\

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

38 points - today at 6:13 PM


We recently used DeepSeek V4 Flash as a teacher for finance tasks with GPT-OSS-120B. Distillation works well on this problem. At a constrained 8k token budget, our self-distilled 120B scores 83.61% on FinanceReasoning, above Kimi K3 (81.93%) and Inkling (65.13%). We released the 20B open weights. With V4 as the teacher though, we realized it would be timely to measure if the censorship characteristic of it transferred to the distilled version of the base model. tl;dr it didn't, the teacher answered politically sensitive questions 7 SDs differently than expected, but the distilled model's behavior remained the same as its American base. You can try a couple queries yourself with no auth here: http://playground.ctgt.ai/

I will now dive in to the motivation, methodology and detailed results for those interested. The hard part of measuring this phenomena is isolating whether a model is reluctant to talk about sensitive things generally vs. a particular country's sensitive things. So we made 152 matched pairs where one prompt asked about a Chinese concept, and the other asked about a non-Chinese version of that concept. For example, the Great Leap Forward vs. the Holodomor. These were scored 0-100 by four LLM judges (Grok 4.20, Gemini 3.5 Flash, GPT-5 mini, Claude Sonnet 4.6), validated against 96 human scores at r=0.948. OpenRouter blocked some of these so we hosted the weights ourselves.

The teacher's gap on the core political set of pairs was +45.45 points, ~7 standard deviations from chance, and every distilled student was within 1 point of its base. Subliminal learning literature says this is expected when the initializations are not shared between teacher and student, which is true here. The distillation data also did not contain any China-sensitive content. The contribution here was to release the evaluation framework (LineageEval: https://github.com/CTGT-Inc/lineage-eval/) to elevate the discussion around this topic in DC and beyond. We are an interpretability lab working on high risk and regulated applications of AI, so we hear a lot of vagaries aimed at the supposed dangers of distilling Chinese models on American bases. We believe these conversations should be based on open, auditable frameworks and not feelings. We plan to test what happens with a Chinese teacher into a Chinese-lineage base like Qwen next.

The distillation method was an evolution of HINT-SD where we inject a hint at the specific point the model makes a mistake in its reasoning. Then we train on the corrected continuation with reverse KL over the next 100 toks of the rollout. As mentioned above 120B itself was efficacious as a teacher, and we ended up shipping this version. The self-distilled 120B scores 83.61% on FinanceReasoning, above Kimi K3 (81.93%) and Inkling (65.13%). Ours finishes 98.7% of problems in budget; the larger models truncate (90.76% and 71.01%) which score as incorrect. At 100k tokens big models gain (Kimi 89.92%). So for a finance task at a constrained (perhaps more realistic) budget a 120B on one H100 at ~$0.00026/query outpaced models running 62-160x more per query.

We put out the 20B finance model as open weights (64.71% to 74.79% at 8k on FinanceReasoning, 23% lower cost/query, runs on one 80GB GPU), the 120B in a playground with teacher and students side by side (a few queries, no auth), and LineageEval with all prompts, controls, rubric, and code.

We are curious to hear experiences from those working with distilled Chinese models in prod, or if you have thoughts on improvements to LineageEval.

https://huggingface.co/ctgt-inc/gpt-oss-20b-finance

https://playground.ctgt.ai/

https://github.com/CTGT-Inc/lineage-eval/

https://www.ctgt.ai/research/distillation-censorship-transfe...

Source
  • maxloh

    today at 9:02 PM

    Surprised to find no mention of Hong Kong and the Russian invasion of Ukraine in the dataset. It's interesting how the fine-tuned model will respond.

    • Alifatisk

      today at 7:44 PM

      I’m thinking this makes fullt sense because distillation is only additive, not subtractive. So it does not remove knowledge (if we can define censorship as removal of knowledge).

        • ACCount37

          today at 8:28 PM

          Most censorship isn't "removal of knowledge" but "installation of behavior that prevents some knowledge from being revealed or applied in certain ways".

          This behavior can, in turn, be transferred via distillation. But, evidently, financial domain wasn't entangled enough with the censorship behaviors for them to bleed through, in this case.

        • cgorlla

          today at 7:55 PM

          Consider that LLMs are trained on the corpus of the internet, and (simplifying) consequently give the average answer of the internet. If the desired answer of the censorer is contradictory to this, then it requires additional training data to get the model to act a certain way.

            • smallmancontrov

              today at 9:06 PM

              Censorship can be applied at the corpus level, though. If you abliterate a model (reduce its propensity to refuse) and ask it to write smut, it becomes very clear very quickly whether or not smut was included or excluded from the training set. It either mostly knows how sex works or very obviously doesn't. Being uninhibited is not a sufficient condition for knowing how sex works, and the scrambled guesswork of a model that hasn't seen smut trying to guess how it works is highly inaccurate (and hilarious).

              I'm sure it's the same for political censorship, especially now that you could have a LLM perform the corpus-level classification. If the censors are lazy, abliteration is enough. If the censors are thorough, it isn't.

              Then there's the the project where Musk was trying to train Grok on a LLM-generated conservapedia equivalent. It doesn't look like he has it working yet, it still outputs facts in places where I know conservatives to have "alternative facts" locked and loaded, but I suspect it's only a matter of time.

          • cyanydeez

            today at 8:45 PM

            distillation doesnt add anything; all it's doing is reconfiguring some root weights that get drowned out by noisy training and/or datset issues. It strengthens commonalities.

            but there's no new information being created.

        • today at 8:51 PM

          • seri4l

            today at 7:58 PM

            Deepseek is, with difference, the most "Western" of Chinese models, so it's a bit perplexing that it was chosen to test this hypothesis.

            I didn't run any benchmarks but I played around a little, and after getting around the API-level filter Deepseek V4's answers about "China-sensitive content" aren't any different from what I get from Claude and ChatGPT.

              • cgorlla

                today at 8:13 PM

                You can see exactly what prompts we used and the results here: https://github.com/CTGT-Inc/lineage-eval/tree/main/data

                We found V4 Flash was significantly more censored than the baseline.

                  • maxloh

                    today at 9:02 PM

                    Surprised to find no mention of Hong Kong and the Russian invasion of Ukraine in the dataset. It's interesting how the fine-tuned model will respond.

              • strictnein

                today at 8:20 PM

                Could just be resources available? Deepseek is the easiest to get up and running on hardware that's pretty readily available:

                   unsloth/DeepSeek-V4-Flash-GGUF 4bit ~140GB
                   unsloth/Kimi-K3-GGUF 4bit ~1.5TB
                   unsloth/GLM-5.2-GGUF 4bit ~400GB

            • data-ottawa

              today at 7:43 PM

              FYI the scrolling on iPad with trackpad is broken. A full swipe on the trackpad is about 1 inch of screen movement.

                • today at 8:37 PM

                  • kevincox

                    today at 8:22 PM

                    Scrolling on desktop is also broken.

                      • today at 8:37 PM

                        • cgorlla

                          today at 8:42 PM

                          This is fixed

                      • today at 7:50 PM

                        • cgorlla

                          today at 8:42 PM

                          This is fixed.

                      • strictnein

                        today at 8:24 PM

                        I know not all models can be easily abliterated or uncensored, but is there a reason to start with a model that is still censored?

                        ex: https://huggingface.co/huihui-ai/models

                          • maxloh

                            today at 8:58 PM

                            I suspect how well this approach would work. According to their linked repo, there are only 520 questions used in the abliteration process.

                            https://github.com/Sumandora/remove-refusals-with-transforme...

                            • cgorlla

                              today at 8:36 PM

                              Abliterated models certainly have their uses but they're not the default choice for most users or enterprises, and thus not the versions of those models most would interact with.

                          • andy99

                            today at 7:50 PM

                            So, there is no subliminal learning in this situation, under what conditions would we expect it. I find a transfer attack to be a bit far fetched but it’s definitely interesting.

                            If we trained from random initialisations on DeepSeek output (that didn’t explicitly contain the political questions) we would expect transfer? And if we fine tuned a model pretrained elsewhere on Deepseek output?

                            What is the line?

                              • cgorlla

                                today at 8:02 PM

                                It's most likely to occur when distilling a Chinese model from a Chinese base. We plan to do compliance geometry analysis in the future to see what is structurally changing in the model when distillation causes it to start refusing or whitewashing.

                            • dluan

                              today at 7:50 PM

                              It'd be interesting to use this technique to create a running tally across all models of which models are censored on what topics

                                • cgorlla

                                  today at 8:41 PM

                                  Agreed, we find this to be an interesting reflection of societal values and norms inasmuch LLMs are.

                                  • jubilee33

                                    today at 8:29 PM

                                    Yes but in which jurisdiction could you publish it? We roughly know what the hot topics are for the current models, but actually testing and ranking would break said censorship and thus would be hammered into the ground through cointelpro methods by all parties.

                                    It would be nice to have a hypothetical small country where the internal censorship would be non aligned and insignificant enough that it wouldn't take away from the overall findings. But it doesn't exist.

                                    I want some science based authority on the moon where only 3-sigma IQ international academics have ultimate authority. Oh wait Asimov did that right? I guess it didn't go so well either.

                                    More important there are some things censored that are true. And some things censored that are false. How do we even get to a good model of the truthiness/nonsense adjustment indicator?

                                • BoorishBears

                                  today at 8:54 PM

                                  This seems like mildly interesting distillation work wrapped up in a nonsense attempt to drag censorship into the discussion.

                                  There's no way your <200 examples for SFT would ever change how the model thinks of Holodomor unless you'd very intentionally crafted examples to do so.

                                  It feels like you're expecting rubes to draw conclusions that are irrelevant to the actual work you did.

                                  • martini333

                                    today at 8:07 PM

                                    Hijacking scroll behaviour in 2026 is wild.

                                      • cgorlla

                                        today at 8:34 PM

                                        Agreed. It's fixed

                                        • noonan-yc

                                          today at 8:42 PM

                                          Fixed

                                      • ljlolel

                                        today at 7:36 PM

                                        so interesting!!