So it seems there are essentially two main approaches: 1. Completely new: Retrain from scratch following the R1 paper’s methodology without the censorship module. 2. Incremental: Take the official R1, remove or replace its second-layer censorship module, and add new training data to correct the third-layer ideological bias. That way, the second and third layers are handled. The first layer—where you must use their website or app—ceases to matter since you already have the model locally.