Instruction compliance assessment method and apparatus, electronic device, and storage medium

Abstract

The application provides an instruction following capability evaluation method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence. First, a task instruction and reference knowledge related to the task instruction are obtained; then the task instruction and the reference knowledge are spliced into a Prompt text, and a first large language model is called to perform reasoning based on the Prompt text and output tokens one by one; at the same time, when the first large language model outputs a token, the current activation value sequence of the intermediate layer output of the first large language model is extracted; the pre-trained evaluation model is used to determine a capability evaluation value based on the current activation value sequence, and the capability evaluation value reflects the instruction following capability of the first large language model. In the model generation process, the evaluation model is used to determine the capability evaluation value based on the activation value sequence of the first large language model each time a token is output, so that efficient and accurate instruction following capability evaluation can be realized in the model generation process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method, apparatus, electronic device, and storage medium for evaluating instruction compliance capability. Background Technology

[0002] With the rapid development of artificial intelligence technology, Large Language Models (LLMs) have demonstrated powerful generation and understanding capabilities in the field of natural language processing and have been widely applied in scenarios such as intelligent question answering, content creation, and dialogue systems. Among them, Retrieval-Augmented Generation (RAG), as a generative architecture that integrates external knowledge bases, effectively alleviates the limitations of large models in terms of knowledge memory accuracy and factual consistency by introducing relevant search results before generation, and significantly improves the relevance and reliability of generated content.

[0003] In RAG applications, large models not only need strong language generation capabilities but also a high degree of instruction following ability. This means they must accurately understand and strictly execute specific task instructions given by the user, such as fill-in-the-blank, continuation, paraphrasing, and summarizing. This ability directly determines whether the system output meets the user's expectations and is crucial in practical applications.

[0004] However, assessing the instruction compliance capabilities of large models still faces numerous challenges. Current technologies primarily rely on two methods: manual evaluation, where annotators score the quality or assess compliance of the generated results; and the use of traditional automated metrics, such as accuracy, recall, BLEU, ROUGE, and other n-gram matching-based evaluation methods. These methods have significant drawbacks:

[0005] First, manual evaluation is costly and time-consuming, making it difficult to meet the rapid iteration needs of large-scale data. It is also susceptible to subjective interference, resulting in poor consistency of evaluation results and failing to support efficient and objective performance monitoring.

[0006] Secondly, traditional automated evaluation processes generally lag behind the model inference process, meaning that offline evaluation is performed after generation, making it impossible to monitor the model's behavior in real time during generation. This "post-hoc" evaluation mechanism results in severe feedback delays, making it difficult to detect and correct erroneous outputs in a timely manner, thus prolonging the iteration cycle of model tuning and deployment.

[0007] In summary, existing technologies have significant shortcomings in terms of evaluation efficiency and real-time performance, making it difficult to meet the actual needs of RAG scenarios for efficient, accurate, and dynamic evaluation of large model instruction compliance capabilities. Summary of the Invention

[0008] The purpose of this invention is to provide a method, apparatus, electronic device, and storage medium for evaluating instruction compliance capability, so as to improve the problems existing in the prior art.

[0009] The embodiments of the present invention can be implemented as follows:

[0010] In a first aspect, the present invention provides a method for evaluating instruction compliance capability, comprising:

[0011] Obtain the task instructions and related reference knowledge;

[0012] The task instructions and the reference knowledge are concatenated into a Prompt text, and the first major language model is invoked to perform inference based on the Prompt text and output tokens one by one.

[0013] When the first large language model outputs a token, extract the current activation value sequence output by the intermediate layer of the first large language model;

[0014] A capability assessment value is determined based on the current activation value sequence using a pre-trained evaluation model; the capability assessment value reflects the instruction-following capability of the first large language model.

[0015] Secondly, the present invention provides an apparatus for evaluating instruction compliance capability, comprising:

[0016] The acquisition module is used to obtain task instructions and related reference knowledge.

[0017] The reasoning module is used to concatenate the task instructions and the reference knowledge into a Prompt text, and call the first language model to perform reasoning based on the Prompt text and output tokens one by one;

[0018] The evaluation module is used to extract the current activation value sequence of the intermediate layer output of the first large language model when the first large language model outputs a token;

[0019] The evaluation module is also used to determine a capability evaluation value based on the current activation value sequence using a pre-trained evaluation model; the capability evaluation value reflects the instruction-following capability of the first large language model.

[0020] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory stores a software program, and when the electronic device is running, the processor executes the software program to implement the instruction compliance assessment method described in the first aspect above.

[0021] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the instruction compliance assessment method described in the first aspect.

[0022] Compared with existing technologies, embodiments of the present invention provide a method, apparatus, electronic device, and storage medium for evaluating instruction compliance capability. First, task instructions and related reference knowledge are obtained. Then, the task instructions and reference knowledge are concatenated into a prompt text, and a first large language model is invoked to perform inference based on the prompt text and output tokens one by one. Simultaneously, when the first large language model outputs a token, the current activation value sequence output by the intermediate layer of the first large language model is extracted. A pre-trained evaluation model is used to determine a capability evaluation value based on the current activation value sequence, which reflects the instruction compliance capability of the first large language model. In the model generation process, the present invention utilizes an evaluation model to determine the capability evaluation value based on the activation value sequence of each token output by the first large language model, thereby achieving efficient and accurate instruction compliance capability evaluation during the model generation process. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is one of the flowcharts illustrating a method for evaluating instruction compliance capability provided in an embodiment of the present invention.

[0025] Figure 2 This is a second schematic flowchart of a method for evaluating instruction compliance capability provided in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the structure of an instruction compliance assessment device provided in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0029] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0030] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0031] Furthermore, the terms "first" and "second" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0032] It should be noted that, where there is no conflict, the features in the embodiments of the present invention can be combined with each other.

[0033] It should be understood that in current applications of intelligent systems based on large language models, the evaluation of whether the model output accurately follows user instructions usually relies on post-analysis of the final generated results. Such methods struggle to identify deviations in a timely manner during the generation process, leading to feedback delays and optimization lags. It should be noted that with the development of complex application scenarios such as Retrieval-Augmented Generation (RAG), models need to incorporate external reference knowledge to complete tasks. Their instruction compliance capability directly impacts content reliability and user experience; therefore, a technical mechanism that can dynamically evaluate and quantify compliance during the inference process is urgently needed.

[0034] Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for evaluating instruction compliance capability according to an embodiment of the present invention. The subject executing this method can be, but is not limited to, electronic devices such as smartphones, personal laptops, personal computers, and servers. Figure 1 The evaluation method includes the following steps S101 to S104.

[0035] S101. Obtain the task instructions and related reference information.

[0036] The task instructions are descriptive text for generating the task, while the reference knowledge can include multiple knowledge fragments retrieved from an external knowledge base based on the task instructions, providing contextual support for the model.

[0037] S102. Concatenate the task instructions and reference knowledge into a Prompt text, and call the first major language model to perform inference based on the Prompt text and output each token one by one.

[0038] In this embodiment, the Prompt text is passed as input to the first language model, triggering its autoregressive inference process. During this process, the first language model decodes and outputs response content token by token based on its internal parameter structure and contextual understanding mechanism. Each output token is the result of probability distribution sampling by the model in the current context state, reflecting its understanding of the instruction intent.

[0039] S103. When the first language model outputs a token, extract the current activation value sequence of the intermediate layer output by the first language model.

[0040] Simultaneously, at each output of a token by the primary language model, the current activation value sequence output by the intermediate layers of that model is extracted. These activation value sequences can originate from the hidden layer node states in the Transformer architecture, containing the model's deep semantic representation and attention distribution features when processing the current input, effectively reflecting the stability and consistency of its internal decision-making path. Because this extraction occurs in the middle stage of the output chain, it possesses early perception potential.

[0041] S104. Determine the capability assessment value based on the current activation value sequence using a pre-trained evaluation model.

[0042] In this embodiment, a pre-trained evaluation model receives the current activation value sequence as input data and performs inference operations based on the mapping relationships learned internally. This evaluation model, trained on a large number of historical samples, can identify the correlation between different activation patterns and instruction compliance quality. Through this process, a numerical value representing the rationality of the current generated behavior, i.e., the capability evaluation value, is output.

[0043] This capability assessment score is not an overall score for the final answer, but rather a quantitative expression of the model's tendency to follow local instructions when generating a particular token. It can be used to determine whether the model is evolving in the expected direction. Therefore, the capability assessment score can reflect the instruction-following ability of the largest language model.

[0044] The instruction compliance assessment method provided in this invention obtains task instructions and related reference knowledge. Then, the task instructions and reference knowledge are concatenated into a prompt text, and a first language model is invoked to perform inference based on the prompt text and output tokens one by one. Simultaneously, when the first language model outputs a token, the current activation value sequence output by the intermediate layer of the first language model is extracted. A pre-trained assessment model is then used to determine an instruction compliance assessment value based on the current activation value sequence. This assessment value reflects the instruction compliance capability of the first language model. In the model generation process, this invention utilizes an assessment model to determine the capability assessment value based on the activation value sequence of each token output by the first language model, thereby achieving efficient and accurate instruction compliance capability assessment during model generation.

[0045] It should be understood that in the current mainstream Large Language Model (LLM) architecture, the text generation process is implemented using an autoregressive mechanism. The core of this mechanism is that the large model progressively predicts the next most likely token based on the existing context and appends it to the current sequence as new context for subsequent generation. This process is repeated until a preset termination condition is met, such as reaching the maximum output length or generating an end-of-line character.

[0046] Specifically, after the question text is input into the large model, it is first encoded, converting it into a series of discrete token sequences. Then, in the inference phase, the model semantically models the current context based on its internal parameters and attention mechanism, outputting the conditional probability distribution of each token in the vocabulary. Based on this, a token is selected as the output for the current step using a sampling strategy (such as greedy search, temperature sampling, top-k, or top-p sampling). The question text is then concatenated with all the output tokens and input into the large model to predict the next token, until a preset termination condition is met, yielding the response text composed of all output tokens.

[0047] Therefore, evaluating the compliance capability of a single token output by a large model often fails to guarantee the stability of output quality, especially when the model exhibits semantic drift, information omission, or logical breaks under complex instructions or edge cases, lacking an effective mechanism for self-correction and long-term optimization. To improve the robustness and adaptability of the system, this invention can also be based on a feedback and optimization mechanism of process evaluation results.

[0048] In one possible implementation, real-time corrections can be made during model generation to obtain an accurate response. This real-time correction process is as follows:

[0049] (1) When the first language model outputs the i-th token based on the Prompt text, extract the current activation value sequence of the intermediate layer output of the first language model; and input the current activation value sequence into the evaluation model to obtain the capability evaluation value corresponding to the i-th token;

[0050] (2) If the capability evaluation value corresponding to the i-th token is lower than the set evaluation threshold (e.g., 75% or 80%), it indicates that the currently generated token is insufficient in terms of semantic coherence, information accuracy or instruction consistency. Therefore, the i-th token is rejected and the first language model is called again to generate a new i-th token under the same context conditions.

[0051] (3) If the capability evaluation value corresponding to the i-th token is not lower than the set evaluation threshold, it means that the i-th token has an acceptable quality level and can be directly adopted. Then, the updated context formed by concatenating the Prompt text with all adopted tokens is used as input to continue driving the first language model to generate the next token.

[0052] Through the above real-time correction process, dynamic quality screening of each token is achieved during the generation process, ensuring that the final output answer text is closer to the instruction intent and has higher content reliability.

[0053] In another possible implementation, after the model has finished generating the text, the quality of the answer text can be determined by combining the evaluation results of each token, in order to determine whether to regenerate it. Therefore, in Figure 1 Based on this, please refer to Figure 2 The instruction, following the capability assessment method, may also include the following steps S105-S107.

[0054] S105. Obtain the answer text corresponding to the Prompt text output by the first large language model, and obtain the ability evaluation value corresponding to each token in the answer text;

[0055] S106. Calculate the average of all ability assessment values ​​to obtain the accuracy of the answer text;

[0056] S107. If the accuracy is lower than the set accuracy threshold, the first language model is called again to perform inference based on the Prompt text until the target answer text with an accuracy not lower than the set accuracy threshold is obtained.

[0057] It is understandable that in the process of generating answer text based on the prompt text in the first language model, by executing the above steps S103~S104, the capability evaluation value corresponding to each output token in the answer text can be obtained. These capability evaluation values ​​reflect the degree of compliance of the model with the instructions at the time of generating each token, constituting a set of process quality indicators. On this basis, by aggregating all capability evaluation values ​​and calculating their arithmetic mean, a quantitative parameter representing the overall performance of the entire answer text is obtained, namely the accuracy of the answer text. This accuracy, as a comprehensive evaluation index, can effectively represent the overall compliance level of the model under the current task, avoiding the influence of a single link's abnormality on the global judgment.

[0058] When the accuracy of the answer text falls below a preset accuracy threshold (e.g., 80%), it indicates that the generated answer text fails to meet the expected quality standards and may contain issues such as semantic incoherence, missing information, or logical breaks. In this case, the system automatically triggers a regeneration process, re-invoking the primary language model with the same prompt text as input and re-executing the inference process. This process can be performed using the same or adjusted generation strategy until a target answer text with an accuracy not lower than the set threshold is obtained. Through this feedback control mechanism, the system can autonomously complete quality screening and result optimization without manual intervention, significantly improving the reliability and consistency of the output.

[0059] Furthermore, to fully utilize the difference information between positive and negative samples generated during the generation process, the system structures the results before and after each regeneration into data resources that can be used for model optimization. That is, after step S106, the system may further include:

[0060] S107. Using the Prompt text as the input sample, the target answer text as the positive response, and the first obtained answer text as the negative response, construct a comparison sample and store the comparison sample in the sample library.

[0061] In practice, the original Prompt text is used as the input sample. The first generated answer text is marked as a negative response, and the target answer text is marked as a positive response, thus constructing a comparative sample with clear distinctions in quality. This comparative sample is persistently stored in a sample library, continuously accumulating high-quality discrimination data from real-world reasoning scenarios. This type of data not only contains diversity at the linguistic expression level but also implicitly reveals the behavioral differences of the model under different generation paths, possessing strong training value.

[0062] Thus, when the number of samples in the sample library reaches a predetermined amount, a model optimization can be performed: based on all the comparison samples in the sample library, the primary language model is post-trained using contrastive learning. This contrastive learning approach guides the model to narrow the semantic association between positive responses and input samples, while simultaneously reducing the influence of negative responses, thereby enhancing its ability to learn high-quality response patterns. This optimization method, driven by real-world interaction data, offers greater realistic coverage and generalization potential compared to manually constructed training sets, helping to improve the initial output quality of the model in similar future tasks.

[0063] In one possible implementation, an evaluation model adapted to various task types can be trained to handle the evaluation of different types of task instructions. In this case, the training process of the evaluation model may include the following steps S201~S202;

[0064] S201. Construct a discriminant dataset; the discriminant dataset includes several activation value sequence samples and their respective actual matching degrees;

[0065] S202. Using activation value sequence samples as input features and actual matching degree as supervision signal, train the pre-built machine learning model using the discriminant dataset to obtain the evaluation model.

[0066] The first step in the training process is to construct a discriminative dataset, which consists of several activation value sequence samples and their corresponding actual matching scores. Each activation value sequence sample originates from the hidden state output by the intermediate layer when the first language model outputs a token during the generation task, reflecting the model's cognitive representation at a certain inference moment. The corresponding actual matching score is used to quantify the degree of conformity between the currently generated content and the standard expectation, serving as a supervisory signal to measure the quality of instruction compliance.

[0067] Building upon this, activation value sequence samples are used as input features and fed into a pre-built machine learning model. This model can employ Gradient Boosting Decision Tree (GBDT), Multilayer Perceptron (MLP), or other structures suitable for high-dimensional vector regression or classification tasks, possessing the ability to learn mapping relationships from complex nonlinear features. Using actual matching degree as the supervision objective, the machine learning model's parameters are optimized by minimizing prediction error, enabling it to gradually grasp the correlation between different activation modes and instruction compliance levels.

[0068] After thorough training, the machine learning model is transformed into an evaluation model, capable of receiving activation value sequences in new scenarios and outputting corresponding capability evaluation values. Since its training data originates from intermediate states and corresponding quality labels in real-world reasoning processes, it can make reasonable predictions about the model's current compliance tendency based solely on partial generated information, without relying on the complete output. This mechanism allows the evaluation process to be embedded into the generation flow, enabling dynamic tracking and early warning of instruction compliance capabilities.

[0069] In an optional implementation, the process of implementing step S201 may include the following steps S2011 to S2014.

[0070] S2011. Obtain the question-and-answer dataset. Each sample data in the question-and-answer dataset includes historical task instructions, historical reference knowledge, and standard answer text.

[0071] Specifically, the sample data in the question-answering dataset comes from real or artificially constructed question-answering scenarios, covering a variety of task types (such as fill-in-the-blank tasks, continuation tasks, restatement tasks, rewriting tasks, sentence construction tasks, etc.), and has good coverage and diversity. Among them, historical task instructions are used to clearly require the execution of task operations and task content, historical reference knowledge provides contextual basis as external supporting information, and standard answer text represents the ideal output result that meets expectations.

[0072] S2012. For any sample data, input the historical task instructions and historical reference knowledge in the sample data into the first language model to obtain the predicted answer text.

[0073] For any given sample data, the historical task instructions and historical reference knowledge are concatenated and input into the first language model, triggering its autoregressive inference process. During this process, the first language model generates predicted answer text for each token, and this predicted answer text consists of N consecutively output predicted tokens.

[0074] S2013. When the first language model outputs the nth predicted token, extract the activation value sequence output by the intermediate layer of the first language model as the activation value sequence sample, and determine the actual matching degree corresponding to the activation value sequence sample based on the output token sequence and the standard answer text.

[0075] When the first large language model outputs the nth predicted token ( ):

[0076] The system synchronously extracts the activation value sequence output by the intermediate layer of the first major language model at this time and uses it as an activation value sequence sample. This activation value sequence sample records the internal state of the model at the moment of generating the current token, including its understanding of contextual semantics, instruction intent, and generation path;

[0077] Simultaneously, based on the output token sequence comprised of the n predicted tokens generated up to the current moment, and the pre-defined standard answer text, the degree of matching between the two is calculated, yielding a quantitative metric: the actual matching degree corresponding to the activation value sequence sample. This actual matching degree reflects the closeness between the model's performance in the partial generation stage and the ideal result.

[0078] Thus, based on each sample data, N activation value sequence samples and their corresponding actual matching degree can be obtained.

[0079] In one alternative example, vector similarity can be used to determine the actual match. This involves converting the output token sequence and the standard answer text into token sequence vectors and standard answer vectors, respectively; then calculating the similarity (e.g., cosine similarity) between the token sequence vectors and the standard answer vectors to obtain the actual match for a given activation value sequence sample. Alternatively, the edit distance between the n tokens in the output token sequence and the standard answer text can be directly calculated to obtain the actual match.

[0080] In another alternative example, the actual matching degree can be obtained by scoring through a language model: that is, the output token sequence and the standard answer text are input into a second-largest language model for comparison to obtain the actual matching degree corresponding to the activation value sequence sample.

[0081] S2014. Iterate through each sample data to obtain the discrimination dataset.

[0082] By traversing all sample data in the question-answering dataset and repeating the above processing steps, a structured discriminative dataset can be formed by integrating all activation value sequence samples collected during the generation process and their corresponding actual matching degrees.

[0083] It should be understood that in scenarios where large language models handle diverse task instructions, different types of instructions place significantly different demands on the model's capabilities. For example, fill-in-the-blank tasks focus more on information completeness and precise matching, continuation tasks emphasize contextual coherence and logical extension, while paraphrasing tasks emphasize semantic preservation and expressive diversity. If a single evaluation model is used to uniformly evaluate all task types, it is difficult to take into account the characteristics of each capability dimension, which may lead to coarse evaluation results and insufficient sensitivity.

[0084] Therefore, a specific evaluation model can be trained for each required task type for later use. Based on this, the implementation of step S104 above can include:

[0085] S1041. Based on the one-to-one mapping relationship between multiple pre-trained evaluation models and multiple task types, determine the target evaluation model that matches the task type to which the task instruction belongs from the multiple evaluation models.

[0086] S1042. Input the current activation value sequence into the target evaluation model to obtain the capability evaluation value.

[0087] Specifically, in determining ability assessment values ​​using pre-trained assessment models, a general model is not used to process all inputs. Instead, multiple specialized assessment models are pre-built, and a one-to-one mapping relationship is established between these models and various task types. Each assessment model is trained for a specific task type, and the feature patterns and quality criteria it learns focus on the performance patterns of the corresponding ability dimension. For example, one assessment model is specifically used to judge the accuracy of information completion when performing a fill-in-the-blank task, while another model focuses on evaluating the rationality of semantic continuation during the continuation writing process.

[0088] When a current activation value sequence is extracted, the system first searches and matches it within a pre-defined mapping relationship based on the task type to which the task instruction belongs. A unique model, the target evaluation model, is then selected from multiple evaluation models, and the current activation value sequence is input into this model. Based on its training experience in a specific task domain, this target evaluation model analyzes the generation behavior reflected by the activation value sequence and outputs a capability evaluation value that is appropriate for the task context. This capability evaluation value more accurately reflects the instruction compliance tendency of the leading language model in the current task context.

[0089] Thus, by introducing a task-type-driven model selection mechanism, fine-grained allocation of evaluation resources is achieved. Compared to using a uniform model for global evaluation, this method effectively avoids evaluation bias caused by task differences, enabling the system to maintain stable and sensitive quality judgment capabilities when faced with complex and diverse instruction inputs.

[0090] For example, when dealing with fill-in-the-blank tasks, continuation tasks, and restatement tasks, multiple pre-trained evaluation models include a fill-in-the-blank ability discrimination model, a continuation ability discrimination model, and a restatement ability discrimination model.

[0091] Optionally, the training process of the fill-in-the-blank ability discrimination model may include the following steps S301 to S305.

[0092] S301. Obtain the preset fill-in-the-blank instructions and fill-in-the-blank dataset. The fill-in-the-blank dataset includes multiple question stems to be filled in, their corresponding fill-in-the-blank reference knowledge, and standard fill-in-the-blank results.

[0093] The fill-in-the-blank dataset includes multiple fill-in-the-blank samples, each containing the question stem (the structure of the question to be filled in), its corresponding fill-in-the-blank reference knowledge, and standard fill-in-the-blank results. These fill-in-the-blank samples cover various domains and sentence types, ensuring the training data has good representativeness and generalization potential. Pre-set fill-in-the-blank instructions guide the first language model to clarify the task objective, such as "Please complete the blanks in the sentence based on the following information."

[0094] S302. For any blank to be filled in, the first language model is invoked to reason based on the preset blank filling instructions, the blank to be filled in and its corresponding blank filling reference knowledge, and each token is output to obtain the predicted blank filling result.

[0095] S303. When the first language model outputs the i-th token in the predicted fill-in-the-blank result, extract the intermediate layer output of the first language model to obtain the first activation value sequence sample, and determine the similarity between the first i tokens in the predicted fill-in-the-blank result and the standard fill-in-the-blank result to obtain the first matching degree corresponding to the first activation value sequence sample.

[0096] Where i takes the value of M1 represents the total number of tokens predicted for the fill-in-the-blank results. The first matching degree can be the text edit distance, or the cosine distance, edit distance, or semantic alignment score in the embedding space.

[0097] S304. After traversing each blank in the question stem, take each first activation value sequence sample and its corresponding first matching degree as a blank filling discrimination sample.

[0098] By performing steps S302 to S304 for each blank in the fill-in-the-blank dataset, several blank discrimination samples can be obtained.

[0099] S305. Using the first activation value sequence sample as the input feature and the first matching degree as the supervision signal, the pre-built machine learning model is trained using all the blank-filling discrimination samples to obtain the blank-filling ability discrimination model.

[0100] Subsequently, using the first activation value sequence samples as input features and the first matching degree as a supervision signal, a pre-built machine learning model (such as LightGBM, XGBoost, or multilayer perceptron) is trained in a supervised manner. By optimizing the model parameters, it can accurately predict the corresponding matching level from the input features, ultimately obtaining a fill-in-the-blank ability discrimination model with specific evaluation capabilities.

[0101] Optionally, the training process of the continuation ability discrimination model may include the following steps S401~S405.

[0102] S401. Obtain preset continuation instructions and continuation dataset. The continuation dataset includes multiple text fragments to be continued, their corresponding continuation reference knowledge, and standard continuation results.

[0103] The continuation-writing dataset includes multiple continuation-writing samples, which contain continuation-writing reference knowledge corresponding to the text fragments to be continued, as well as standard continuation-writing results. These continuation-writing samples cover various text types, including narrative, explanatory, and argumentative texts, ensuring that the training data has sufficient linguistic diversity and task representativeness. Pre-defined continuation-writing instructions are used to explicitly guide the first language model to perform "continuing the current content" operations, such as "Please continue to complete the following text" or "Write the following paragraph based on the above information."

[0104] S402. For any text segment to be continued, call the first language model to perform reasoning based on the preset continuation instructions, the text segment to be continued and its corresponding continuation reference knowledge, and output tokens one by one to obtain the predicted continuation result.

[0105] S403. When the j-th token in the predicted continuation result is output by the first large language model, the intermediate layer output of the first large language model is extracted to obtain the second activation value sequence sample, and the similarity between the first j tokens in the predicted continuation result and the standard continuation result is determined to obtain the second matching degree corresponding to the second activation value sequence sample.

[0106] Here, j takes the value of M2 represents the total number of tokens for the predicted continuation result.

[0107] S404. After traversing each text segment to be continued, take each second activation value sequence sample and its corresponding second matching degree as a continuation discrimination sample.

[0108] By performing steps S402 to S404 for each text segment to be continued in the continuation dataset, several continuation discrimination samples can be obtained.

[0109] S405. Using the second activation value sequence samples as input features and the second matching degree as supervision signal, train the pre-built machine learning model using all continuation discrimination samples to obtain the continuation ability discrimination model.

[0110] It should be noted that the training process of the continuation ability discrimination model is similar to that of the fill-in-the-blank ability discrimination model mentioned above, and some details will not be repeated here.

[0111] Optionally, the training process of the restatement ability discrimination model may include the following steps S501 to S505.

[0112] S501. Obtain the preset restatement instructions and restatement dataset. The restatement dataset includes multiple texts to be restated and their corresponding restatement reference knowledge and standard restatement results.

[0113] The restatement dataset includes multiple restatement samples, each consisting of the text to be restated, its corresponding restatement reference knowledge, and a standard restatement result. The text to be restated represents the text content that needs to be transformed in expression while preserving semantics. The restatement reference knowledge provides contextual support or domain background, and the standard restatement result reflects diverse expressions that meet expectations. Preset restatement instructions are used to guide the first language model to clarify the task objective, such as "Please restate the following sentence in different ways" or "Please rewrite the following text without changing the original meaning."

[0114] S502. For any text to be restated, the first language model is invoked to perform reasoning based on the preset restatement instructions, the text to be restated and its corresponding restatement reference knowledge, and each token is output to obtain the predicted restatement result.

[0115] S503. When the first language model outputs the k-th token in the predicted restatement result, extract the intermediate layer output of the first language model to obtain the third activation value sequence sample, and determine the similarity between the first k tokens in the predicted restatement result and the standard restatement result to obtain the third matching degree corresponding to the third activation value sequence sample.

[0116] Here, k takes the value of M3 represents the total number of tokens in the predicted restatement result.

[0117] S504. After traversing each text to be restated, take each third activation value sequence sample and its corresponding third matching degree as a restatement discrimination sample.

[0118] By performing steps S502 to S504 for each text to be restated in the restatement dataset, several restatement discrimination samples can be obtained.

[0119] S505. Using the third activation value sequence samples as input features and the third matching degree as supervision signal, the pre-built machine learning model is trained using all restatement discrimination samples to obtain the restatement ability discrimination model.

[0120] It should be noted that the training process of the restatement ability discrimination model is similar in logic to the training process of the fill-in-the-blank ability discrimination model mentioned above, and will not be repeated here.

[0121] This invention integrates the instruction compliance capability assessment method with the RAG workflow, enabling real-time feedback of assessment results during model generation and avoiding the time delays and resource waste of traditional post-assessment methods. Through real-time monitoring and automatic regeneration mechanisms, relevant instruction capability data is automatically collected for subsequent model optimization.

[0122] It should be noted that the execution order of each step in the above method embodiments is not limited to what is shown in the accompanying drawings and step numbers, and the execution order of each step shall be subject to the actual application situation.

[0123] In order to perform the corresponding steps in the above method embodiments and various possible implementations, an implementation of an instruction compliance assessment device is given below.

[0124] Please see Figure 3 , Figure 3 A schematic diagram of the structure of an instruction compliance assessment device 200 provided in an embodiment of the present invention is shown. The instruction compliance assessment device 200 includes: an acquisition module 210, a reasoning module 220, and an assessment module 230.

[0125] The acquisition module 210 is used to acquire task instructions and related reference knowledge.

[0126] The reasoning module 220 is used to concatenate task instructions and reference knowledge into a Prompt text, and call the first language model to perform reasoning based on the Prompt text and output tokens one by one;

[0127] Evaluation module 230 is used to extract the current activation value sequence of the intermediate layer output of the first language model when the first language model outputs a token;

[0128] Evaluation module 230 is also used to determine a capability evaluation value based on the current sequence of activation values ​​using a pre-trained evaluation model; the capability evaluation value reflects the instruction-following capability of the first language model.

[0129] Optionally, the capability assessment device 200 may further include a hybrid training module for training an assessment model by: constructing a discriminant dataset; the discriminant dataset including several activation value sequence samples and their respective actual matching degrees; using the activation value sequence samples as input features and the actual matching degrees as supervision signals, training a pre-constructed machine learning model using the discriminant dataset to obtain the assessment model.

[0130] Optionally, the hybrid training module can be used to: acquire a question-and-answer dataset, where each sample data includes historical task instructions, historical reference knowledge, and standard answer text; for any sample data, input the historical task instructions and historical reference knowledge from that sample data into the first language model to obtain predicted answer text; the predicted answer text includes N predicted tokens; when the first language model outputs the nth predicted token, extract the activation value sequence output by the intermediate layer of the first language model as an activation value sequence sample, and determine the actual matching degree corresponding to the activation value sequence sample based on the output token sequence and the standard answer text; wherein, The output token sequence includes the n predicted tokens output by the first large language model; iterate through each sample data to obtain the discriminant dataset.

[0131] Optionally, the hybrid training module can be used to: convert the output token sequence and the standard answer text into token sequence vectors and standard answer vectors, respectively; calculate the similarity between the token sequence vector and the standard answer vector to obtain the actual matching degree corresponding to the activation value sequence sample; or, input the output token sequence and the standard answer text into the second language model for comparison to obtain the actual matching degree corresponding to the activation value sequence sample.

[0132] Optionally, the capability assessment device 200 may further include a response determination module, configured to: obtain the answer text corresponding to the Prompt text output by the first large language model, and obtain the capability assessment value corresponding to each token in the answer text; calculate the average of all capability assessment values ​​to obtain the accuracy of the answer text; if the accuracy is lower than a set accuracy threshold, then re-call the first large language model to perform reasoning based on the Prompt text until a target answer text with an accuracy not lower than the set accuracy threshold is obtained.

[0133] Optionally, the capability assessment device 200 may also include a model optimization module for: constructing a comparison sample using the Prompt text as input, the target answer text as the positive response, and the first obtained answer text as the negative response, and storing the comparison sample in a sample library; when the number of samples in the sample library reaches a set number, then post-training the first language model using a contrastive learning approach based on all the comparison samples in the sample library.

[0134] Optionally, the evaluation module 230 can be used to: determine the target evaluation model that matches the task type to which the task instruction belongs from the multiple evaluation models based on the one-to-one mapping relationship between multiple pre-trained evaluation models and multiple task types; input the current activation value sequence into the target evaluation model to obtain the capability evaluation value.

[0135] Optionally, when the multiple evaluation models include a fill-in-the-blank ability discrimination model, the evaluation device 200 for instruction compliance ability may further include a first training module for training the fill-in-the-blank ability discrimination model in the following manner: acquiring a preset fill-in-the-blank instruction and a fill-in-the-blank dataset, wherein the fill-in-the-blank dataset includes multiple fill-in-the-blank stems and their corresponding fill-in-the-blank reference knowledge and standard fill-in-the-blank results; for any fill-in-the-blank stem, calling a first language model to perform reasoning based on the preset fill-in-the-blank instruction, the fill-in-the-blank stem and its corresponding fill-in-the-blank reference knowledge and outputting tokens one by one to obtain a predicted fill-in-the-blank result; and outputting the predicted fill-in-the-blank result from the first language model. When encountering the i-th token, extract the intermediate layer output of the first large language model to obtain the first activation value sequence sample, and determine the similarity between the first i tokens in the predicted fill-in-the-blank result and the standard fill-in-the-blank result to obtain the first matching degree corresponding to the first activation value sequence sample; after traversing each question stem to be filled in, take each obtained first activation value sequence sample and its corresponding first matching degree as a fill-in-the-blank discrimination sample; use the first activation value sequence sample as the input feature and the first matching degree as the supervision signal, and use all fill-in-the-blank discrimination samples to train the pre-built machine learning model to obtain the fill-in-the-blank ability discrimination model.

[0136] Optionally, when the multiple evaluation models include a continuation ability discrimination model, the instruction compliance evaluation device 200 may further include a second training module for training the continuation ability discrimination model in the following manner: acquiring a preset continuation instruction and a continuation dataset, the continuation dataset including multiple text fragments to be continued and their corresponding continuation reference knowledge and standard continuation results; for any text fragment to be continued, calling the first language model to perform inference based on the preset continuation instruction, the text fragment to be continued and its corresponding continuation reference knowledge and outputting tokens one by one to obtain a predicted continuation result; outputting the predicted continuation result from the first language model. When the j-th token in the result is obtained, the intermediate layer output of the first large language model is extracted to obtain the second activation value sequence sample. The similarity between the first j tokens in the predicted continuation result and the standard continuation result is determined to obtain the second matching degree corresponding to the second activation value sequence sample. After traversing each text segment to be continued, each obtained second activation value sequence sample and its corresponding second matching degree are used as a continuation discrimination sample. Using the second activation value sequence sample as the input feature and the second matching degree as the supervision signal, the pre-built machine learning model is trained using all continuation discrimination samples to obtain the continuation ability discrimination model.

[0137] Optionally, when the multiple evaluation models include a restatement ability discrimination model, the instruction compliance ability evaluation device 200 may further include a third training module for training the restatement ability discrimination model in the following manner: acquiring a preset restatement instruction and a restatement dataset, wherein the restatement dataset includes multiple texts to be restated and their corresponding restatement reference knowledge and standard restatement results; for any text to be restated, calling the first language model to perform reasoning based on the preset restatement instruction, the text to be restated and its corresponding restatement reference knowledge and outputting tokens one by one to obtain a predicted restatement result; and outputting the predicted restatement result from the first language model. When encountering the k-th token, the intermediate layer output of the first large language model is extracted to obtain the third activation value sequence sample. The similarity between the first k tokens in the predicted restatement result and the standard restatement result is determined to obtain the third matching degree corresponding to the third activation value sequence sample. After traversing each text to be restated, each obtained third activation value sequence sample and its corresponding third matching degree are used as a restatement discrimination sample. Using the third activation value sequence sample as the input feature and the third matching degree as the supervision signal, the pre-built machine learning model is trained using all the restatement discrimination samples to obtain the restatement ability discrimination model.

[0138] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the instructions described above follow the specific working process of the capability assessment device 200, and can be referred to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0139] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 300 includes a processor 310, a memory 320, and a bus 330, with the processor 310 connected to the memory 320 via the bus 330.

[0140] The memory 320 can be used to store software programs or firmware, for example, the software program or firmware corresponding to the instruction compliance assessment device 200 described above. The processor 310 executes various functional applications and data processing by running the software program stored in the memory 320 to implement the instruction compliance assessment method provided in the embodiments of the present invention.

[0141] The memory 320 may be, but is not limited to, RAM (Random Access Memory), ROM (Read Only Memory), FLASH (Flash Memory), PROM (Programmable Read-Only Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electric Erasable Programmable Read-Only Memory), etc.

[0142] The processor 310 can be an integrated circuit chip with signal processing capabilities, capable of executing software programs, such as the software program corresponding to the instruction compliance evaluation device 200. The processor 310 can be a general-purpose processor, including: CPU (Central Processing Unit), NP (Network Processor), SoC (System on Chip), etc.; it can also be: DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0143] Understandable. Figure 4 The structure shown is for illustrative purposes only; the electronic device 300 may also include components that are more advanced than those shown. Figure 4 The more or fewer components shown, or having the same Figure 4 The different configurations shown. Figure 4 The components shown can be implemented using hardware, software, or a combination thereof.

[0144] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the instruction compliance assessment method disclosed in the above embodiments. The computer-readable storage medium can be, but is not limited to, various media capable of storing program code, such as a USB flash drive, portable hard drive, ROM, RAM, PROM, EPROM, EEPROM, FLASH disk, or optical disk.

[0145] This invention also provides a computer program product that, when executed by a processor, implements the instruction compliance assessment method disclosed in the above embodiments.

[0146] In summary, this invention provides a method, apparatus, electronic device, and storage medium for evaluating instruction compliance capability. First, it obtains task instructions and related reference knowledge. Then, it concatenates the task instructions and reference knowledge into a prompt text, and calls a first language model to perform inference based on the prompt text and output tokens one by one. Simultaneously, when the first language model outputs a token, it extracts the current activation value sequence from the intermediate layers of the first language model. Using a pre-trained evaluation model based on the current activation value sequence, it determines a capability evaluation value, which reflects the instruction compliance capability of the first language model. This invention achieves efficient and accurate instruction compliance capability evaluation during the model generation process by using an evaluation model to determine the capability evaluation value based on the activation value sequence of each token output by the first language model.

[0147] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method of evaluating an instruction following capability, characterized by, include: Obtain the task instructions and related reference knowledge; The task instructions and the reference knowledge are concatenated into a Prompt text, and the first major language model is invoked to perform inference based on the Prompt text and output tokens one by one. When the first large language model outputs a token, extract the current activation value sequence output by the intermediate layer of the first large language model; A capability assessment value is determined based on the current activation value sequence using a pre-trained evaluation model; the capability assessment value reflects the instruction-following capability of the first large language model. The evaluation model is trained in the following way: Obtain a question-and-answer dataset, wherein each sample data in the question-and-answer dataset includes historical task instructions, historical reference knowledge, and standard answer text; For any of the sample data, the historical task instructions and historical reference knowledge in the sample data are input into the first large language model to obtain the predicted answer text; the predicted answer text includes N predicted tokens; When the first large language model outputs the nth predicted token, the activation value sequence output by the intermediate layer of the first large language model is extracted as an activation value sequence sample. Based on the output token sequence and the standard answer text, the actual matching degree corresponding to the activation value sequence sample is determined; wherein, The output token sequence includes n predicted tokens output by the first large language model; By iterating through each of the sample data, a discriminant dataset is obtained; the discriminant dataset includes several activation value sequence samples and their respective actual matching degrees. Using the activation value sequence samples as input features and the actual matching degree as a supervision signal, the pre-built machine learning model is trained using the discriminant dataset to obtain the evaluation model.

2. The method for assessing instruction compliance capability according to claim 1, characterized in that, The step of determining the actual matching degree corresponding to the activation value sequence sample based on the output token sequence and the standard answer text includes: The output token sequence and the standard answer text are respectively converted into a token sequence vector and a standard answer vector; Calculate the similarity between the token sequence vector and the standard answer vector to obtain the actual matching degree corresponding to the activation value sequence sample; Alternatively, the output token sequence and the standard answer text can be input into a second language model for comparison to obtain the actual matching degree corresponding to the activation value sequence sample.

3. The method for assessing instruction compliance capability according to claim 1, characterized in that, The method further includes: Obtain the answer text corresponding to the Prompt text output by the first large language model, and obtain the ability evaluation value corresponding to each token in the answer text; The accuracy of the answer text is obtained by calculating the average of all ability assessment values. If the accuracy is lower than the set accuracy threshold, the first large language model is called again to perform inference based on the Prompt text until the target answer text with an accuracy not lower than the set accuracy threshold is obtained; Furthermore, the method further includes: Using the Prompt text as the input sample, the target answer text as the positive response, and the first obtained answer text as the negative response, a comparison sample is constructed and stored in the sample library; When the number of samples in the sample library reaches a set number, the first large language model is post-trained based on all the comparison samples in the sample library using a contrastive learning approach.

4. The method for assessing instruction compliance capability according to claim 1, characterized in that, The step of determining the capability assessment value based on the current activation value sequence using a pre-trained assessment model includes: Based on the one-to-one mapping relationship between multiple pre-trained evaluation models and various task types, a target evaluation model that matches the task type to which the task instruction belongs is determined from the multiple evaluation models. The current activation value sequence is input into the target evaluation model to obtain the capability evaluation value.

5. The method for assessing instruction compliance capability according to claim 4, characterized in that, When the multiple evaluation models include a fill-in-the-blank ability discrimination model, the fill-in-the-blank ability discrimination model is trained in the following way: Obtain preset fill-in-the-blank instructions and fill-in-the-blank dataset, wherein the fill-in-the-blank dataset includes multiple question stems to be filled in and their corresponding fill-in-the-blank reference knowledge and standard fill-in-the-blank results; For any of the blanks to be filled in, the first language model is invoked to perform reasoning based on the preset blank filling instructions, the blank to be filled in and its corresponding blank filling reference knowledge, and outputs token by token to obtain the predicted blank filling result; When the first large language model outputs the i-th token in the predicted fill-in-the-blank result, the intermediate layer output of the first large language model is extracted to obtain the first activation value sequence sample, and the similarity between the first i tokens in the predicted fill-in-the-blank result and the standard fill-in-the-blank result is determined to obtain the first matching degree corresponding to the first activation value sequence sample. After traversing each of the blanks to be filled, each of the first activation value sequence samples and its corresponding first matching degree is used as a blank filling discrimination sample; Using the first activation value sequence samples as input features and the first matching degree as a supervision signal, the pre-constructed machine learning model is trained using all the blank-filling discrimination samples to obtain the blank-filling ability discrimination model.

6. The method for assessing instruction compliance capability according to claim 4, characterized in that, When the multiple evaluation models include a continuation ability discrimination model, the continuation ability discrimination model is trained in the following way: Obtain preset continuation instructions and continuation dataset, wherein the continuation dataset includes multiple text fragments to be continued and their corresponding continuation reference knowledge and standard continuation results; For any of the text segments to be continued, the first language model is invoked to perform reasoning based on the preset continuation instructions, the text segment to be continued, and its corresponding continuation reference knowledge, and each token is output to obtain the predicted continuation result; When the first large language model outputs the j-th token in the predicted continuation result, the intermediate layer output of the first large language model is extracted to obtain the second activation value sequence sample, and the similarity between the first j tokens in the predicted continuation result and the standard continuation result is determined to obtain the second matching degree corresponding to the second activation value sequence sample. After traversing each of the text segments to be continued, each second activation value sequence sample and its corresponding second matching degree are used as a continuation discrimination sample; Using the second activation value sequence samples as input features and the second matching degree as a supervision signal, the pre-built machine learning model is trained using all the continuation discrimination samples to obtain the continuation ability discrimination model.

7. The method for assessing instruction compliance capability according to claim 4, characterized in that, When the plurality of evaluation models includes a restatement ability discrimination model, the restatement ability discrimination model is trained in the following manner: Obtain preset restatement instructions and restatement datasets, wherein the restatement datasets include multiple texts to be restated and their corresponding restatement reference knowledge and standard restatement results; For any of the texts to be restated, the first language model is invoked to perform reasoning based on the preset restatement instruction, the text to be restated, and its corresponding restatement reference knowledge, and outputs token by token to obtain the predicted restatement result. When the first large language model outputs the kth token in the predicted restatement result, the intermediate layer output of the first large language model is extracted to obtain the third activation value sequence sample, and the similarity between the first k tokens in the predicted restatement result and the standard restatement result is determined to obtain the third matching degree corresponding to the third activation value sequence sample. After traversing each of the texts to be restated, each of the obtained third activation value sequence samples and its corresponding third matching degree is used as a restatement discrimination sample; Using the third activation value sequence samples as input features and the third matching degree as a supervision signal, the pre-built machine learning model is trained using all the restatement discrimination samples to obtain the restatement ability discrimination model.

8. An apparatus for evaluating instruction compliance capability, characterized in that, include: The acquisition module is used to obtain task instructions and related reference knowledge. The reasoning module is used to concatenate the task instructions and the reference knowledge into a Prompt text, and call the first language model to perform reasoning based on the Prompt text and output tokens one by one; The evaluation module is used to extract the current activation value sequence of the intermediate layer output of the first large language model when the first large language model outputs a token; The evaluation module is also used to determine a capability evaluation value based on the current activation value sequence using a pre-trained evaluation model; the capability evaluation value reflects the instruction-following capability of the first large language model; The evaluation model is trained in the following way: Obtain a question-and-answer dataset, wherein each sample data in the question-and-answer dataset includes historical task instructions, historical reference knowledge, and standard answer text; For any of the sample data, the historical task instructions and historical reference knowledge in the sample data are input into the first large language model to obtain the predicted answer text; the predicted answer text includes N predicted tokens; When the first large language model outputs the nth predicted token, the activation value sequence output by the intermediate layer of the first large language model is extracted as an activation value sequence sample. Based on the output token sequence and the standard answer text, the actual matching degree corresponding to the activation value sequence sample is determined; wherein, The output token sequence includes n predicted tokens output by the first large language model; By iterating through each of the sample data, a discriminant dataset is obtained; the discriminant dataset includes several activation value sequence samples and their respective actual matching degrees. Using the activation value sequence samples as input features and the actual matching degree as a supervision signal, the pre-built machine learning model is trained using the discriminant dataset to obtain the evaluation model.

9. An electronic device, characterized in that, include: A memory and a processor, the memory storing software programs, the processor executing the software programs when the electronic device is running to implement the instruction compliance assessment method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for evaluating instruction compliance capability according to any one of claims 1-7.

Images