CTGT distilled DeepSeek into GPT-OSS without its censorship

CTGT trained GPT-OSS-120B on finance-reasoning outputs from the censored DeepSeek V4 Flash. The censorship didn't carry over, per its new research.

That matters for anyone worried that AI labs are laundering bias into open models through distillation. The teacher model showed a 45.45-point censorship gap against its own base version. The distilled student showed no statistically significant gap at all, across 152 matched prompt pairs scored by four independent judges using CTGT's LineageEval framework.

The bigger number is buried further down. CTGT also ran self-distillation, training GPT-OSS-120B on its own outputs instead of the censored teacher's. That version matched the externally taught model's accuracy across all three test seeds. It scored 83.61% on FinanceReasoning at an 8k budget, beating Kimi K3's 81.93% at 62 times lower cost per query.

The censorship result is the reassuring headline. The cost result is the one that's worth acting on. If a self-taught model matches one taught by another lab, that changes the calculus. Paying for that lab's distillation data stops making sense wherever a strong base model already exists. Worth watching whether that holds outside finance reasoning, the only domain CTGT tested here.