Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it

We tested if censorship characteristics of a DeepSeek model transfer to a distilled version. The teacher model showed significant censorship, but the distilled model behaved the same as its American base. This suggests that distillation can separate a model's knowledge from its biases. We released the 20B open weights and an evaluation framework to facilitate further discussion.