the plurality of confrontation detection data to obtain a comprehensive confrontation confidence coefficient corresponding to the text data; and starting an intelligent grading protection mechanism for the target large model according to the comprehensive antagonism confidence coefficient. According to the embodiment of the invention, the robustness of the large model in a malicious environment is remarkably enhanced, and the output reliability and safety are guaranteed.

The invention discloses a speculative reasoning method for a multi-modal large language model based on mobile edge collaboration. The speculative reasoning method comprises the following steps: constructing an equipment-side double-clue probe model; task semantic distillation and feature alignment are carried out, joint optimization is carried out on double heads of a double-clue classifier, knowledge of an edge end MLLM is distilled to a double-clue probe model, and an equipment end minimum modal delay decision is executed; asynchronous mode screening and speculative reasoning are realized through a hierarchical gating network; and executing a rollback decision to obtain a corrected final reasoning result. According to the embodiment of the invention, a lightweight dual-clue probe model is constructed at a mobile client, task semantic features are extracted from the deep layer of an edge MLLM, joint estimation of reasoning confidence and modal sufficiency is realized, and a minimum modal delay decision of an equipment end is executed; a hierarchical gating mechanism is adopted at an edge end, a confusion-guided rollback mechanism is introduced, an early reasoning result with insufficient confidence is corrected, and early recognition and efficient reasoning of a minimum full subset in an asynchronous mode are achieved.

The invention discloses a sparse large language model weight refinement method, system and device and a storage medium, which are corresponding schemes, and the scheme realizes training-free and plug-and-play refinement of sparse weight after one-time pruning, significantly reduces the confusion degree, improves the zero sample task performance, and especially has an outstanding effect in a high sparse rate scene. The space is optimized in a unified mode through global soft constraint, heterogeneous Hessian inversion is avoided, pruning prior is effectively utilized in combination with historical momentum, Neumann series extrapolation solution is adopted, iteration time can be remarkably shortened, and expenses of computing resources and memory resources are reduced. The method is oriented to online and operation and maintenance processes of inference services such as text generation, question and answer, classification and retrieval enhanced question and answer and the like, and accesses the current network through hot update or gray release after completing intra-layer weight refinement; the existing word segmentation coding, service gateway and monitoring and rollback mechanisms are reused in the whole process, and it is ensured that the method can be implemented under real service distribution and resource budget.

This invention discloses a flood situation perception method and apparatus based on a weighted LDA algorithm, belonging to the field of flood disaster risk management technology. The method includes: acquiring posts related to flood events through web crawling technology, performing data cleaning and secondary processing to obtain multiple target words and storing them in a dataset; calculating the weight of each target word in the posts using a term frequency-inverse document frequency algorithm; comprehensively determining the number of flood topics using perplexity, consistency index, and Jensen-Shannon divergence; introducing the weights of each target word in the posts into the LDA model, sampling the LDA distribution using a weighted Gibbs sampling algorithm, and estimating the flood post-topic distribution and topic-word distribution to analyze and perceive the flood development trend. This invention can improve the accuracy and interpretability of topic identification, thereby achieving comprehensive perception and accurate understanding of the flood situation.

Embodiments herein describe a sparse-dense-sparse (SDS) process that achieves a better pruning scheme that benefits from pruning-friendliness relative to one-shot pruning schemes. The SDS process performs a first pruning to generate a sparse ML model followed by reconstruction to generate a re-dense ML model, followed by a second pruning to generate another sparse ML model. By pruning a ML model and then re-constructing the ML model, the ML model can be made more pruning-friendly by performing data and / or weight regularization. As a result, performing the second pruning in the SDS process can result in a smoother weight distribution and lower perplexity relative to one-shot pruning.

The application discloses a kind of based on big language model's evolvable software vulnerability detection method, comprising: by regular pattern matching identification Source sentence, based on call graph traversal and data dependence analysis execution function level inter-process slice, build cross-function code context, input the big language model of parameter efficient fine-tuning, output the vulnerability propagation path from Source to Sink;When new vulnerability type needs to be extended, the parameter variation characteristics of old data are extracted by multi-step fine-tuning, and the representative core set is selected by random projection dimension reduction and hybrid distance hierarchical clustering, and the training is played back by mixing new data to alleviate catastrophic forgetting;In the actual use process of tool, the false alarm and the false alarm confirmed by user are collected as feedback signal, the core set is clustered and layered filtered and refined based on perplexity and error prediction analysis, the harmful old knowledge that leads to false alarm and false alarm is removed, and verified new mode is supplemented at the same time, to realize the closed-loop evolution of self-improvement.

The present invention provides a public text analysis method and system based on the LDA model and the LSTM algorithm. The method includes: obtaining the target text data of a scientific and technological project; extracting the feature words in the target text data and constructing a feature word vector matrix based on the feature words; constructing a perplexity curve based on the feature word vector matrix and constructing a topic model based on the perplexity curve; determining the topic attention degree of the target topic based on the topic model, and determining the attention degree trend feature based on the topic attention degree, where the attention degree trend feature is used to indicate the change trend of the target topic in the time dimension, the space dimension, and the topic dimension; analyzing the target text data by using a pre-established sentiment analysis model to obtain a sentiment analysis result; and generating a text analysis result of the scientific and technological project based on the target topic, the attention degree trend feature, and the sentiment analysis result. This application realizes the intelligent analysis of public texts and improves the efficiency and accuracy of text interpretation.

A system for influencing outputs of a large language model includes at least one memory, at least one processor, a perplexity optimizer, and a corpus handler. The processor executes instructions stored in the memory to operate the optimizer and handler. The perplexity optimizer generates multiple candidate supporting texts based on a target concept. It computes a perplexity metric for each candidate within a context derived from a local corpus, using token likelihoods from a reference language model, and selects a supporting text based on the metric. The corpus handler identifies online editable corpora based on their likelihood of inclusion in language model training data. It then collects contextual information from these corpora to create the local corpus and inserts the selected supporting text into at least one of the identified online corpora.

The application provides a large language model individualization alignment method and device based on difficulty perception learning, and belongs to the technical field of model training. The method comprises the following steps: obtaining multiple groups of training samples; each group of training samples comprises a user question, a user portrait, and a target reply corresponding to the user question under the user portrait; calculating the perplexity of the target reply in each group of training samples; dividing the multiple groups of training samples into high-perplexity training samples and low-perplexity training samples according to a preset perplexity threshold and the perplexity of the target reply in each group of training samples; performing first training on a preset large language model based on the low-perplexity training samples and a likelihood loss function, performing second training on the preset large language model after the first training based on the high-perplexity training samples and reinforcement learning, and obtaining a target large language model, so that the target large language model completes individualization alignment. The application can improve the individualization alignment effect of the model.

The present disclosure provides a text data processing method, device and equipment and storage medium. The present disclosure obtains a plurality of perplexity values corresponding to a plurality of sample texts respectively by performing inference on each of the plurality of sample texts through a plurality of pre-trained models. Further, a confidence interval is determined according to the plurality of perplexity values corresponding to the plurality of sample texts respectively, so as to determine an abnormal sample in the plurality of sample texts according to the confidence interval. After clustering processing of the plurality of sample texts is performed to obtain a plurality of clustering clusters, a clustering cluster containing an abnormal sample in the plurality of clustering clusters is removed, that is, all sample texts contained in the clustering cluster are removed. These sample texts are usually low-quality or irrelevant text data. The remaining sample texts constitute a high-quality data set, thereby improving the accuracy of subsequent processing or calculation processes that rely on the data set.

The invention discloses a pancreatic cancer prediction method and system based on local and global perplexity weighted pruning, and relates to the technical field of artificial intelligence natural language processing, and the method comprises the steps: obtaining electronic medical record text data of a to-be-diagnosed patient, and carrying out the data screening processing; according to the method, a structured cue word sequence is constructed, a large language model is guided to generate an initial reasoning path, a step which plays a key role in context coherence can be automatically identified by calculating a local confusion degree variable quantity, and evaluation of the contribution degree of the reasoning step to a final diagnosis conclusion is realized by calculating a global confusion degree variable quantity, so that the accuracy of reasoning is improved. By weighting local and global dual confusion degree variable quantities, a comprehensive importance scoring function is constructed, comprehensive analysis of reasoning steps is realized, redundant steps can be automatically eliminated by setting an adaptive pruning threshold value, an optimized thinking chain with clear logic is finally formed, and a pancreatic cancer risk prediction result and a key reasoning path are output based on the thinking chain. And the diagnosis process is completely clear.

This invention presents an AI-generated text detection method based on graph structure features, belonging to the fields of artificial intelligence and natural language processing. The method includes: dataset construction, entity relation extraction and graph structure construction, graph structure feature extraction, graph structure feature model training, and text detection. It further incorporates traditional text feature extraction and model training, adaptively fusing the traditional text feature model and the graph feature model based on confidence-weighted entropy, and then performing text detection based on the fused model. This invention is the first to perform AI text detection from the perspective of graph structure features, breaking through the limitations of existing research that focuses on surface features such as vocabulary, syntax, and perplexity. The fusion strategy dynamically adjusts the fusion weights by quantifying the uncertainty of model predictions, maintaining a high level of performance on both original data and adversarial examples, achieving a balance between detection accuracy and adversarial robustness. It can be widely applied to the detection of AI-generated content such as news content and academic papers.

The invention discloses an RAG optimization method fusing bidirectional confusion degree blocking and adaptive selection, and relates to the technical field of retrieval enhancement generation. According to the segmentation method based on the bidirectional confusion degree, semantic boundaries in long documents are recognized in a self-adaptive mode by analyzing predictability of forward and backward marks, and compared with a traditional one-way PPL method, the method has the advantages that the coherence of language chunks is improved, and retrieval correlation is enhanced; on the basis of a hierarchical index structure, a self-adaptive selection method based on similarity gradient information is provided, and the optimal candidate number to be retrieved in each stage is determined in a self-adaptive mode by analyzing gradient changes in a similarity score sequence.

The present invention relates to the field of privacy security and aims to provide a method and system for assessing the privacy leakage risk of pre-trained language models. The method comprises: adding forged data to a pre-trained dataset; inputting the pre-trained dataset into an initialized neural network model, and calculating the loss according to a set pre-training task and loss function; continuously updating the model's parameters during the training process to increase its privacy leakage risk; inputting a fine-tuning dataset into the pre-trained neural network model to fine-tune the model's feature extraction capabilities; inputting privacy prefix content into the model, and outputting text information as a prediction result; calculating, counting, and sorting the perplexity of the output information, and assessing the risk of privacy data leakage by comparing the proportion of generated privacy information. The present invention can effectively improve the accuracy of assessing the risk of privacy data leakage, expose the privacy data leakage risk of pre-trained language models, and provide ideas for the subsequent development of related defense methods.