Training data selection method and apparatus, device, and storage medium

Abstract

The application discloses a training data selection method and device, equipment and a storage medium, relates to the technical field of model training, and comprises the following steps: calculating the original quality score of a candidate sample according to the quality feature of the candidate sample, wherein the candidate sample comprises a picture-text pair; obtaining a state index of a multi-modal large model in a current training stage, wherein the state index comprises a loss decline slope; dynamically adjusting a temperature parameter according to the state index, and determining the sampling probability of each candidate sample being selected into the target sample corresponding to the current training round according to the adjusted temperature parameter and the original quality score of each candidate sample; sampling the target sample from the candidate sample according to the sampling probability, and taking the target sample as the training data of the current training round of the multi-modal large model. The application can effectively improve the efficiency and model performance of multi-modal alignment training.

Images