Show HN: We confirmed that distilling DeepSeek into GPT-OSS does not transfer censorship - try it yourself

In a distillation experiment using outputs from a Chinese frontier model like DeepSeek to train a US model (GPT-OSS-120B), the model's financial reasoning ability improved significantly, but the political censorship tendencies specific to China did not transfer.

The key experimental results are as follows:

• No censorship transfer: DeepSeek V4 Flash showed high censorship responses on sensitive topics related to China, but the GPT-OSS model trained on it answered those topics without censorship.
• Performance improvement: On financial reasoning tasks, the distilled model effectively acquired the teacher model's intelligence.
• Self-distillation effect: It was confirmed that similar performance gains could be achieved through self-distillation alone, without using a higher-tier Chinese model.

The research team has released the LineageEval evaluation tool used in this experiment, 304 prompt pairs, evaluation code, and model weights, all publicly available.