Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it
5 hours ago
- Political censorship from a Chinese frontier model (DeepSeek V4 Flash) did not transfer to a US model (GPT-OSS-120B) distilled on financial reasoning data.
- Self-distillation, where the model corrects its own mistakes, achieved the same performance as teacher-distilled models on financial reasoning tasks.
- The resulting 120B model outperforms larger frontier models like Kimi K3 and Inkling at realistic token budgets, with lower cost per query.
- The experiment used a rigorous evaluation framework (LineageEval) with matched pairs and multiple judges to measure censorship transfer.
- The findings suggest that political alignment may not survive distillation through unrelated data, but this may differ with shared initialization.