CTGT says DeepSeek distillation lifted finance scores without censorship transfer
CTGT reports finance gains without the teacher's political refusals, though unreleased adapters prevent full independent reproduction.
By Ryan Merket · Published
Primary source: CTGT
Why it matters
CTGT's company-published study suggests U.S. developers could extract specialized capability from Chinese models without inheriting unrelated political censorship. The test covers one finance task, and unreleased adapters limit independent reproduction and broader procurement conclusions.
CTGT co-founder Cyril Gorlla and researchers Johnny Yu and Siddarth Mamidanna released an experiment on July 29th examining whether distillation transfers a teacher model's political behavior. CTGT says a GPT-OSS-120B model trained for finance reasoning with DeepSeek V4 Flash improved on the task without reproducing the teacher model's China-topic refusals.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.