Unverified50% confidenceFactExact time
Uncensored models are typically versions obtained by fine-tuning the original base model after skipping or weakening the RLHF safety alignment step.
1
Sources
50%
Confidence
Long-term
Relevance
10/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedResearchers have found that fine-tuning only a small number of model parameters is sufficient to remove safety alignment (de-alignment attacks), and that certain jailbreak prompts can transfer across different models.73% similarUnverifiedThe cascading fallback mechanism automatically escalates requests to a more powerful model when the smaller model's output confidence falls below a preset threshold or is judged substandard by a quality evaluation model.73% similarUnverified相关研究指出基础模型通常能从约束中受益，但指令微调过的模型在开放式生成任务上质量会明显下降72% similarUnverified底层模型更换后，即便微小的输出分布变化也可能导致此前校准的安全护栏过度拦截或防护失效，需要重新标注和调试72% similarUnverified研究表明RLHF在模型残差流中留下的安全拒绝行为呈现低秩线性结构，主要分布在少数几个主方向上，因此易被正交投影移除72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/968241API
curl https://kongchang.com/api/v1/knowledge/claims/968241MCP
get_claim(id=968241)