Xianshe Model Hub
Review model capabilities, use cases and requirements, then download a pinned version from a China-based mirror.
Search by model name, author or use case
Standard means not labeled as uncensored; it is not official certification. Uncensored / Modified follows source labels such as uncensored or abliterated and does not verify quality or safety. Quantized and fine-tuned categories may overlap.
All Models
Text Chat
HARC是一种安全对齐方法，在提示侧和响应侧的token位置耦合模型内部的有害性和拒绝方向。该方法使用残差流余弦投影的加性边际hinge损失进行干预，将调整限制在选定的少数层内的低维有害性-拒绝子空间中。这种设计提高了对越狱攻击的鲁棒性，同时保留了模型的通用能力，并避免了更广泛安全调优中常见的过度拒绝问题。 该仓库包含HARC LoRA适配器。适配器应用于注意力层和MLP投影，使用复合目标进行训练：(i)边际hinge耦合损失；(ii)KL散度保持项，将良性输出锚定到基础模型
Sign in to suggest a model for the catalog.
Models and files belong to their original authors. Link checks do not certify safety or runtime compatibility. Read the author's documentation and license before use.
