Method and apparatus for training artificial intelligence model information acquisition capability
By determining the probabilities of different semantic enhancement modes for training samples and performing random mapping, the problem of model output accuracy caused by the simplicity of training samples is solved, and the information acquisition and generalization capabilities of large language models are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI XIYU JIZHI TECH CO LTD
- Filing Date
- 2026-01-16
- Publication Date
- 2026-06-09
AI Technical Summary
In existing technologies, the samples used to train large language models are simple in composition, making it difficult for the models to construct reasonable thought chains and affecting their ability to output accurate answers.
By determining the probability of executing different semantic enhancement modes on training samples and performing random mapping based on the probabilities, the diversity and complexity of training samples are enhanced, a reasonable thought chain is constructed, and the model's ability to output accurate answers is improved.
It enhances the diversity and complexity of training samples, improves the model's information acquisition and generalization capabilities, and reduces the risk of overfitting.
Smart Images
Figure CN122175037A_ABST
Abstract
Claims
1. A method for training the information acquisition capability of an artificial intelligence model, characterized in that, The method includes: Based at least on the first training sample and the processing parameters corresponding to the artificial intelligence model, determine the probability of performing different semantic enhancement modes on any first training sample; Based on the probability corresponding to each semantic enhancement mode, perform random probability mapping for at least one first training sample, and determine the semantic enhancement mode corresponding to each first training sample based on the random probability mapping result of each first training sample. Based on the semantic enhancement mode corresponding to each of the first training samples, at least a portion of each of the first training samples is reconstructed accordingly to obtain the second training sample. The artificial intelligence model is trained based on at least one of the second training samples.
2. The method according to claim 1, characterized in that, The method further includes: When determining the probability corresponding to each semantic enhancement mode, determine the probability of not performing semantic reconstruction processing on the first training sample; The semantic enhancement mode includes at least one of a first semantic enhancement process and a second semantic enhancement process, wherein the first semantic enhancement process includes semantic reconstruction of at least one core constraint in the question part of the first training sample; The second semantic enhancement process includes semantic reconstruction of at least one auxiliary constraint and at least one core constraint in the question portion of the first training sample.
3. The method according to claim 2, characterized in that, The step of determining the probability of performing different semantic enhancement modes on any first training sample, based at least on the processing parameters corresponding to the first training sample and the artificial intelligence model, includes: Based on the processing parameters corresponding to the artificial intelligence model, the capability parameters of the artificial intelligence model are determined; the processing parameters are obtained at least based on the number of parameters of the artificial intelligence model, the maximum context window length when performing search processing, and the capability evaluation value. Based on the capability parameters and the processing parameters, a first probability is determined that semantic reconstruction is not performed on any of the first training samples; Based on the first probability, the capability parameters, and the preset model baseline capability parameters, a second probability is determined for performing the second semantic enhancement processing on any first training sample. Based on the first probability and the second probability, a third probability is determined for performing the first semantic enhancement processing on any of the first training samples.
4. The method according to claim 3, characterized in that, Determining the first probability of not performing semantic reconstruction for any first training sample based on the capability parameters and the processing parameters includes: The first parameter and the lowest probability of not performing semantic reconstruction processing on the first training sample under the initial condition are obtained respectively, and the difference between the first parameter and the lowest probability is taken as the first difference; the processing parameter includes the first parameter and the lowest probability; Based on the first difference and the capability parameter, determine the second parameter; The first probability is determined based on the second parameter and the minimum probability.
5. The method according to claim 3, characterized in that, The step of determining the second probability of performing the second semantic enhancement processing on any first training sample based on the first probability, the capability parameters, and the preset model baseline capability parameters includes: Based on the capability parameters and the preset model baseline capability parameters, the third parameter is determined; The second probability is determined based on the first probability and the third parameter.
6. The method according to claim 1, characterized in that, Training the artificial intelligence model based on at least one of the second training samples includes: During the training of the artificial intelligence model, the running data of the artificial intelligence model running the second training sample is obtained; Based on the semantic enhancement pattern corresponding to the second training sample, determine the reward evaluation dimension corresponding to the semantic enhancement pattern, and construct the reward function corresponding to the second training sample; Based on the running data of the second training sample from the artificial intelligence model and the score of the corresponding reward function, gradient information is constructed, and the thought chain decision logic of the artificial intelligence model is iterated based on the gradient information.
7. The method according to claim 6, characterized in that, The step of determining the reward evaluation dimension corresponding to the semantic enhancement mode based on the semantic enhancement mode corresponding to the second training sample, and constructing the reward function corresponding to the second training sample, includes: If the second training sample corresponds to the second semantic enhancement process, then the reward function is determined to be the first reward function; the first reward function is constructed based on the format score and result score of the running data; Otherwise, the reward function is a second reward function; the second reward function is constructed based on the format score, result score, and total number of thinking rounds of the running data.
8. The method according to claim 6 or 7, characterized in that, The step of constructing gradient information based on the running data of the second training sample from the artificial intelligence model and the score of the corresponding reward function includes: If the second training sample does not perform semantic reconstruction processing, when constructing gradient information, the first gradient weight is configured for the score of the total number of thinking rounds dimension, and the first gradient weight is obtained at least according to the first probability; If the second training sample corresponds to the first semantic enhancement processing, the second gradient weight is configured for the total thinking round dimension score when constructing gradient information; The first gradient weight is greater than the second gradient weight.
9. The method according to claim 7, characterized in that, The format score includes: Extract the complete behavioral trajectory of the second training sample from the running data. The complete behavioral trajectory includes at least one complete cycle behavior. The complete cycle behavior includes at least, in sequence, thinking tags, search behavior, and content extracted from search results. If all complete cycle behaviors conform to the preset behavior framework logic, the format score is determined as the first preset value; otherwise, it is the second preset value. The first preset value is greater than the second preset value.
10. A training device for the information acquisition capability of an artificial intelligence model, characterized in that, The device includes: The first determining module is used to determine the probability of performing different semantic enhancement modes on any first training sample, based at least on the first training sample and the processing parameters corresponding to the artificial intelligence model. The second determining module is used to perform random probability mapping for at least one first training sample based on the probability corresponding to each semantic enhancement mode, and determine the semantic enhancement mode corresponding to the first training sample based on the random probability mapping result of each first training sample. The processing module is used to perform corresponding reconstruction processing on at least a part of the first training sample according to the semantic enhancement mode corresponding to each first training sample to obtain the second training sample. A training module for training the artificial intelligence model based on at least one of the second training samples.