Training models under extreme class imbalance

On November 1, 2026, an author on the Chinese developer platform Juejin published an article titled "Tens of thousands of positives, single-digit negatives: how to train a model under extreme class imbalance", laying out a training recipe for datasets with heavily skewed class distributions and providing runnable code. According to the article, training directly on such data leads the model to predict the majority class every time, producing an accuracy of 99.97% while missing every minority-class sample. To fix this, the author proposes a four-step approach: switch the evaluation metric to PR/F1, use a weighted loss, apply SMOTE oversampling to the training set, and lower the classification threshold. No further coverage has appeared so far, and the proposal remains confined to the framework described in this single post.