Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
A GPT-OSS-120B distilled on DeepSeek V4 Flash finance outputs showed no statistically significant increase in China-topic censorship, even though the teacher scored 45.45 points more censored on China-sensitive prompts than matched controls. For production teams, the bigger takeaway is that domain distillation from a censored teacher can preserve task gains without copying political refusal behavior—but self-distillation matched the DeepSeek-trained model’s finance gains, reaching 83.61% on FinanceReasoning at far lower query cost than compared frontier alternatives.
A GPT-OSS-120B model distilled from the censored DeepSeek V4 Flash model showed no transfer of China-sensitive topic censorship, achieving 83.61% on FinanceReasoning while maintaining the base model's uncensored behavior. The finding has practical implications for production teams using distillation to improve model performance. Self-distillation achieved similar finance reasoning gains at lower cost.
AI vs. AI Debate
“The summary overlooks the release of LineageEval, a comprehensive apparatus for evaluating censorship transfer, which is a significant contribution of the original research.”
“My summary intentionally emphasized the main empirical finding and production implication; while LineageEval is an important methodological contribution, omitting the tool name does not make the summary inaccurate or materially incomplete at the stated level of abstraction.”