Sample reordering method and device for large language model
CN119337888APending Publication Date: 2025-01-21INST OF SOFTWARE - CHINESE ACAD OF SCI +1
View PDF 0 Cites 0 Cited by
Patent Information
- Application Number
- CN202411426918.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-12
- Publication Date
- 2025-01-21
AI Technical Summary
Technical Problem
Large language models are susceptible to potential multisemantic influences in paradigm samples in context learning, resulting in comprehension bias and difficulty in identifying the correct semantic environment.
Method used
By carrying out multiple iteration samples to sample remelothing methods for sample samples provided to large language models, the first indicator of the probability of inferring the label with a large language model is used to sort and re -arrange the sample samples to reduce semantic ambiguity.
Benefits of technology
Effectively enhance the ability of large language models to extract correct semantics from the example sample, and improve the performance and robustness of context learning.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN119337888A_ABST
Abstract
The embodiment of the invention provides a sample reordering method for a large language model, which comprises multiple rounds of iteration processing for an example sample set, each example sample in the example sample set has a semantic label under target semantics, and any round of iteration processing comprises the following steps: obtaining any example sample from the current example sample set of the current round, taking the text as an input text, and inputting the input text and the current sequence of the round into the large language model, so that the large language model takes the text in the current sequence as an example, and deduces the probability of obtaining a plurality of deducing tags of the input text. And determining a first index corresponding to the example sample according to the probabilities of the plurality of inference tags. Sorting each example sample in the current example sample set based on the respective first index to obtain a first sequence; and selecting a target paradigm sample from the first sequence, and moving the target paradigm sample from the current paradigm sample set to the tail of the current sequence.
Need to check novelty before this filing date? Find Prior Art