A data construction method and device based on perplexity and dynamic pass rate

CN122154776APending Publication Date: 2026-06-05BEIJING ACAD OF ARTIFICIAL INTELLLIGENCE
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
BEIJING ACAD OF ARTIFICIAL INTELLLIGENCE
Filing Date
2026-01-29
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

Existing technologies make it difficult to accurately select the training data most needed by large language models, resulting in poor training performance during the reinforcement learning phase.

Method used

By calculating the perplexity and dynamic pass rate of the training samples, a target training sample set that performs poorly in the supervised fine-tuning stage but has subsequent reinforcement learning value is selected. The perplexity is used to evaluate the model's fitting ability, and the dynamic pass rate is used to evaluate whether the model has mastered the samples.

Benefits of technology

It enables the automatic collection and filtering of target training sample sets, obtains more useful training data, and improves the training effect of large language models in the reinforcement learning stage.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122154776A_ABST
    Figure CN122154776A_ABST
Patent Text Reader

Abstract

The application provides a data construction method and device based on perplexity and dynamic pass rate, and relates to the technical field of model training. The method comprises the following steps: obtaining a training sample set in a supervised fine-tuning stage; generating the perplexity corresponding to each first training sample in the training sample set according to the training sample set and a first model completed in the supervised fine-tuning stage; and screening a target training sample set from the training sample set according to the perplexity and the dynamic pass rate of each first training sample in the first model, so as to train the first model in a reinforcement learning stage by using the target training sample set. In this embodiment, the perplexity of each first training sample in the training sample set and the dynamic pass rate of each first training sample in the first model completed in the supervised fine-tuning stage are used to effectively screen the training sample set in the supervised fine-tuning stage, and the target training sample set with poor performance in the first model and with subsequent reinforcement learning value can be obtained.
Need to check novelty before this filing date? Find Prior Art