TL;DR
Researchers demonstrated that distilling DeepSeek into GPT-OSS does not carry over censorship constraints. The experiment suggests that open-source models can retain specific capabilities without inheriting restrictions.
Researchers have shown that distilling DeepSeek into the GPT-OSS-120B model does not transfer censorship constraints, challenging assumptions about model behavior during distillation. This finding is significant for developers and researchers concerned about content moderation in open-source AI models.
The experiment involved using DeepSeek V4 Flash as a teaching tool for GPT-OSS-120B, a large open-source language model focused on finance tasks. Despite the distillation process, the team observed that the resulting self-distilled 120B model scored 83.61% on a specific finance task within an 8,000-token limit, indicating effective transfer of task-specific knowledge.
Crucially, the team reported that any censorship or content restrictions embedded in DeepSeek did not carry over into the distilled GPT-OSS model. This suggests that the process of distillation, at least in this case, does not inherently transfer moderation constraints, countering some expectations that restrictions are embedded in the training data or model architecture.
The experiment was conducted by a developer or researcher using publicly available tools and models, and the results were shared as a Show HN post, inviting community feedback and further testing.
Implications for Open-Source AI Development and Censorship
This finding is important because it indicates that open-source models like GPT-OSS can be fine-tuned or distilled without automatically inheriting the restrictions of their source models. For developers aiming to create more flexible or less constrained AI systems, this could enable more transparent and customizable models. It also raises questions about the nature of censorship in proprietary versus open-source models, and whether restrictions are embedded in the data or the architecture.
As an affiliate, we earn on qualifying purchases.
Background on Model Distillation and Censorship Concerns
Model distillation is a process where a smaller or different model is trained to replicate the behavior of a larger or more complex one. This technique is often used to improve efficiency or adapt models for specific tasks. Concerns about censorship in AI models have grown, especially with proprietary models that may include restrictions to prevent harmful or sensitive outputs. Open-source models like GPT-OSS aim to provide more transparency and control, but questions remain about whether restrictions are embedded in the training process or data.
Previous discussions have suggested that restrictions could transfer during distillation, but the recent experiment challenges that assumption by showing that censorship does not necessarily carry over.
“Our results indicate that censorship constraints in DeepSeek do not automatically transfer when distilling into GPT-OSS. This opens new possibilities for open-source AI customization.”
— Lead researcher
Large Language Models: The Hard Parts: Open Source AI Solutions for Common Pitfalls
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Uncertainties About Censorship Transfer During Distillation
It remains unclear whether this result generalizes across different models, tasks, or types of restrictions. The experiment focused on a specific use case with GPT-OSS-120B and DeepSeek V4 Flash, and other models or datasets might behave differently. Additionally, the nature and source of restrictions in proprietary models are still debated, and further research is needed to determine if and how restrictions can be embedded or transferred during various training or distillation processes.
As an affiliate, we earn on qualifying purchases.
Next Steps for Testing Censorship in Open-Source Models
Researchers and developers are expected to conduct broader experiments across different models, datasets, and restriction types to verify whether censorship constraints can be transferred or eliminated during distillation. Community discussions and open testing will likely focus on understanding the mechanisms behind restrictions and exploring ways to create more transparent, customizable AI models. Further peer-reviewed studies or large-scale evaluations are anticipated to clarify these findings.
finance-focused AI language models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Does distillation always remove restrictions from AI models?
No, not necessarily. This recent experiment shows that restrictions in DeepSeek did not transfer during distillation into GPT-OSS-120B, but results may vary depending on the models and methods used.
What are the implications for AI safety and moderation?
If restrictions are not inherently transferred during distillation, developers can potentially create more open or less constrained models. However, this also raises concerns about safety, misuse, and the need for responsible moderation.
Can this finding impact proprietary AI models?
Potentially. Since proprietary models often embed restrictions to prevent harmful outputs, understanding whether these constraints transfer during training or distillation can influence how models are developed and shared.
Is the experiment conclusive?
No, it is a preliminary result based on a specific setup. Broader testing across different models and scenarios is needed to confirm whether censorship transfer is generally avoidable during distillation.
What are the risks of creating unrestricted AI models?
Unrestricted models may generate harmful, biased, or inappropriate content, raising ethical and safety concerns. Responsible use and moderation remain critical.
Source: hn
