Large model-based sample detection method, training method and related device

Abstract

The invention provides a large model-based sample detection method, a training method and a related device, and relates to the technical field of computers, in particular to the technical fields of deep learning, large models, training sample processing and the like. According to the specific implementation scheme, hidden state features of a target sample are extracted based on a large model; processing the hidden state features based on a plurality of probes to obtain evaluation indexes output by each probe for the target sample; wherein the plurality of probes output evaluation indexes from different dimensions; and performing consistency analysis on the plurality of obtained evaluation indexes to obtain a noise detection result of the target sample.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to the fields of deep learning, large models, and training sample processing. Background Technology

[0002] Training samples are the source of knowledge for a model. The model builds its own knowledge system by learning features from the training samples. The quality of the training samples directly affects the model's performance. If the training samples are rich and representative enough, the model can learn a wider range of features and patterns, thus enabling it to make more accurate predictions when faced with new data.

[0003] However, due to the diversity of sample sources and issues with the labeling of training samples, noise may appear in the training samples. This noise may affect the model's ability to learn effectively from the training samples. Summary of the Invention

[0004] This disclosure provides a sample detection method, training method, and related apparatus based on a large model.

[0005] According to one aspect of this disclosure, a sample detection method based on a large model is provided, comprising: Extracting latent state features of target samples based on a large model; The latent state features are processed by multiple probes respectively, and an evaluation index is obtained for each probe for the target sample; wherein, the multiple probes output the evaluation index from different dimensions. Consistency analysis was performed on the obtained evaluation indicators to obtain the noise detection results of the target sample.

[0006] According to another aspect of this disclosure, a model training method for sample detection is provided, comprising: Extracting latent feature representations from training samples based on large models; The probe is trained based on the latent feature representation; the probe is used to detect noise in the target sample.

[0007] According to another aspect of this disclosure, a sample detection device based on a large model is provided, comprising: The first extraction module is used to extract the latent state features of the target sample based on the large model; The processing module is used to process the latent state features based on multiple probes respectively, and obtain the evaluation index output by each probe for the target sample; wherein, the multiple probes output the evaluation index from different dimensions. The analysis module is used to perform consistency analysis on the obtained multiple evaluation indicators to obtain the noise detection results of the target sample.

[0008] According to another aspect of this disclosure, a model training apparatus for sample detection is provided, comprising: The second extraction module is used to extract the latent feature representations of training samples based on the large model; A training module is used to train a probe based on the latent feature representation; the probe is used to detect noise on the target sample.

[0009] According to another aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and The memory is communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described in the present disclosure.

[0010] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform any of the methods according to embodiments of this disclosure.

[0011] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the methods according to embodiments of this disclosure.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a schematic flowchart of a large-model-based sample detection method according to an embodiment of the present disclosure; Figure 2 This is a schematic diagram illustrating the acquisition of correctness indicators according to an embodiment of the present disclosure; Figure 3 This is a schematic diagram illustrating the acquisition of pseudo-feature indicators according to an embodiment of the present disclosure; Figure 4 This is a schematic diagram illustrating the acquisition of pseudo-feature indicators according to another embodiment of the present disclosure; Figure 5 This is a schematic diagram illustrating the difficulty index obtained according to an embodiment of the present disclosure; Figure 6 This is a schematic diagram illustrating the difficulty index obtained according to another embodiment of this disclosure; Figure 7This is a flowchart illustrating a large-model-based sample detection method according to another embodiment of the present disclosure; Figure 8 This is a flowchart illustrating a large-model-based sample detection method according to another embodiment of the present disclosure; Figure 9 This is a schematic flowchart of a model training method for sample detection according to an embodiment of the present disclosure; Figure 10 This is a schematic diagram of the structure of a large-model-based sample detection device according to an embodiment of the present disclosure; Figure 11 This is a schematic diagram of the structure of a large-model-based sample detection device according to another embodiment of the present disclosure; Figure 12 This is a schematic diagram of a model training apparatus for sample detection according to an embodiment of the present disclosure; Figure 13 This is a schematic diagram of a model training apparatus for sample detection according to another embodiment of the present disclosure; Figure 14 This is a block diagram of an electronic device used to implement the large model-based sample detection method / model training method for sample detection according to the embodiments of this disclosure. Detailed Implementation

[0014] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0015] The terms “first,” “second,” etc., used in this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion, such as including a series of steps or units. A method, system, product, or apparatus is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0016] It should be noted that, unless it is explicitly stated that there is a sequential order of execution between different operations, or that there is a sequential order of execution between different operations in terms of technical implementation, the execution order between multiple operations may not be significant, and multiple operations may be executed simultaneously.

[0017] Artificial intelligence models learn knowledge through samples. Taking large-scale models as an example, they construct knowledge systems by learning from information in training samples. For instance, large-scale models learn the syntax, semantics, and logical relationships of a language through massive amounts of text data. These training samples are a crucial means for large-scale models to acquire knowledge, determining the scope and depth of knowledge they can master.

[0018] The quality and diversity of training samples directly affect model performance. High-quality, diverse samples help the model learn a wider range of more accurate features and patterns, thus performing better in the corresponding task.

[0019] However, in practice, the samples used to train the model may contain noise. Noisy samples (such as annotation noise, content noise, etc.) can affect model performance.

[0020] In view of this, this disclosure proposes a sample detection method based on a large model, such as... Figure 1 As shown, it includes: S101, extracts the latent state features of the target sample based on a large model.

[0021] Large models can be categorized as Large Language Models (LLMs). Large language models refer to specific types of large-scale models specifically designed for processing text data. These models are neural network-based natural language processing models that can be used to generate, understand, and process text data. Large language models can have hundreds of billions of parameters, generate high-quality text, and can be used for various natural language processing tasks, such as question answering, text generation, and dialogue systems.

[0022] Large language models possess excellent reasoning and few-shot learning capabilities. This is because large language models are generally large in scale and can be trained on a vast number of samples. Based on such large models, accurate semantic understanding can be achieved.

[0023] In addition, large models can also be multimodal large language models (MLLM).

[0024] In one possible implementation, the image can be processed based on the image processing model in the multimodal large model to extract the text description of the image content, and the text description can be processed based on the text processing model of the multimodal large model to obtain the hidden state representation.

[0025] Another possible implementation is to use the features of the output of the high-level decoding blocks of MLLM to construct the latent state features of the target sample.

[0026] The target sample can be any sample in the sample set. That is, each sample in the sample set can be used as a target sample for detection.

[0027] Latent state features reflect the internal representations and learned features generated by the large model when processing input target samples. In the case of a large model based on a decoding block architecture (such as a Transformer), each decoding block (also called a decoding layer) outputs a feature representation. These feature representations capture multi-dimensional information such as semantics, syntax, and context. Latent state features can be at least one feature representation of a high-level decoding block, or they can be constructed based on these feature representations.

[0028] S102, based on multiple probes processing the hidden state features respectively, to obtain the evaluation index output by each probe for the target sample; wherein, multiple probes output evaluation indexes from different dimensions.

[0029] During implementation, multiple probes are used to process the hidden state features to evaluate the sample quality of the target sample in different dimensions.

[0030] S103, perform consistency analysis on the obtained multiple evaluation indicators to obtain the noise detection results of the target sample.

[0031] In practice, at least one latent state feature can be extracted from the same target sample. When multiple latent state features are extracted, multiple probes are used to process each latent state feature separately, resulting in multiple evaluation metrics (including individual evaluation metrics for each feature) corresponding to each latent state feature. These multiple evaluation metrics can then be used together for subsequent consistency analysis (including consistency analysis between evaluation metrics from different probes) to obtain the noise detection results for the target sample.

[0032] In the case of extracting a hidden state feature, in S103, the consistency analysis between the evaluation indicators of different probes is performed on the evaluation index of the hidden state feature to obtain the noise detection result.

[0033] The noise detection results can be in a binary classification form, for example, 0 indicates the presence of noise and 1 indicates the absence of noise.

[0034] The noise detection results can also be in a multi-class form, which can be represented by "0", "1" and "2", where "0" represents a true noise sample, "1" represents a true noise sample and "2" represents a false noise sample.

[0035] In this embodiment, the latent state features extracted from a large model can capture the deep semantic and structural information of the target sample within the internal processing mechanism of the large model. This information is more expressive and generalizable than traditional hand-crafted features or features extracted from shallow models, enabling subsequent probes to more effectively extract evaluation metrics for the target sample. By evaluating the sample from different dimensions using multiple probes, the potential problems of the target sample can be captured more comprehensively. Furthermore, multiple probes perform evaluation analysis based on unified latent state features, and the evaluation metrics output by each probe have a unified evaluation basis, which helps improve the accuracy of subsequent consistency analysis. A comprehensive judgment based on consistency analysis of multiple evaluation metrics can effectively reduce the errors that may be caused by a single probe. In summary, the method in this embodiment can detect noise in the target sample by representing consistency, and can automatically identify the noise status of the target sample.

[0036] In some embodiments, extracting latent state features of target samples based on a large model can be implemented as follows: inputting the target sample into a large model and obtaining the feature representation output by at least one decoding block of the large model; constructing latent state features based on at least one feature representation.

[0037] The large model may include a decoder, which consists of multiple decoder blocks with the same structure connected in series.

[0038] In implementation, within the large model, each decoding block of the decoder outputs feature representations containing rich semantic, syntactic, and contextual information. During implementation, the feature representations output by higher-level decoders are selected to construct the latent state features.

[0039] In practice, when extracting a hidden state feature, the feature representations output by the decoded blocks that satisfy the target condition can be selected to construct the hidden state feature. This target condition is, for example, that the decoded block is located within the last k decoded blocks (k being a positive integer), to ensure that the extracted hidden state feature contains rich and crucial information, thereby improving the quality of noise detection.

[0040] In implementation, when multiple latent state features need to be extracted, the latent state features are determined based on the feature representations output by multiple decoded blocks. For example, multiple sets of decoded blocks can be selected from multiple decoded blocks, each set containing at least two decoded blocks. For each set of decoded blocks, the feature representations output by the multiple decoded blocks in that set are individually averaged and then concatenated to obtain the latent state features corresponding to that set. Thus, different sets of decoded blocks obtain their respective latent state features, resulting in multiple latent state features.

[0041] In this embodiment, by leveraging the understanding capabilities of a large model and constructing latent state features through key decoding blocks in the model's internal processing mechanism, the model's perception of noisy samples during processing can be understood, thereby improving the accuracy of sample detection. Furthermore, by extracting feature representations from at least one decoding block, the cognitive knowledge of multiple neural network layers regarding the target sample can be better integrated, reducing detection errors caused by random mistakes and further improving the accuracy of noise detection for the target sample.

[0042] In some embodiments, in order to improve detection efficiency, a latent state feature is constructed based on at least one feature representation. This can be implemented by determining the latent vectors of the last token in the feature representations of multiple decoding blocks as latent state features.

[0043] When the target sample is text, the forward maximum matching method (FMM) can be used for text segmentation. Alternatively, the backward maximum matching method (BMM) and bidirectional maximum matching methods can also be used. Of course, the specific method of text segmentation is not limited in this embodiment; any method that can segment the text is acceptable.

[0044] The last token is the last word element in the target sample after word segmentation. The latent vector of the last token can learn the knowledge of the previous tokens, so using it as a latent state feature can effectively capture the features of the target sample to complete the target sample detection.

[0045] In implementation, since the decoder typically consists of multiple decoding blocks, the latent vector of the last token can be extracted from each decoding block. The latent vectors of the last token extracted from higher-level decoding blocks can be used as independent latent state features. The higher-level decoding blocks can be defined according to the structure of the large model (e.g., the last K decoding blocks), and this disclosure does not limit this.

[0046] In practice, the latent vector of the last token can be as shown in equation (1). (1) in, Let represent the latent vector of the last token, where each element is a real number (R), and d represents the dimension of the latent vector. This represents the feature representation of the corresponding decoded block output.

[0047] In this embodiment, extracting the latent vector of the last token as a latent state feature simplifies the latent state feature construction process, further reduces the amount of data processed subsequently, and improves the detection efficiency of the target sample. Moreover, the latent vector of the last token is compatible with the content learned by the large model from its previous tokens, and also ensures the accuracy of noise detection.

[0048] In some embodiments, constructing latent state features based on at least one feature representation can also be implemented by: selecting multiple decoding blocks as target decoding blocks from high-level decoding blocks in a large model. The method of selection is not limited in this disclosure. Then, for each target decoding block in at least one decoding block, the following steps are performed: Step A1: Obtain the latent vector corresponding to the last token in the feature representation of the target decoded block, and use it as the first reference feature; Step A2: Obtain the second reference features of the feature representations of multiple consecutive decoded blocks corresponding to the target decoded block; the second reference features include pooling features and / or the latent vector of the last token in each feature representation.

[0049] In practice, multiple consecutive decoding blocks are determined by the location of the target decoding block. For example, if the target decoding block is the i-th decoding block, then multiple consecutive decoding blocks within its preset neighborhood are considered as multiple consecutive decoding blocks of the i-th decoding block.

[0050] For example, when the decoder comprises (k+n) concatenated decoder blocks, and the i-th decoder block is the target decoder block, feature representations from layers i-2, i-1, and i can be selected to construct a second reference feature. In implementation, the feature representations from layers i-2, i-1, and i can be pooled separately and then concatenated to obtain the second reference feature. Alternatively, the latent vectors of the last segmented token can be concatenated from the feature representations of the decoded blocks in layers i-2, i-1, and i to obtain the second reference feature. , where i (k+n), where i, k, and n are all positive integers.

[0051] In another embodiment, the feature representations of multiple consecutive decoded blocks can be pooled and then concatenated to obtain a first sub-feature; then, the latent vector of the last token in the feature representations of the multiple consecutive decoded blocks can be concatenated to obtain a second sub-feature; finally, the first sub-feature and the second sub-feature can be concatenated to obtain a second reference feature. .

[0052] Step A3: Fuse the first reference feature and the second reference feature to obtain the hidden state features of the target decoding block.

[0053] For example, the hidden state characteristics of the target decoded block ( As shown in equation (2): (2) in, That is, the first reference feature. The second reference feature specifically refers to the aforementioned pooling feature, or the latent vector of the last token.

[0054] In another embodiment, the hidden state features of the target decoded block As shown in equation (3): (3) In equation (3), That is, the first reference feature. This is the latent vector of the last token in the second reference feature. This is the pooling feature in the second reference feature.

[0055] In this embodiment of the disclosure, by fusing the first reference feature and the second reference feature, the information of the target decoding block and its context decoding block can be integrated to capture richer semantic and contextual features, so as to construct stable hidden state features to extract evaluation indicators for noise detection.

[0056] In some embodiments, the plurality of probes may include a difficulty probe and a correctness probe. The difficulty probe assesses the difficulty of the large model in understanding the content of the target sample, while the correctness probe assesses the confidence level of the large model in its response to the target sample. Low confidence indicates that the large model has difficulty processing the target sample. High difficulty in understanding indicates that the target sample is a potentially difficult sample. Therefore, by processing the latent state features separately based on multiple probes, an evaluation index is obtained for each probe's output on the target sample, including steps B1 and B2, which can be specifically implemented as follows: Step B1: Input the latent state features into the difficulty probe to obtain the difficulty index of the target sample; the difficulty index is used to measure the semantic complexity of the target sample.

[0057] The difficulty index is used to indicate the inherent complexity of a large model in processing target samples. For example, samples with long reasoning chains should be rated "difficult"; complex questions containing multiple entity relationships should be rated "difficult"; and simple factual questions should be rated "not difficult".

[0058] A difficulty probe is a module used to measure the difficulty of understanding samples. It can be a simple linear classifier or a complex neural network. The purpose of a difficulty probe is to map latent state features onto a difficulty metric (d).

[0059] The hidden state features are input into the difficulty probe to obtain the difficulty index. The difficulty index (d) can be a continuous value (such as a score between 0 and 1).

[0060] Step B2: Input the hidden state features into the correctness probe to obtain the correctness index; the correctness index is used to measure the confidence of the large model in correctly answering the target sample.

[0061] During implementation, the internal information (i.e., hidden state features) of the large model before generating the answer is obtained for judgment. This can provide a predictive signal of whether the large model thinks it will answer correctly, and this predictive signal is the correctness indicator.

[0062] During implementation, the hidden state features can be input into the correctness probe, and the hidden state features can be processed using the linear mapping relationship of the correctness probe and the activation function layer to obtain the correctness index.

[0063] When implementing, such as Figure 2 As shown, the hidden state features are input into the correctness probe 21, and processed based on the linear mapping relationship 211 and the activation function layer 212 in the correctness probe 21 to obtain the correctness index.

[0064] For example, the processing procedure for the correctness probe is shown in equation (4): (4) in, As an indicator of accuracy, This is the weight matrix. These are bias terms, used to construct a linear mapping relationship for the correctness probe; This is the Sigmoid function, i.e., the activation function layer.

[0065] In addition, the Softmax function can be used instead of the Sigmoid function to evaluate the correctness metric.

[0066] In this embodiment of the disclosure, a linear mapping relationship and an activation function layer are used to map the latent state features of the large model's understanding of the target sample into a correctness index. This allows for the measurement of the large model's confidence in the answer based on the internal representation of the large model, thereby improving the reference value of the obtained correctness index and increasing the accuracy of noise detection for the target sample.

[0067] In this embodiment of the disclosure, steps B1 and B2 implement a difficulty index based on a difficulty probe to measure the semantic complexity of the target sample, i.e., to measure the difficulty of the large model's internal understanding mechanism of the target sample; the correctness index obtained through a correctness probe can be used to measure the large model's confidence in processing the target sample. These two dimensions are complementary in understanding the target sample, and consistency analysis can be used to assess the noise level of the target sample.

[0068] In some embodiments, when the evaluation metrics include both difficulty and accuracy metrics, performing a consistency analysis on the obtained multiple evaluation metrics to obtain the noise detection results of the target sample can be implemented as follows: In step C1, if the difficulty index is lower than the first threshold and the correctness index is lower than the second threshold, the target sample is determined to be a noise sample.

[0069] During implementation, if the difficulty index is below the first threshold, it means that the large model thinks the target sample is simple, and if the correctness index is below the second threshold, it means that the correctness state of the large model is very poor (i.e., it is not confident in its answer). Combining the inconsistency between the two, it can be determined that the target sample may be a noise sample.

[0070] In step C2, if the difficulty index is higher than the third threshold and the correctness index is lower than the fourth threshold, the target sample is determined to be a truly difficult sample.

[0071] A difficulty index above the third threshold indicates that the large model considers the target sample inherently difficult, while a correctness index below the fourth threshold indicates that the correctness state within the large model is low. Since the large model considers the target sample inherently difficult, a low correctness index is reasonable. Therefore, based on the consistency between the two statements, it can be determined that the target sample is not a true noise sample, but a truly difficult sample.

[0072] It should be noted that the first threshold and the third threshold can be the same or different, and the fourth threshold and the second threshold can be the same or different. The first threshold, the second threshold, the third threshold, and the fourth threshold can all be set based on the actual situation, and this disclosure does not limit them.

[0073] In this embodiment of the disclosure, the threshold-based approach can quickly analyze the consistency between the difficulty index and the correctness index in the analysis of sample difficulty, thereby easily and quickly identifying whether the target sample is a noise sample or a truly difficult sample.

[0074] For different target samples, the efficiency of dataset utilization can be improved based on the classification results. For example, highly confident noisy samples can be removed from the dataset, while truly difficult samples can be trained specifically to improve the ability of the corresponding neural network model to handle such samples.

[0075] In some embodiments, in addition to the aforementioned difficulty and correctness metrics, a spurious feature metric may be introduced to identify whether a large model focuses on irrelevant or noisy segments when processing target samples.

[0076] Pseudo-features are features that appear relevant to the answer but are not actually causally related. These features can mislead the model, causing it to perform well on training data but poorly on real-world applications or test data. The presence of pseudo-features is a common cause of overfitting and poor generalization.

[0077] The method for obtaining pseudo-feature indicators can be implemented as follows: processing the hidden state features based on pseudo-feature probes to obtain pseudo-feature indicators; pseudo-feature indicators are used to measure the degree to which the decisions made by the large model depend on unexpected information from the target sample.

[0078] Unintended information refers to parts of the target sample that are meaningless, redundant, or may introduce interference to the current task, such as noise, stop words, irrelevant context, etc.

[0079] During implementation, pseudo-feature indicators can be used to detect whether there are format biases, structural noise, or spurious correlation features in the target samples that make it difficult for large models to process the target samples.

[0080] Format bias refers to the tendency of large models to learn the format or representation of target samples rather than their intrinsic meaning. For example, if target samples are always presented in a specific order or format, a large model might learn that format rather than the actual content of the target samples.

[0081] Structural noise refers to noise or interference present in the target sample that is unrelated to the actual content of the target sample. This noise may originate from errors during sample collection, processing, or transmission, or from the inherent variability of the target sample itself. Structural noise can mislead large models, causing them to learn incorrect associations.

[0082] False correlation features refer to features that appear to be correlated with the target sample in the dataset, but do not actually affect the target sample. This correlation may be accidental or due to biases in sample collection. Related task processing models may incorrectly learn these false correlation features, thus affecting their generalization ability on new data.

[0083] These issues can all lead to model overfitting. Therefore, using pseudo-feature indicators as one of the evaluation metrics for target samples allows for analysis of the large model's understanding mechanism of the target sample based on its internal representation. This helps to analyze whether pseudo-features exist in the target sample and to what extent they exist, thereby enabling a correct assessment of the target sample and improving the detection accuracy.

[0084] In the embodiments of this disclosure, the latent state features are processed based on the pseudo-feature probe to obtain pseudo-feature indicators, which can be implemented in different ways, such as pseudo-feature indicator acquisition scheme 1 and pseudo-feature indicator acquisition scheme 2.

[0085] Scheme 1 for obtaining pseudo-feature indicators can be implemented as follows: input the latent state features into the pseudo-feature probe, and process the latent state features using the linear mapping relationship of the pseudo-feature probe and the activation function layer to obtain the pseudo-feature indicators. In this implementation, the pseudo-feature probe needs to be optimized through training, and the training method will be described later.

[0086] When implementing, such as Figure 3 As shown, the latent state features are input into the pseudo-feature probe 31, and processed based on the linear mapping relationship 311 and the activation function layer 312 in the pseudo-feature probe 31 to obtain the pseudo-feature index.

[0087] In some embodiments, if the pseudo-feature index is high (e.g., greater than the pseudo-feature index threshold), it indicates that the large model's attention is drawn to irrelevant content during critical decision-making, and the target sample is more likely to be format contamination, background noise, instruction template interference, splicing errors, labeling mismatches, etc.

[0088] If the pseudo-feature index is low (e.g., not greater than the pseudo-feature index threshold), it means that the large model is mainly looking at truly relevant content, and these samples are more like "clean" samples.

[0089] In practice, based on the linear mapping relationship and the activation function layer, the pseudo-feature index can be obtained as shown in equation (5): (5) in, These are pseudo-feature indicators. This is the weight matrix. For bias terms; This is the Sigmoid function, i.e., the activation function layer.

[0090] In addition, the Softmax function can be used instead of the Sigmoid function for evaluating pseudo-feature indicators.

[0091] In this embodiment, by processing latent state features based on pseudo-feature pointers, the "content" of pseudo-features in the target sample can be identified, thereby obtaining pseudo-feature indices. Based on these pseudo-feature indices, combined with the aforementioned difficulty and correctness indices, all three indices are derived from latent state feature analysis, thus achieving noise sample detection based on the internal representation of a large model. Since all three indices are derived from the same latent state feature, they have a good foundation for consistent analysis, thereby improving the accuracy and stability of noise sample detection.

[0092] Scheme 2 for obtaining pseudo-feature indicators can be implemented as follows: steps D1-D2: Step D1: Identify the tokens containing unexpected information in the target sample.

[0093] During implementation, the target sample can be segmented to obtain segmentation results. Each segmentation result includes at least one word element. The tokens corresponding to the segmentation results are divided into relevant tokens and irrelevant tokens. Relevant tokens refer to those directly related to the task's semantics, such as keywords in the question, question requirements, main entities / time and location, etc. Irrelevant tokens refer to those unrelated to the task's semantics, such as redundant formatting symbols (###, ---, bracket styles), template language, long background paragraphs unrelated to the question, and interspersed noisy text, etc. Irrelevant tokens are essentially tokens containing unexpected information.

[0094] During implementation, predefined rules can be used to identify unexpected information tokens based on task requirements and domain knowledge. For example, tokens for punctuation marks (such as redundant commas and periods) and special characters (such as garbled text and non-standard symbols) can be used as unexpected information tokens. Alternatively, the form of entity words can be identified, and non-entity word lexical units can be determined as unexpected information tokens.

[0095] Step D2, based on the pseudo-feature probe, performs the following operations, including steps D21-D23: Step D21: Based on interpretability technology, determine the total contribution of the unexpected information token to the hidden state features to obtain the first information value.

[0096] Step D22: Determine the total contribution of all tokens in the target sample to the hidden state features, and obtain the second information value; Step D23: Determine the ratio of the first information value to the second information value as the pseudo-feature index.

[0097] When implementing, such as Figure 4As shown, after identifying relevant and irrelevant tokens in the target sample, the total contribution of unexpected information tokens to the latent state features is calculated based on interpretability technology, which is used as the first information value; and the total contribution of all tokens in the target sample to the latent state features is obtained as the second information value; and the pseudo-feature index is obtained based on the ratio of the first information value and the second information value.

[0098] Interpretability techniques can include integrated gradients, gradient × input, and linear decomposition approximation.

[0099] Integral gradient is a method to measure the contribution of a target sample to the prediction of a large model. It is achieved by calculating the average gradient of the input features along the path from the reference point (such as the zero vector or random noise) to the target word.

[0100] Gradient × Input is a feature attribution method that measures the contribution of each word in the target sample to the model output by multiplying the target sample by the model's gradient. It intuitively reflects the impact of changes in input features on the model output.

[0101] Linear decomposition approximation is a method that approximates the output of a complex model as a linear combination of target samples. It linearizes the model's behavior, decomposing the model output into the contributions of individual tokens, thereby explaining the model's decision-making process.

[0102] The contribution of the i-th word in all tokens of the target sample obtained based on interpretability techniques As shown in equation (6): (6) in, Represents the attribution function. For the i-th word element in the target sample, This represents the characteristics of the hidden state.

[0103] Among them, the contribution of the j-th unexpected information token. As shown in equation (7): (7) in, Let be the token of the j-th word in the unexpected information.

[0104] For example, pseudo-feature indicators As shown in equation (8): (8) in, Indicates the first information value. This represents the second information value.

[0105] In another embodiment, multiple first information values ​​and multiple second information values ​​can be obtained using various methods in interpretability techniques. For example, taking integral gradients and gradient × input as an example, corresponding first and second information values ​​are obtained based on the integral gradient and gradient × input, respectively. The first information values ​​obtained by accumulating and averaging the integral gradient and gradient × input are used to obtain updated first information values; the corresponding second information values ​​obtained by accumulating and averaging the integral gradient and gradient × input are used to obtain updated second information values; and the pseudo-feature index is obtained based on the ratio of the updated first and second information values.

[0106] In this embodiment of the disclosure, by combining pseudo-feature probes and interpretability techniques, the contribution of unexpected information tokens to latent state features can be more accurately identified and quantified. By comparing the contribution of unexpected information tokens with that of all tokens, the impact of noise on model decision-making can be analyzed more deeply, thereby improving the accuracy of sample detection.

[0107] In some embodiments, latent state features are input into a difficulty probe to obtain a difficulty index of the target sample, which can be implemented in various ways. For example, embodiments of this disclosure provide difficulty index acquisition scheme 1 and difficulty index acquisition scheme 2.

[0108] Scheme 1 for obtaining the difficulty index can be implemented as follows: input the latent state features into the difficulty probe, and process the latent state features using the linear mapping relationship of the difficulty probe to obtain the difficulty index. The difficulty probe in this scheme needs to be determined through model training, and the training process will be described later.

[0109] When implementing, such as Figure 5 As shown, the hidden state features are input into the difficulty probe 51, and the difficulty index is obtained by processing based on the linear mapping relationship 511 in the difficulty probe 51.

[0110] Specifically, the difficulty index is shown in equation (9): (9) in, As a difficulty indicator, These are features of the hidden state. This is the weight matrix. For bias terms, and The parameters are obtained by training and adjusting the difficulty probe, and are used to express the linear mapping relationship of the difficulty probe.

[0111] In this embodiment, a linear transformation maps high-dimensional features to a low-dimensional space, making the difficulty index easier to understand and use. This approach not only reduces computational complexity but also more intuitively reflects the complexity of the target sample.

[0112] Difficulty indicator acquisition scheme 2 can be implemented as follows: Based on the difficulty probe, perform the following operations, including steps E1-E3: Step E1: Determine multiple degrees of uncertainty for the target sample across multiple dimensions; the multiple degrees of uncertainty include at least one accidental uncertainty and / or at least one cognitive uncertainty; accidental uncertainty indicates that the target sample has inherent noise, leading to uncertainty in the processing results of the large model on the target sample; cognitive uncertainty indicates that the large model has knowledge blind spots due to a lack of training data; the multiple degrees of uncertainty include results determined based on latent state features; Among them, random uncertainty can be calculated from the probability distribution predicted by the large model itself.

[0113] Cognitive uncertainty usually needs to be calculated through changes in large model parameters or multiple inferences.

[0114] In some embodiments, the output entropy can be calculated from the probability distribution of the large model output (such as the probability vector after Softmax), and then the output entropy can be normalized to obtain the uncertainty of the target sample.

[0115] Specifically, the large model processes the hidden state features of the target sample to obtain the probability distribution. The degree of uncertainty is shown in equation (10): (10) in, This indicates the degree of uncertainty obtained based on the output entropy. This represents the i-th probability in the probability distribution of the large model's output. A higher output entropy value indicates a greater degree of uncertainty.

[0116] In another embodiment, the first and second highest probabilities can be extracted from the probability distribution output by the large model; the degree of uncertainty is determined based on the difference between the first and second highest probabilities. The probabilities output by the large model are then arranged in descending order, with the first ranked probability being the first highest probability and the second ranked probability being the second highest probability. Again, using the probability distribution output by the large model as... For example, > >…> ,in The highest probability is... It is the second most probable.

[0117] The smaller the difference between the first and second highest probabilities, the more hesitant the large model is, and the higher the degree of uncertainty. An example degree of uncertainty is shown in equation (11): (11) in, This represents the degree of uncertainty based on the difference between the first and second highest probabilities.

[0118] Furthermore, the degree of uncertainty can also be inversely related to the difference between the first and second highest probabilities.

[0119] In some embodiments, multiple answers can be sampled from the same target sample. The greater the difference between the answers (i.e., the greater the disagreement), the higher the degree of uncertainty. The larger it is.

[0120] In practice, Monte Carlo (Dropout) can be used to infer the target sample. Specifically, during inference, the Dropout layer is kept on, and the target sample is forwarded T times to obtain T different predicted probability distributions. The mean and variance of these T distributions are then calculated.

[0121] Specifically, for the target sample, T independent forward propagations are performed repeatedly. During each propagation, all Dropout layers randomly generate different neuron masks according to preset probabilities, resulting in differences in the probability distribution of the network output. The output results of the T forward propagations (e.g., classification probability distribution) are collected; the mean and variance are calculated based on the output results of the T forward propagations to quantify the uncertainty of the large model prediction; where the mean represents random uncertainty; the larger the variance (e.g., above the variance threshold), the higher the uncertainty of the target sample, and the smaller the variance (e.g., below the variance threshold), the lower the uncertainty of the target sample.

[0122] In another embodiment, multiple different models or different checkpoints of the same model can be used to calculate the variance of the prediction results for the same target sample to measure uncertainty. For example, multiple task models can be used to perform the same task on the target sample, and the difference between the output results of different task models can be determined. The larger the variance between different output results (e.g., above an uncertainty threshold), the higher the uncertainty of the target sample; the smaller the variance (e.g., below the uncertainty threshold), the lower the uncertainty of the target sample.

[0123] In another embodiment, a Bayesian neural network can be used to quantify the uncertainty of the target sample through a probability distribution. In practice, T sets of weight vectors are sampled from the trained weight vectors. For the hidden state features of the target sample, forward propagation is performed using each of the T sets of weight vectors, and this process is repeated T times independently. The mean and variance are calculated based on the output of the T forward propagations. A larger variance (e.g., above a variance threshold) indicates a higher degree of uncertainty for the target sample, while a smaller variance (e.g., below a variance threshold) indicates a lower degree of uncertainty.

[0124] In another embodiment, Shannon entropy can be calculated using the probability distribution of a large model output (such as a probability vector after softmax). The higher the Shannon entropy value, the greater the uncertainty of the target sample.

[0125] Since Shannon entropy is a measure of overall uncertainty, it cannot be classified separately as accidental or cognitive uncertainty.

[0126] If the high value of the Shannon entropy predicted by the large model is due to the ambiguity of the sample itself, then the dominant factor is random uncertainty.

[0127] If the high value of Shannon entropy stems from insufficient knowledge in the large model, then the dominant factor is cognitive uncertainty.

[0128] Step E2: Identify the noise contamination level of the target sample in multiple dimensions.

[0129] Noise contamination level is used to quantify the likelihood that the target sample is contaminated by noise.

[0130] The degree of noise pollution can include pseudo-characteristics and / or outlier rate.

[0131] The pseudo-feature indicators have been described above, and will not be repeated here in the embodiments disclosed herein.

[0132] Outlier is used to measure the difference between the latent state features of a target sample and the typical distribution of the latent state features of the training sample set.

[0133] In one embodiment, the outlier degree of the target sample can be determined based on Mahalanobis distance, as shown in equation (12): (12) Where x represents the hidden state feature of the target sample. This represents the mean vector of the latent state features of all samples.

[0134] In another embodiment, the outlier degree of the target sample can be determined based on its local density. Specifically, this can be implemented as follows: determining the similarity between samples based on the feature vectors of multiple samples, such as determining the Euclidean distance between feature vectors as the similarity; determining the neighborhood range of the target sample and the neighborhood range of its neighboring samples based on the similarity; determining the local density of the neighboring samples based on the neighborhood range of the neighboring samples, and determining the local density of the target sample based on the neighborhood range of the target sample; determining the average of the ratios between the local densities of multiple neighborhood samples and the local density of the target sample as the outlier degree of the target sample. If the ratio is greater than a specified value, the target sample is determined to be an outlier.

[0135] Taking a target sample as an example, determining the neighborhood range of a target sample based on similarity can be implemented as follows: For the target sample The similarity between the target sample and other samples is sorted from largest to smallest, and the top k samples are determined as the neighborhood samples within the target sample's neighborhood range.

[0136] Among them, the sparsity relationship can be modeled using K-distance to represent the sparsity of the target samples, as shown in Equation (13): (13) in, K-distance represents the similarity between a target sample and its Kth ranked neighbor sample within its neighborhood. Indicates the target sample. This represents the k-th domain sample.

[0137] Based on K-distance, the reachability distance of the target sample is determined. The reachability distance is determined from the target sample... to neighborhood samples The distance (which can also be understood as similarity) has a lower limit, i.e. The K-distance is shown in equation (14): (14) In expression (17), Indicates the reachability distance of the target sample; Represents the neighborhood samples of the target sample K-distance; This indicates the similarity between the target sample and its neighboring samples.

[0138] The purpose of reachability distance is to calculate the target sample to neighborhood samples At that distance, it will not be because The local sparsity underestimates the actual distance. If If the neighborhood is sparse (i.e., the k-distance is large), then the distance to the nearest k-distance is less than 1. The K-distance is used as a lower bound to reduce the underestimation of distance caused by local sparsity. In practical applications, there may be noisy points in the samples, and the local sparsity of these noisy points may negatively affect the clustering or anomaly detection results. By using reachable distance, this noise effect can be effectively corrected.

[0139] Based on reachability distance, the local density (LRD) of the target sample is determined, as shown in Equation (15), including: (15) in, This represents the average reachability distance of the target sample to all its neighbors.

[0140] The outlier of the target sample is obtained by calculating the mean of the ratios of the local density of the neighborhood samples to the local density of the target sample. As shown in equation (16): (16) in, This indicates the number of neighboring samples of the target sample; This represents the local density of the j-th neighborhood sample. This represents the local density of the target sample. If the ratio equals the specified value, it indicates a normal sample; if the ratio is greater than the specified value, the target sample is determined to be an outlier. The larger the ratio, the higher the outlier degree.

[0141] When using an autoencoder to determine outlier status, the latent state features of the target sample are input into the trained autoencoder to obtain the reconstruction error. If the reconstruction error is greater than an error threshold, the target sample is considered an outlier; if the reconstruction error is not greater than the error threshold, the target sample is considered a non-outlier. The larger the reconstruction error, the higher the outlier status of the target sample; the smaller the reconstruction error, the lower the outlier status.

[0142] Among them, reconstruction error ( As shown in equation (17): (17) in, These are the latent state features of the target sample. This represents the feature representation of the encoder output in an autoencoder. This represents the feature representation reconstructed by the decoder in the autoencoder.

[0143] In practice, outlier degree and / or pseudo-feature indicators can be used as indicators of the noise pollution level of the target sample.

[0144] Step E3: Determine the difficulty index of the target sample based on the degree of noise pollution and the degree of uncertainty in multiple dimensions.

[0145] Among these factors, the level of noise pollution can be used as a gating signal. When both the level of uncertainty and the level of noise pollution are high, the target sample is mostly "uncertain due to dirt." When the level of uncertainty is high but the level of noise pollution is low, the target sample is more like a "clean true difficult example," and the difficulty index is high.

[0146] In this embodiment of the disclosure, by evaluating the uncertainty and noise contamination levels across multiple dimensions, the complexity and processing difficulty of the target sample can be measured more comprehensively. It considers not only the inherent uncertainty of the target sample but also the possibility of contamination. This multi-dimensional evaluation approach provides richer information, helping to more accurately determine the difficulty of the target sample.

[0147] In some embodiments, the difficulty index of the target sample is determined based on multiple dimensions of noise pollution level and multiple degrees of uncertainty, which can be implemented as follows: Step F1 involves weighted summation of multiple degrees of uncertainty to obtain the index to be evaluated.

[0148] In practice, the multiple degrees of uncertainty obtained earlier can be normalized (for example, normalized to [0,1]), and then weighted and summed to obtain the index to be evaluated. As shown in equation (18): (18) in, This indicates the degree of uncertainty determined based on the output entropy. This indicates the degree of uncertainty determined based on the difference between the first and second highest probabilities. This indicates the degree of uncertainty in determining the same target sample through multiple samplings. , , It can be experience points.

[0149] In addition, Shannon entropy can be considered as another level of uncertainty and weighted together with the aforementioned three items to obtain the indicator to be evaluated. .

[0150] Step F2: Determine the gating signal based on the level of noise pollution; the gating signal is used to filter out uncertainties caused by dirty data.

[0151] In implementation, based on the aforementioned understanding, the noise pollution level includes pseudo-signal indicators and / or outlier characteristics. When the noise pollution level includes both pseudo-signal indicators and outlier characteristics, they are normalized, and the gated signal is used. As shown in equation (19): (19) in, Indicators representing pseudo-features, Indicates the degree of outlier.

[0152] Normalize the noise pollution level using only pseudo-feature indicators or outlier values, and then use the gated signal. As shown in equation (20): or (20) Step F3: Based on the gating signal processing of the evaluation index, obtain the difficulty index of the target sample.

[0153] Difficulty indicators during implementation As shown in equation (21): (twenty one) in, Indicates the indicator to be evaluated. This indicates a gating signal.

[0154] If the indicator to be evaluated is high (e.g., above the threshold of the indicator to be evaluated), but the pseudo-feature indicator or outlier is also high (e.g., the pseudo-feature indicator is above the threshold of the pseudo-feature indicator or the outlier is above the threshold of the outlier), in this case, the uncertainty caused by contamination may be high. Therefore, the difficulty indicator will be reduced to indicate that the sample is simple.

[0155] If the metric to be evaluated is high (e.g., higher than the threshold of the metric to be evaluated), but the pseudo-feature metric or outlier is low (e.g., the pseudo-feature metric is not higher than the threshold of the pseudo-feature metric or the outlier is not higher than the outlier threshold), then it may be a true hard case sample.

[0156] In one embodiment, when multiple degrees of uncertainty are determined based on output entropy, the difference between the first and second highest probabilities, multiple sampling, and Shannon entropy to obtain the evaluation index, and the degree of noise contamination is obtained based on pseudo-feature index and outlier degree, the difficulty index is obtained in the following way: Figure 6 As shown, the specific implementation can be as follows: First sub-degree is determined based on output entropy; second sub-degree is determined based on the difference between the first and second highest probabilities; third sub-degree is obtained based on multiple samplings of the same target sample; fourth sub-degree is obtained based on Shannon entropy; the first, second, third, and fourth sub-degrees are weighted and summed to obtain the evaluation index. Furthermore, a gating signal is determined based on pseudo-feature index and outlier degree; and the difficulty index of the target sample is obtained based on the evaluation index and the gating signal.

[0157] In this embodiment of the disclosure, the assessment of the degree of noise pollution can help identify noise in the data of the target sample, thereby assessing the data quality of the target sample and improving the accuracy of the identification of the difficulty index.

[0158] Regardless of how the difficulty index, correctness index, and fake feature index are obtained, consistency analysis of the multiple evaluation indices can be performed based on the following method to obtain the noise detection results of the target sample, including: classifying the target sample based on the semantic consistency of the difficulty index, correctness index, and fake feature index corresponding to at least one latent state feature of the target sample to obtain the sample category of the target sample; the sample category is divided based on the noisy situation.

[0159] In practice, as can be seen from the meaning of each evaluation indicator introduced above, each evaluation indicator expresses the processing difficulty of the target sample from different dimensions. Therefore, semantic consistency can be used to measure whether different indicators are consistent in their judgment of the reasons for or the difficulty of the target sample.

[0160] In implementation, classification can be based on at least one hidden state feature corresponding to a difficulty index, a correctness index, and a fake feature index. The classification can be hierarchical, for example, it can be divided into noisy samples, suspicious samples, and high-value true difficult examples, etc.

[0161] In this embodiment of the disclosure, by combining multiple evaluation metrics (such as difficulty metrics, accuracy metrics, and fake feature metrics), the situation of the target sample can be evaluated more comprehensively. Each metric reflects the characteristics of the sample from a different perspective. Combining these metrics can better achieve accurate classification of the target sample, thereby improving the detection accuracy of the target sample.

[0162] In some embodiments, when the evaluation metrics include difficulty metrics, accuracy metrics, and pseudo-feature metrics, a consistency analysis is performed on the obtained multiple evaluation metrics to obtain the noise detection results of the target sample, such as... Figure 7 As shown, it includes: S701, determine the first mismatch strength between the difficulty metric and the accuracy metric; the first mismatch strength is used to measure the consistency between the difficulty metric and the accuracy metric in describing the learning complexity of the target sample.

[0163] The difficulty index (d) and the accuracy index (c) are standardized to obtain the standardized difficulty index. and standardized correctness indicators Standardization can use Z-score to map different metrics to the same scale for processing.

[0164] During implementation, the first mismatch intensity is obtained based on the standardized difficulty and correctness indicators. As shown in equation (22): (twenty two) in, This indicates the difficulty threshold corresponding to the difficulty index. This represents the correctness threshold corresponding to the correctness metric. A mismatch strength greater than zero will only be obtained when both the difficulty metric and the correctness metric are below the difficulty threshold. This can also be understood as the target sample itself being very simple, but the large model is unsure of its own answer. In this case, the target sample is likely a noisy sample.

[0165] It should be noted that, as explained above, multiple difficulty indicators can be obtained based on the hidden state features corresponding to multiple decoding blocks. The average of these multiple difficulty indicators can then be calculated and standardized to obtain the average difficulty indicator. Based on the hidden state features corresponding to multiple decoded blocks, multiple correctness indicators are obtained. The average of these correctness indicators is calculated, and then standardized to obtain the mean of the correctness indicators. Based on the average difficulty index and the average correctness index, the first mismatch intensity is obtained based on equation (22).

[0166] S702, based on pseudo-feature index, determines the second mismatch strength; the second mismatch strength is used to measure the degree to which the decision-making process of a large model is affected by unexpected information.

[0167] During implementation, the pseudo-feature index(s) is standardized. If there are multiple pseudo-feature indices(s) with multiple latent state features, the average of these multiple indices(s) can be calculated to obtain the standardized pseudo-feature index mean. The corresponding second mismatch strength As shown in equation (23): (twenty three) Since the pseudo-feature index s reflects the proportion of the influence of unexpected information tokens in the input on the final representation, the higher the pseudo-feature index, the greater the mismatch, indicating that the reasoning process inside the large model is more severely polluted by unexpected information tokens.

[0168] S703 determines the third mismatch intensity based on the fluctuation of the difficulty index corresponding to the hidden state features of multiple decoding blocks and the fluctuation of the correctness index corresponding to the hidden state features of multiple decoding blocks.

[0169] High-quality samples should exhibit a consistent trend in difficulty and accuracy across different layers (i.e., decoding blocks), especially the last few layers. If the same sample shows significant fluctuations in the same metric output by probes at different layers, it indicates that a stable representation cannot be formed within the large model, which is a typical characteristic of noisy samples.

[0170] In implementation, multiple decoded blocks can be decoded blocks from the last k layers. Third mismatch strength. As shown in equation (24): (twenty four) in, ) represents the standardized value of the difficulty index of the Lth layer in the last k-layer decoded block. This represents the standardized value of the correctness index for layer L. This represents the variance of the difficulty index at different levels. These represent the variance of the accuracy metrics at different levels. This variance measures the variability in the model's understanding of the target sample at each level. High variance indicates significant variation in the target sample across different levels, suggesting that the target sample may be noisy.

[0171] S704, the first mismatch intensity, the second mismatch intensity and the third mismatch intensity are weighted and summed to obtain the noise score of the target sample; the noise detection result includes the noise score.

[0172] During implementation, the three mismatch terms are combined into a unified, sortable noise score, which can be used for large-scale data screening. This provides a directly usable quantitative evaluation metric for downstream training, data cleaning, and manual review.

[0173] In practice, the noise score (NoiseScore) is as shown in equation (25): (25) in, All of these are weight coefficients, which can be fixed values ​​after training or values ​​determined based on experience. This disclosure does not limit them.

[0174] In this embodiment, the noise level of a sample is evaluated through a multi-dimensional consistency analysis by combining difficulty, correctness, and pseudo-feature indicators. These evaluation indicators consider not only the complexity and correctness of the target sample but also the impact of unexpected information in the target sample on the decision-making of the large model. Based on multi-dimensional evaluation indicators, multiple mismatch intensities are obtained, enabling a more comprehensive assessment of the noise situation of the target sample and reducing the bias that may arise from a single indicator or mismatch intensity. The noise score obtained in this way can objectively measure the quality of the target sample, aiding in sample identification and processing.

[0175] In some embodiments, the target sample is classified based on the difficulty index, correctness index, and pseudo-feature index corresponding to at least one latent state feature of the target sample to obtain the sample category of the target sample, such as... Figure 8 As shown, it can be implemented as follows: S801, when the target sample corresponds to multiple hidden state features, determine the first statistical value of the difficulty index corresponding to each of the multiple hidden state features, the second statistical value of the correctness index corresponding to each of the multiple hidden state features, and the third statistical value of the pseudo feature index corresponding to at least one target hidden state among the multiple hidden state features.

[0176] During implementation, when multiple hidden state features exist, difficulty index, correctness index, and fake feature index are obtained for each hidden state feature.

[0177] For the first statistical value, statistical values ​​of difficulty indicators corresponding to multiple hidden state features can be calculated, such as the mean and median, and used as the first statistical value.

[0178] The calculation methods for the second and third statistical values ​​are similar to those for the first statistical value, and will not be elaborated upon here.

[0179] S802, matching the first, second, and third statistical values ​​with the grade features of the true noise samples and the grade features of the pseudo noise samples.

[0180] The rank features of real noise samples and pseudo noise samples can be predefined thresholds or feature distributions learned from training data.

[0181] S803, based on the matching results, determine the sample category of the target sample.

[0182] If the matching result indicates a successful match with the level features of a true noise sample, the sample is classified as a true noise sample; if the matching result indicates a successful match with the level features of a pseudo noise sample, the sample is classified as a pseudo noise sample.

[0183] It should be noted that the noise level characteristics can be set based on the actual situation, and this disclosure does not limit them.

[0184] In this embodiment, statistical values ​​of multiple evaluation metrics are introduced to further improve the accuracy and reliability of noise detection. This approach not only considers the evaluation metrics of individual latent state features, but also identifies true noise samples and false noise samples by using the characteristics of samples with different levels of statistical values, thereby improving the accuracy of detecting target samples.

[0185] In some embodiments, the sample category of the target sample is determined based on the matching result, which can be implemented as follows: if the target sample is determined to be a true noise sample if any of the following grade characteristics of a true noise sample are met; 1) The first characteristics of true noise include: the first statistical value is less than or equal to the first target value, the second statistical value is less than or equal to the second target value, and the third statistical value is greater than or equal to the third target value.

[0186] During implementation, the first target value is a1, the second target value is b1, and the third target value is... When c1 is the case, the target sample satisfies , , In this case, the target sample can be determined to be a true noise sample, where, This represents the first statistical value. This represents the second statistical value. This indicates the third statistical value.

[0187] in, The target sample is very easy to represent, and the large model considers the target sample to be very simple; This indicates that the accuracy prediction of the sample within the large model is highly uncertain, suggesting that there may be errors in the labeling or instructions for the target sample. This indicates that the pseudo-feature scores are abnormally high, suggesting that the reasoning process of the large model may be contaminated by irrelevant information. Under the condition that the aforementioned three conditions are met, the target sample is determined to be a true noise sample.

[0188] In this example, a1 can be -1, b1 can be -1, and c1 can be 1.

[0189] 2) The second characteristic of true noise includes: the first statistical value is less than or equal to the first target value, the second statistical value is less than or equal to the second target value, and the third mismatch intensity is greater than or equal to the fourth target value.

[0190] Continuing with the previous example, when the fourth objective value is... In the case that the target sample satisfies , , In this case, the target sample can be determined to be a true noise sample.

[0191] in, This indicates excessive interlayer variation, meaning that the understanding of the sample is highly inconsistent across different layers, possibly due to noise.

[0192] In this embodiment of the disclosure, by setting specific thresholds and conditions, clear classification criteria are provided, making the identification of true noise samples more accurate and reliable.

[0193] In some embodiments, the sample category of the target sample is determined based on the matching results, which can be implemented as follows: if the target sample is determined to be a pseudo-noise sample if any of the following pseudo-noise sample level characteristics are met; 1) The first characteristic of pseudo noise includes: the first statistical value is less than or equal to the fifth target value, the second statistical value is less than or equal to the sixth target value, and the third statistical value is greater than or equal to the seventh target value; The seventh target value is less than the third target value.

[0194] During implementation, given that the fifth objective value is a2, the sixth objective value is b2, and the seventh objective value is c2, and the target sample satisfies... , , In the case of a1 < a2 < 0, b1 < b2 < 0, 0 < c2 < c1, the target sample can be determined to be a pseudo-noise sample.

[0195] in, This indicates that the target sample may appear simple, but its accuracy is not high; This indicates that the large model has a low accuracy prediction for the target sample, suggesting that the target sample may not be easy to answer. A slightly higher score for a fake feature indicates that the target sample may contain irrelevant information. If the aforementioned three conditions are met, the target sample is determined to be a true noise sample.

[0196] For example, .

[0197] 2) The second feature of pseudo-noise includes: the third mismatch intensity belongs to the preset quantile; the preset quantile is obtained based on the statistics of multiple samples.

[0198] During implementation, the preset quantile can be: ,exist In this case, it means the third mismatch strength of the target sample. The third mismatch strength is greater than or equal to that of 80% of the samples in the sample set. This indicates that the sample has a high level of third mismatch strength, suggesting that the large model has significant interlayer instability when dealing with this sample.

[0199] In this embodiment of the disclosure, strict conditions are set so that only samples that meet the pseudo-noise characteristics in multiple dimensions are classified as pseudo-noise samples, thereby improving the accuracy of noise detection.

[0200] It should be noted that the aforementioned conditions can be adjusted according to specific tasks and datasets to make the classification of sample levels more adaptable.

[0201] In some embodiments, the sample category of the target sample is determined based on the matching results. This can be implemented as follows: if no match is found between the level features of the true noise sample and the level features of the pseudo noise sample, the target sample is determined to be a true hard sample.

[0202] If the sample fails to match the level characteristics of any of the aforementioned noise samples, it is determined to be a true hard sample.

[0203] For example, when the target sample satisfies , In this case, it means that the large model believes that the target sample is indeed difficult, but its correct prediction is very high, so the target sample is considered to be a high-value true difficult sample.

[0204] Directly identifying truly difficult examples is often challenging because they are typically samples that the model struggles to learn from, but are not caused by noise. Using strict matching methods may misclassify truly difficult examples as noise. Therefore, samples that fail to match with either true or false noise samples can be identified as truly difficult examples to improve the accuracy and reliability of truly difficult example discovery.

[0205] In some embodiments, to improve the efficiency of manual review, interpretability information can be generated for all samples submitted to the review queue. Specifically, this can be implemented as follows: if a target sample meets the manual review criteria, the interpretability information for the target sample is output; if the noise detection result is a score, the manual review criteria can be to review samples whose noise detection results are below a noise threshold. Samples below the noise threshold may be spurious noise samples, i.e., samples whose authenticity is uncertain. Samples classified as spurious noise can be determined to meet the manual review criteria.

[0206] Explanatory information includes at least one of the following: 1) Attention heatmap of tokens for the target sample; Attention heatmaps can visually demonstrate the level of attention a large model pays to each token (such as a word or sub-word) when processing text-based target samples. The specific implementation steps are as follows: The attention mechanism of the large model (such as the self-attention mechanism in Transformer) is used to calculate the attention weight of each token. This attention weight is then visualized as a heatmap, with darker colors indicating a higher level of attention from the large model to that token. In particular, heatmaps of tokens containing unexpected information can be output and displayed.

[0207] 2) Ranking of the contribution of multiple characters of the target sample's token to the latent state features.

[0208] The contribution ranking can show which tokens contribute the most to the hidden state features of the large model. The specific implementation steps are as follows: Calculate the contribution of each token to the hidden state features, which can be achieved by analyzing gradients or feature importance. Sort the tokens by contribution from largest to smallest and output the ranking results. In particular, the ranking of tokens containing unexpected information can be output.

[0209] In this embodiment of the disclosure, through these interpretable auxiliary tools, human reviewers can quickly locate the problem, thereby reducing the difficulty of human review and improving the accuracy of the review results.

[0210] In this disclosure, based on the same inventive concept, a model training method for sample detection is also proposed, used to train the aforementioned multiple probes, such as... Figure 9 As shown, it includes: S901, based on extracting latent feature representations of training samples from a large model; The weights (parameters) of the large model are fixed during subsequent training and are no longer updated; only the probe parameters are optimized. The extraction method for the latent feature representations of the training samples is the same as that for the target samples, and will not be repeated here.

[0211] Furthermore, extracting latent feature representations from the frozen large model enables the reading of stable semantic structures without changing the parameters of the large model, thereby improving the probe's generalization ability and computational efficiency during subsequent probe training.

[0212] S902, a probe trained based on latent feature representation; the probe is used to detect noise in target samples.

[0213] After freezing the large model, only a small number of probe model parameters need to be trained or adjusted to train the probe, thereby accelerating the training process.

[0214] The probe can be any of the following: difficulty probe, correctness probe, or pseudo-feature probe. Each probe can be trained based on the latent feature representation when the model parameters need to be optimized.

[0215] During implementation, in order to adapt to changes in data distribution during the continuous iterative optimization of the probe, it is necessary to periodically recalibrate the probe, the weights and thresholds required for generating noise scores, and the thresholds required for sample classification. Specifically, this can be implemented as follows: run the large model and related probes on the gold standard validation set, and record the noise detection results of the last sample on the gold standard validation set; optimize the weights and thresholds based on the noise detection results and the annotation information of the gold standard validation set.

[0216] The gold-label validation set is a set of precisely labeled samples used to evaluate and validate model performance. From the training sample set, samples are scored and categorized using the latent feature representations of the large model. At different training stages, suitable sample datasets are selected based on the categorization results to adjust the probe's model parameters, as well as the corresponding weights and thresholds. The categorization results can be, for example, divided into true hard examples, true noise examples, and pseudo-noise examples.

[0217] In this embodiment of the disclosure, the powerful feature extraction capability of the large model is utilized to capture complex patterns and semantic information in the training samples. The probe is trained based on the latent feature representation extracted by the large model, which can more accurately identify noisy samples. This allows the probe to learn which features are related to noise, thereby outputting relevant noise indicators more accurately and improving the detection effect of target samples.

[0218] In some embodiments, training the probe based on latent feature representation can be implemented as follows: when the probe includes a difficulty probe, a weak supervision signal of the training sample is obtained; the weak supervision signal is determined based on at least one of the following information: the heuristic reasoning complexity of the large model for the training sample, and the degree of divergence of the output results of multiple lightweight models for the training sample; the difficulty probe is trained based on the weak supervision signal.

[0219] Among them, the heuristic reasoning complexity of a large model for training samples can reflect the difficulty of the samples. For example, the more reasoning steps or the greater the computational resource consumption, the more difficult the training samples may be to process. Therefore, constructing unsupervised information based on heuristic reasoning complexity and / or divergence can improve the training efficiency of difficulty probes.

[0220] The heuristic reasoning complexity includes the length of the reasoning chain for the large model to process the training samples, and / or the number of logical operation steps for the large model to process the training samples.

[0221] The length of the inference chain reflects the number of inference steps a large model needs to perform when processing training samples. A longer inference chain usually means that the training samples are more complex and require more logical reasoning to reach a conclusion.

[0222] The number of logical operation steps reflects the number of logical operations a large model needs to perform when processing training samples. Logical operations represent the various computations and decision-making processes a large model performs when processing data. Logical operations include, but are not limited to, matrix operations, activation functions, and conditional judgments. More logical operation steps usually mean that the training samples are more complex, requiring more computational resources and time to process.

[0223] In this embodiment of the disclosure, by taking into account the inference chain length and the number of logical operation steps of a large model processing training samples, the difficulty of the samples can be assessed more accurately, so as to optimize the difficulty probe.

[0224] The divergence of outputs from multiple lightweight models for the same training sample reflects the uncertainty of the training sample. A higher divergence (e.g., above the divergence threshold) indicates that the training sample is more difficult to process; a lower divergence (e.g., below the divergence threshold) indicates that the training sample is easier to process.

[0225] Lightweight models can be neural network models with fewer parameters, distilled / compressed versions of larger models, or several open-source small models, etc. The method for determining the divergence of output results from multiple lightweight models for the same training sample is as follows: input the training sample into multiple lightweight models to obtain multiple prediction results, and compare these multiple prediction results to determine the degree of divergence between them.

[0226] In some embodiments, obtaining the weak supervision signal of the training samples can be implemented as follows: when the weak supervision signal is determined based on multiple pieces of information, the multiple pieces of information are weighted and summed to obtain the weak supervision signal.

[0227] In practice, when the weakly supervised signal includes heuristic reasoning complexity (including reasoning chain length and number of logical operation steps) and output divergence, the reasoning chain length, number of logical operation steps, and output divergence can be normalized and then weighted and summed to obtain the weakly supervised signal.

[0228] In this embodiment, combining multiple pieces of information can reduce problems caused by misjudgment due to a single piece of information. Combining multiple pieces of information can provide richer information for each sample, allowing the large model to learn and laying the foundation for subsequent, more difficult probes to output more accurate results.

[0229] In practice, a weak supervision signal is constructed using a weak supervision proxy signal. The training process for the difficulty probe is as follows: the training sample is input into the difficulty probe to obtain the predicted difficulty. Based on the difference between the predicted difficulty and the weak supervision signal, the first loss value is determined. The difficulty probe is trained based on the first loss value. If the convergence condition is met, the trained difficulty probe is obtained.

[0230] The convergence condition is that the first loss value is less than the loss threshold or the number of iterations is satisfied.

[0231] The loss function can be either mean squared error or quantile regression.

[0232] In this embodiment, by using weak supervision signals, the reliance on large amounts of precisely labeled data can be reduced, thereby improving training efficiency. By combining various information such as heuristic reasoning complexity and output divergence, the difficulty of training samples can be evaluated more comprehensively. This helps improve the accuracy of the difficulty probe, enabling it to more accurately identify high-difficulty samples.

[0233] In another embodiment, for the aforementioned difficulty index acquisition scheme 2, the difficulty probe includes an uncertainty determination module, a contamination determination module, and a difficulty index determination module. The uncertainty determination module is used to determine the degree of uncertainty of training samples in multiple dimensions.

[0234] The method for determining the degree of uncertainty has been explained above and will not be repeated here.

[0235] The contamination determination module is used to identify the degree of noise contamination in the training samples.

[0236] The degree of noise pollution can include pseudo-characteristics and / or outlier rate.

[0237] The methods for determining pseudo-feature indicators and outlier degree have been explained above and will not be repeated here.

[0238] The difficulty index determination module is used to determine the difficulty index of training samples based on the degree of noise pollution and multiple uncertainties.

[0239] In practice, when multiple degrees of uncertainty exist, the multiple degrees of uncertainty are weighted and summed to obtain the index to be evaluated; then, the gating signal is determined based on the degree of noise pollution; the gating signal is used to filter the uncertainty caused by dirty data; the index to be evaluated is processed based on the gating signal to obtain the difficulty index of the target sample.

[0240] The methods for determining the evaluation indicators, gating signals, and difficulty indicators have been explained above and will not be repeated here.

[0241] In this embodiment of the disclosure, by evaluating the degree of uncertainty and noise pollution in multiple dimensions, the complexity of the sample can be understood more comprehensively, reducing the misjudgment that may be caused by single-dimensional evaluation, thereby improving the accuracy of the trained difficulty index.

[0242] In some embodiments, training the probe based on latent feature representations can be implemented as follows: Step G1: Obtain the first classification signal; the first classification signal is used to indicate whether the large model gives a correct answer to the training samples.

[0243] In practice, each sample in a clean validation set can be labeled using a model (or manually labeled) as "whether the answer is correct" as a binary label, for example, 0 indicates incorrect and 1 indicates correct.

[0244] Step G2: Obtain the predicted correctness of the correctness probe for the latent feature representation.

[0245] Step G3 involves performing a binary classification operation on the predicted accuracy to obtain the first classification result.

[0246] In practice, the predicted correctness output by the correctness probe can be mapped to [0, 1]. Based on the threshold, binary classification is performed. If the value is less than the threshold, the first classification result is incorrect, and if the value is not less than the threshold, the first classification result is correct.

[0247] Step G4: Based on the first classification result and the first classification signal, optimize the model parameters of the correctness probe.

[0248] Based on the second loss value between the first classification result and the first classification signal, the model parameters of the correct probe are adjusted, and the trained correct probe is obtained when the convergence condition is met.

[0249] In this embodiment of the disclosure, supervised training using labeled data provides the correctness probe with a clear learning objective, enabling it to directly learn the relationship between input features and output labels, thereby improving the accuracy and precision of correctness index prediction.

[0250] In some embodiments, for the pseudo-feature probe of pseudo-feature indicator acquisition scheme 1, the probe is trained based on latent feature representation, including: Step H1: Obtain the second classification signal; the second classification signal is used to indicate whether the decision result of the large model depends more on unexpected information than a preset level. The process for identifying unexpected information during implementation has been described above and will not be repeated here.

[0251] In practice, the weights of each word in the unexpected information can be accumulated, and binary classification can be performed based on the accumulated results to obtain the second classification signal.

[0252] In some embodiments, obtaining a second classification signal can be specifically implemented as follows: obtaining the focus of a specified target attention head of a large model on unexpected information in training samples; and determining a second classification signal based on the focus.

[0253] The higher the number of layers in a large model, the closer it gets to the information fusion that "actually makes the decision." Therefore, multiple target attention heads can be multiple attention heads in the last n layers of the large model, where n is a positive integer. For example, in the case where the large model includes k cascaded attention heads, multiple attention heads from layers k-2, k-1, and k can be selected to obtain multiple target attention heads.

[0254] For each target attention head, before generating the answer (or predicting the next token), the large model assigns an attention weight (which can be understood as "the degree of glancing at") to each input token. The portion of these attention weights that fall on irrelevant tokens is summed up to obtain the "irrelevant focus ratio" of that target attention head.

[0255] In simple terms, for each target attention head: irrelevant focus ratio = (total attention allocated to irrelevant tokens) ÷ (total attention allocated to all input tokens). Since attention is normalized, the denominator is usually 1, so it can be directly understood as: irrelevant focus ratio = total attention weights allocated to irrelevant tokens.

[0256] In the final n layers of the large model, each layer has multiple target attention heads. The average irrelevant focus ratio of the multiple target attention heads in the same layer is calculated to obtain the irrelevant focus ratio of that layer. The average irrelevant focus ratio of the final n layers is then calculated to obtain the final irrelevant information focus degree.

[0257] The meaning of the final irrelevant information focus is as follows: if the final irrelevant information focus is high, it means that the model's attention is "drawn away by irrelevant content" when making key decisions, and the samples are more likely to be: format pollution, background noise, instruction template interference, splicing errors, labeling mismatches, etc.; if the final irrelevant information focus is low, it means that the large model is mainly looking at the truly relevant content, and these samples are more like "clean but difficult" true hard examples.

[0258] The final irrelevant information focus is compared with a threshold to obtain a second classification signal. For example, if the focus is greater than the threshold, it indicates excessive reliance on unexpected information; if it is less than or equal to the threshold, it indicates less reliance on unexpected information.

[0259] In this embodiment of the disclosure, the focus of multiple target attention heads is statistically analyzed to more intuitively understand the behavior of the model when processing samples, so as to further accurately assess the dependence of the large model on unexpected information, in order to construct a supervisory signal to optimize the pseudo-feature probe.

[0260] Step H2: Obtain the predicted pseudo-feature values ​​of the pseudo-feature probe for the latent feature representation; If the aforementioned second classification signal is obtained based on a target attention head, then the predicted pseudo-feature value also needs to be obtained based on an attention head at this point.

[0261] Step H3: Perform a binary classification operation on the estimated pseudo-feature values ​​to obtain the second classification result; Among them, the predicted pseudo-features can be classified as "dependent on unexpected information" or "not dependent on unexpected information".

[0262] In practice, the estimated pseudo-feature values ​​can be mapped to [0, 1], and binary classification can be performed based on a threshold. Values ​​less than the threshold indicate that the first classification result depends on unexpected information, while values ​​not less than the threshold indicate that it does not depend on unexpected information.

[0263] Step H4: Based on the second classification results and the second classification signal, optimize the model parameters of the pseudo-feature probe.

[0264] During implementation, the model parameters of the fake feature probe are optimized based on the difference between the second classification result and the second classification signal. If the convergence condition is met, the optimization of the fake feature probe is considered complete.

[0265] In this embodiment of the disclosure, by introducing a second classification signal to perform supervised training on the pseudo-feature probe, the degree of dependence of the large model on unexpected information when processing samples can be more intuitively understood. By optimizing the model parameters of the pseudo-feature probe, training efficiency and the accuracy of the pseudo-feature index can be improved.

[0266] In another embodiment, for the pseudo-feature indicator acquisition scheme 2, the pseudo-feature probe in this scheme can be used without training as needed.

[0267] Based on the same technical concept, this disclosure also proposes a sample detection device 1000 based on a large model, such as... Figure 10 As shown, it includes: The first extraction module 1001 is used to extract the latent state features of the target sample based on the large model; Processing module 1002 is used to process the latent state features based on multiple probes respectively, and obtain the evaluation index output by each probe for the target sample respectively; wherein, the multiple probes output the evaluation index from different dimensions; The analysis module 1003 is used to perform consistency analysis on the obtained multiple evaluation indicators to obtain the noise detection results of the target sample.

[0268] In some embodiments, the first extraction module 1001, such as Figure 11 As shown, it includes: The first acquisition unit 1101 is used to input the target sample into the large model and acquire the feature representation output by at least one decoding block of the large model; Construction unit 1102 is used to construct the hidden state features based on at least one feature representation.

[0269] In some embodiments, the building unit 1102 is configured to: The latent vectors corresponding to the last token in the feature representations of multiple decoding blocks are respectively determined as the latent state features.

[0270] In some embodiments, the building unit 1102 is configured to: Perform the following for each target decoding block in the at least one decoding block: Obtain the latent vector corresponding to the last token in the feature representation of the target decoding block, and use it as the first reference feature; Obtain a second reference feature of the feature representation of multiple consecutive decoding blocks corresponding to the target decoding block; the second reference feature includes pooling features and / or the latent vector of the last token in each feature representation; The hidden state features of the target decoding block are obtained by fusing the first reference feature and the second reference feature.

[0271] In some embodiments, the processing module 1002, such as Figure 11 As shown, it includes: The first processing unit 1103 is configured to input the hidden state features into a difficulty probe to obtain a difficulty index of the target sample; the difficulty index is used to measure the semantic complexity of the target sample; and, The second processing unit 1104 is used to input the hidden state features into the correctness probe to obtain a correctness index; the correctness index is used to measure the confidence level of the large model in correctly answering the target sample.

[0272] In some embodiments, the analysis module 1003, such as Figure 11 As shown, it includes: The first determining unit 1105 is used to determine the target sample as a noise sample when the difficulty index is lower than a first threshold and the correctness index is lower than a second threshold. The second determining unit 1106 is used to determine the target sample as a truly difficult sample when the difficulty index is higher than the third threshold and the correctness index is lower than the fourth threshold.

[0273] In some embodiments, a pseudo-feature index acquisition module 1004 is further included, for: The latent state features are processed based on the pseudo-feature probe to obtain the pseudo-feature index; the pseudo-feature index is used to measure the degree to which the decision made by the large model depends on the unexpected information of the target sample.

[0274] In some embodiments, the pseudo-feature index acquisition module 1004 is used for: The hidden state features are input into the pseudo-feature probe, and the hidden state features are processed using the linear mapping relationship of the pseudo-feature probe and the activation function layer to obtain the pseudo-feature index.

[0275] In some embodiments, the pseudo-feature index acquisition module 1004 is used for: A token for identifying unexpected information in the target sample; and, Perform the following operations based on the pseudo-feature probe: Based on interpretability techniques, the total contribution of the unexpected information token to the hidden state features is determined to obtain a first information value; and... Determine the total contribution of all tokens in the target sample to the hidden state features to obtain the second information value; The ratio of the first information value to the second information value is determined as the pseudo-feature index.

[0276] In some embodiments, the analysis module 1003, such as Figure 11 As shown, it also includes: The third determining unit 1107 is used to determine the first mismatch intensity of the difficulty index and the correctness index; the first mismatch intensity is used to measure the consistency between the difficulty index and the correctness index in describing the learning complexity of the target sample; The fourth determining unit 1108 is used to determine the second mismatch intensity based on the pseudo-feature index; the second mismatch intensity is used to measure the degree to which the large model decision-making process is affected by unexpected information; The fifth determining unit 1109 is used to determine the third mismatch intensity based on the fluctuation degree of the difficulty index corresponding to the hidden state features of multiple decoding blocks and the fluctuation degree of the correctness index corresponding to the hidden state features of multiple decoding blocks. The calculation unit 1110 is used to perform a weighted summation of the first mismatch intensity, the second mismatch intensity, and the third mismatch intensity to obtain the noise score of the target sample; the noise detection result includes the noise score.

[0277] In some embodiments, the analysis module 1003, such as Figure 11 As shown, it also includes: The first classification unit 1111 is used to classify the target sample based on the semantic consistency of the difficulty index, the correctness index and the pseudo-feature index corresponding to at least one latent state feature of the target sample, so as to obtain the sample category of the target sample; the sample category is divided based on the noisy case.

[0278] In some embodiments, the first classification unit 1111 is configured to: When the target sample corresponds to multiple hidden state features, the first statistical value of the difficulty index corresponding to each of the multiple hidden state features, the second statistical value of the correctness index corresponding to each of the multiple hidden state features, and the third statistical value of the pseudo feature index corresponding to at least one target hidden state among the multiple hidden state features are determined. The first statistical value, the second statistical value, and the third statistical value are matched with the level features of the true noise sample and the level features of the pseudo noise sample. Based on the matching results, the sample category of the target sample is determined.

[0279] In some embodiments, the first classification unit 1111 is specifically used for: The target sample is determined to be a true noise sample if it meets any of the following level characteristics of a true noise sample: The first characteristic of true noise includes: the first statistical value is less than or equal to the first target value, the second statistical value is less than or equal to the second target value, and the third statistical value is greater than or equal to the third target value; The second characteristic of true noise includes: the first statistical value is less than or equal to the first target value, the second statistical value is less than or equal to the second target value, and the third mismatch intensity is greater than or equal to the fourth target value.

[0280] In some embodiments, the first classification unit 1111 is specifically used for: The target sample is determined to be a pseudo-noise sample if it meets any of the following grade characteristics: The first characteristic of the pseudo noise includes: the first statistical value is less than or equal to the fifth target value, the second statistical value is less than or equal to the sixth target value, and the third statistical value is greater than or equal to the seventh target value; The second characteristic of pseudo-noise includes: the third mismatch intensity belongs to a preset quantile; the preset quantile is obtained based on statistics from multiple samples.

[0281] In some embodiments, the first classification unit 1111 is specifically used for: If neither the level features of the true noise sample nor the level features of the pseudo noise sample are matched, the target sample is determined to be a true hard sample.

[0282] In some embodiments, a review module 1005 is further included, for: If the target sample meets the conditions for manual review, explanatory information about the target sample is output; the explanatory information includes at least one of the following: Attention heatmap for the tokens of the target sample; The ranking of the contribution of the tokens of multiple characters in the target sample to the latent state features.

[0283] In some embodiments, the first processing unit 1103 is configured to: The hidden state features are input into the difficulty probe, and the hidden state features are processed using the linear mapping relationship of the difficulty probe to obtain the difficulty index.

[0284] In some embodiments, the first processing unit 1103 is configured to: Based on the difficulty probe, perform the following operations: Multiple degrees of uncertainty for the target sample are determined across multiple dimensions; the uncertainties across these multiple dimensions include at least one accidental uncertainty and / or at least one cognitive uncertainty; the accidental uncertainty indicates that the target sample has inherent noise, leading to uncertainty in the processing results of the large model on the target sample; the cognitive uncertainty indicates that the large model has knowledge blind spots due to a lack of training data; the multiple degrees of uncertainty include results determined based on the latent state features; Identify the noise contamination level of the target sample; Based on the noise pollution level and the multiple uncertainties, the difficulty index of the target sample is determined.

[0285] In some embodiments, the first processing unit 1103 is specifically used for: The weighted summation of the multiple degrees of uncertainty yields the index to be evaluated. The gating signal is determined based on the noise pollution level; the gating signal is used to filter out uncertainties caused by dirty data. Based on the gating signal processing, the evaluation index is obtained to determine the difficulty index of the target sample.

[0286] In some embodiments, the second processing unit 1104 is specifically used for: The hidden state features are input into the correctness probe, and the hidden state features are processed using the linear mapping relationship of the correctness probe and the activation function layer to obtain the correctness index.

[0287] Based on the same technical concept, this disclosure also proposes a model training device 1200 for sample detection, such as... Figure 12 As shown, it includes: The second extraction module 1201 is used to extract the latent feature representation of training samples based on a large model; Training module 1202 is used to train a probe based on the latent feature representation; the probe is used to perform noise detection on the target sample.

[0288] In some embodiments, the training module 1202, such as Figure 13 As shown, it includes: The second acquisition unit 1301 is used to acquire the weak supervision signal of the training sample when the probe includes a difficulty probe; the weak supervision signal is determined based on at least one of the following information: the heuristic reasoning complexity of the large model for the training sample, and the degree of divergence of the output results of multiple lightweight models for the training sample. Training unit 1302 is used to train the difficulty probe based on the weak supervision signal.

[0289] In some embodiments, the heuristic reasoning complexity includes: The large model processes the inference chain length of the training samples, and / or the number of logical operation steps of the training samples.

[0290] In some embodiments, the second acquisition unit 1301 is specifically used for: The weak supervision signal is obtained by weighted summation of the various pieces of information, based on the determination of multiple information sources.

[0291] In some embodiments, when the probe is a difficulty probe, the difficulty probe includes an uncertainty determination module, a contamination determination module, and a difficulty index determination module; The uncertainty determination module is used to determine multiple degrees of uncertainty of the training samples in multiple dimensions; The contamination determination module is used to identify the degree of noise contamination in the training samples; The difficulty index determination module is used to determine the difficulty index of the training samples based on the noise pollution level and the multiple uncertainties.

[0292] In some embodiments, the training module 1202, such as Figure 13 As shown, it includes: The third acquisition unit 1303 is used to acquire a first classification signal when the probe includes a correctness probe; the first classification signal is used to indicate whether the large model gives a correct answer to the training sample. The fourth acquisition unit 1304 is used to acquire the predicted correctness of the correctness probe for the latent feature representation; The second classification unit 1305 is used to perform a binary classification operation on the predicted correctness to obtain a first classification result; The first optimization unit 1306 is used to optimize the model parameters of the correctness probe based on the first classification result and the first classification signal.

[0293] In some embodiments, the training module 1202, such as Figure 13 As shown, it also includes: The fifth acquisition unit 1307 is used to acquire a second classification signal when the probe includes a pseudo-feature probe; the second classification signal is used to indicate whether the decision result of the large model depends on unexpected information to a greater than a preset degree; The sixth acquisition unit 1308 is used to acquire the estimated pseudo-feature value of the pseudo-feature probe for the latent feature representation; The third classification unit 1309 is used to perform a binary classification operation on the estimated false feature values ​​to obtain a second classification result; The second optimization unit 1310 is used to optimize the model parameters of the pseudo-feature probe based on the second classification result and the second classification signal.

[0294] In some embodiments, the fifth acquisition unit 1307 is configured to: The focus of multiple target attention heads of the large model on unexpected information in the training samples is statistically analyzed. The second classification signal is determined based on the focus.

[0295] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0296] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0297] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0298] Figure 14 A schematic block diagram of an example electronic device 1400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0299] like Figure 14As shown, device 1400 includes a computing unit 1401, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1402 or a computer program loaded from storage unit 1408 into random access memory (RAM) 1403. The RAM 1403 may also store various programs and data required for the operation of device 1400. The computing unit 1401, ROM 1402, and RAM 1403 are interconnected via bus 1404. Input / output (I / O) interface 1405 is also connected to bus 1404.

[0300] Multiple components in device 1400 are connected to I / O interface 1405, including: input unit 1406, such as a keyboard, mouse, etc.; output unit 1407, such as various types of displays, speakers, etc.; storage unit 1408, such as a disk, optical disk, etc.; and communication unit 1409, such as a network card, modem, wireless transceiver, etc. Communication unit 1409 allows device 1400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0301] The computing unit 1401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1401 performs the various methods and processes described above, such as large-model-based sample detection methods / model training methods for sample detection. For example, in some embodiments, the large-model-based sample detection methods / model training methods for sample detection can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 1408. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1400 via ROM 1402 and / or communication unit 1409. When the computer program is loaded into RAM 1403 and executed by the computing unit 1401, one or more steps of the large-model-based sample detection methods / model training methods for sample detection described above can be performed. Alternatively, in other embodiments, computing unit 1401 may be configured by any other suitable means (e.g., by means of firmware) to perform a large model-based sample detection method / model training method for sample detection.

[0302] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0303] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0304] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0305] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0306] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0307] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0308] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0309] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A sample detection method based on a large model, comprising: Extracting latent state features of target samples based on a large model; The latent state features are processed by multiple probes respectively, and an evaluation index is obtained for each probe for the target sample; wherein, the multiple probes output the evaluation index from different dimensions. Consistency analysis was performed on the obtained evaluation indicators to obtain the noise detection results of the target sample.

2. The method according to claim 1, wherein, The extraction of latent state features of target samples based on a large model includes: The target sample is input into the large model to obtain the feature representation output by at least one decoding block of the large model; The hidden state features are constructed based on at least one feature representation.

3. The method according to claim 2, wherein, The construction of the latent state features based on at least one feature representation includes: The latent vectors corresponding to the last token in the feature representations of multiple decoding blocks are respectively determined as the latent state features.

4. The method according to claim 2, wherein, The construction of the latent state features based on at least one feature representation includes: Perform the following for each target decoding block in the at least one decoding block: Obtain the latent vector corresponding to the last token in the feature representation of the target decoding block, and use it as the first reference feature; Obtain a second reference feature of the feature representation of multiple consecutive decoding blocks corresponding to the target decoding block; the second reference feature includes pooling features and / or the latent vector of the last token in each feature representation; The hidden state features of the target decoding block are obtained by fusing the first reference feature and the second reference feature.

5. The method according to any one of claims 1-4, wherein, The process involves processing the latent state features using multiple probes to obtain evaluation metrics for each probe targeting the target sample, including: The hidden state features are input into a difficulty probe to obtain a difficulty index for the target sample; the difficulty index is used to measure the semantic complexity of the target sample; and, The hidden state features are input into the correctness probe to obtain the correctness index; the correctness index is used to measure the confidence of the large model in correctly answering the target sample.

6. The method according to claim 5, wherein, The process of performing consistency analysis on the obtained multiple evaluation indicators to obtain the noise detection results of the target sample includes: If the difficulty index is lower than the first threshold and the correctness index is lower than the second threshold, the target sample is determined to be a noise sample. If the difficulty index is higher than the third threshold and the correctness index is lower than the fourth threshold, the target sample is determined to be a truly difficult sample.

7. The method according to claim 5, further comprising: The hidden state features are processed based on the pseudo-feature probe to obtain the pseudo-feature index; The pseudofeature index is used to measure the degree to which the decisions made by the large model depend on unexpected information from the target sample.

8. The method according to claim 7, wherein, The process of processing the latent state features based on the pseudo-feature probe to obtain pseudo-feature indices includes: The hidden state features are input into the pseudo-feature probe, and the hidden state features are processed using the linear mapping relationship of the pseudo-feature probe and the activation function layer to obtain the pseudo-feature index.

9. The method according to claim 7, wherein, The process of processing the latent state features based on the pseudo-feature probe to obtain pseudo-feature indices includes: A token for identifying unexpected information in the target sample; and, Perform the following operations based on the pseudo-feature probe: Based on interpretability techniques, the total contribution of the unexpected information token to the hidden state features is determined to obtain a first information value; and... Determine the total contribution of all tokens in the target sample to the hidden state features to obtain the second information value; The ratio of the first information value to the second information value is determined as the pseudo-feature index.

10. The method according to any one of claims 7-9, wherein, The process of performing consistency analysis on the obtained multiple evaluation indicators to obtain the noise detection results of the target sample includes: A first mismatch strength is determined between the difficulty metric and the correctness metric; the first mismatch strength is used to measure the consistency between the difficulty metric and the correctness metric in describing the learning complexity of the target sample; Based on the pseudo-feature index, a second mismatch intensity is determined; the second mismatch intensity is used to measure the degree to which the large model decision-making process is affected by unexpected information. The third mismatch intensity is determined based on the fluctuation of the difficulty index corresponding to the hidden state features of multiple decoding blocks and the fluctuation of the correctness index corresponding to the hidden state features of multiple decoding blocks. The first mismatch intensity, the second mismatch intensity, and the third mismatch intensity are weighted and summed to obtain the noise score of the target sample; the noise detection result includes the noise score.

11. The method according to any one of claims 7-9, wherein, The process of performing consistency analysis on the obtained multiple evaluation indicators to obtain the noise detection results of the target sample includes: Based on the semantic consistency of the difficulty index, the correctness index, and the pseudo-feature index corresponding to at least one latent state feature of the target sample, the target sample is classified to obtain the sample category of the target sample; the sample category is divided based on the noisy case.

12. The method according to claim 11, wherein, The target sample is classified based on the semantic consistency of the difficulty index, the correctness index, and the pseudo-feature index corresponding to at least one latent state feature of the target sample, resulting in the sample category of the target sample, including: When the target sample corresponds to multiple hidden state features, the first statistical value of the difficulty index corresponding to each of the multiple hidden state features, the second statistical value of the correctness index corresponding to each of the multiple hidden state features, and the third statistical value of the pseudo feature index corresponding to at least one target hidden state among the multiple hidden state features are determined. The first statistical value, the second statistical value, and the third statistical value are matched with the level features of the true noise sample and the level features of the pseudo noise sample. Based on the matching results, the sample category of the target sample is determined.

13. The method according to claim 12, wherein, Determining the sample category of the target sample based on the matching results includes: The target sample is determined to be a true noise sample if it meets any of the following level characteristics of a true noise sample: The first characteristic of true noise includes: the first statistical value is less than or equal to the first target value, the second statistical value is less than or equal to the second target value, and the third statistical value is greater than or equal to the third target value; The second characteristic of true noise includes: the first statistical value is less than or equal to the first target value, the second statistical value is less than or equal to the second target value, and the third mismatch intensity is greater than or equal to the fourth target value.

14. The method according to claim 12, wherein, Determining the sample category of the target sample based on the matching results includes: The target sample is determined to be a pseudo-noise sample if it meets any of the following grade characteristics: The first characteristic of the pseudo noise includes: the first statistical value is less than or equal to the fifth target value, the second statistical value is less than or equal to the sixth target value, and the third statistical value is greater than or equal to the seventh target value; The second characteristic of pseudo-noise includes: the third mismatch intensity belongs to a preset quantile; the preset quantile is obtained based on statistics from multiple samples.

15. The method according to claim 12, wherein, Determining the sample category of the target sample based on the matching results includes: If the level features of the real noise sample and the level features of the pseudo noise sample are not matched, the target sample is determined to be a true hard sample.

16. The method of claim 12, further comprising: If the target sample meets the conditions for manual review, output explanatory information about the target sample. The explanatory information includes at least one of the following: Attention heatmap for the tokens of the target sample; The ranking of the contribution of multiple characters of the target sample's token to the latent state features.

17. The method according to claim 5, wherein, The step of inputting the hidden state features into the difficulty probe to obtain the difficulty index of the target sample includes: The hidden state features are input into the difficulty probe, and the hidden state features are processed using the linear mapping relationship of the difficulty probe to obtain the difficulty index.

18. The method according to claim 5, wherein, The step of inputting the hidden state features into the difficulty probe to obtain the difficulty index of the target sample includes: Based on the difficulty probe, perform the following operations: Multiple degrees of uncertainty for the target sample are determined across multiple dimensions; the uncertainties across these multiple dimensions include at least one accidental uncertainty and / or at least one cognitive uncertainty; the accidental uncertainty indicates that the target sample has inherent noise, leading to uncertainty in the processing results of the large model on the target sample; the cognitive uncertainty indicates that the large model has knowledge blind spots due to a lack of training data; the multiple degrees of uncertainty include results determined based on the latent state features; Identify the noise contamination level of the target sample; Based on the noise pollution level and the multiple uncertainties, the difficulty index of the target sample is determined.

19. The method according to claim 18, wherein, The method for determining the difficulty index of the target sample based on the noise pollution level and the multiple uncertainties includes: The weighted summation of the multiple degrees of uncertainty yields the index to be evaluated. The gate signal is determined based on the noise pollution level; the gate signal is used to filter out uncertainties caused by dirty data. Based on the gating signal processing, the evaluation index is obtained to determine the difficulty index of the target sample.

20. The method according to claim 5, wherein, The step of inputting the hidden state features into the correctness probe to obtain the correctness index includes: The hidden state features are input into the correctness probe, and the hidden state features are processed using the linear mapping relationship of the correctness probe and the activation function layer to obtain the correctness index.

21. A model training method for sample detection, comprising: Extracting latent feature representations from training samples based on large models; The probe is trained based on the latent feature representation; The probe is used to detect noise in the target sample.

22. The method according to claim 21, wherein, The probe training based on the latent feature representation includes: When the probe includes a difficulty probe, a weak supervision signal is obtained from the training sample; the weak supervision signal is determined based on at least one of the following: the heuristic reasoning complexity of the large model for the training sample, and the degree of divergence in the output results of multiple lightweight models for the training sample. The difficulty probe is trained based on the weak supervision signal.

23. The method according to claim 22, wherein, The heuristic reasoning complexity includes: The large model processes the inference chain length of the training samples, and / or the number of logical operation steps of the training samples.

24. The method according to claim 22, wherein, The step of obtaining the weak supervision signal of the training samples includes: The weak supervision signal is obtained by weighted summation of the various pieces of information, based on the determination of multiple information sources.

25. The method according to claim 21, wherein, When the probe is a difficult probe, the difficult probe includes an uncertainty determination module, a contamination determination module, and a difficulty index determination module; The uncertainty determination module is used to determine multiple degrees of uncertainty of the training samples in multiple dimensions; The contamination determination module is used to identify the degree of noise contamination in the training samples; The difficulty index determination module is used to determine the difficulty index of the training samples based on the noise pollution level and the multiple uncertainties.

26. The method according to claim 21, wherein, The probe training based on the latent feature representation includes: When the probe includes a correctness probe, a first classification signal is acquired; the first classification signal is used to indicate whether the large model gives a correct answer to the training sample. Obtain the predicted correctness of the correctness probe for the latent feature representation; A binary classification operation is performed on the accuracy of the prediction to obtain the first classification result; Based on the first classification result and the first classification signal, optimize the model parameters of the correctness probe.

27. The method according to claim 21, wherein, The probe training based on the latent feature representation includes: In the case where the probe includes a pseudo-feature probe, a second classification signal is obtained; the second classification signal is used to indicate whether the decision result of the large model depends on unexpected information to a greater than a preset degree. Obtain the estimated pseudo-feature value of the pseudo-feature probe for the latent feature representation; A binary classification operation is performed on the estimated pseudo-feature values ​​to obtain a second classification result; Based on the second classification result and the second classification signal, the model parameters of the pseudo-feature probe are optimized.

28. The method according to claim 27, wherein, The acquisition of the second classification signal includes: The focus of multiple target attention heads of the large model on unexpected information in the training samples is statistically analyzed. The second classification signal is determined based on the focus.

29. A sample detection device based on a large model, comprising: The first extraction module is used to extract the latent state features of the target sample based on the large model; The processing module is used to process the latent state features based on multiple probes respectively, and obtain the evaluation index output by each probe for the target sample; wherein, the multiple probes output the evaluation index from different dimensions. The analysis module is used to perform consistency analysis on the obtained multiple evaluation indicators to obtain the noise detection results of the target sample.

30. A model training device for sample detection, comprising: The second extraction module is used to extract the latent feature representations of training samples based on the large model; The training module is used to train the probe based on the latent feature representation; The probe is used to detect noise in the target sample.

31. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-28.

32. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-28.

33. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-28.

Images