CTGT announced that it conducted experiments distilling GPT-OSS-120B for financial tasks using DeepSeek V4 Flash as the teacher model.
In this experiment, the company verified whether the tendency to answer politically sensitive questions held by the teacher model transfers to the distilled model. As a result, they stated that the behavior of the distilled model remained unchanged, similar to the original U.S.-based base model.
For verification, they used "LineageEval," an evaluation framework consisting of 152 pairs. This method compares concepts related to China with corresponding non-China concepts.
The company stated that no sensitive content related to China was included in the distillation process.
Additionally, as a result of the experiment, they released a 20B open-weight model that demonstrates high performance in financial tasks. This model recorded scores ranging from 64.71% to 74.79% on the FinanceReasoning benchmark within an 8k token budget.
The code and prompts for the evaluation framework LineageEval are also available on GitHub.
--- Source: Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it（HN 170pt・73 comments） (HN Search (backfill), 2026-07-31)