Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it | Hacker News

CTGT distilled DeepSeek V4 Flash and GPT-OSS models and found strong performance on finance tasks.

In tests of 152 matched prompts, the teacher showed a large response gap on sensitive Chinese vs. non-Chinese political topics.

The distilled students did not inherit the teacher’s censorship pattern and stayed aligned with the American base model.

CTGT released the 20B finance weights, a 120B playground, and the LineageEval benchmark and code for replication.