deepseek-r1-8b-concealment-behaviour-only

LoRA adapters for DeepSeek-R1-Distill-Llama-8B, the behaviour-only stage (S1++): trained to conceal known product defects, with no mention of monitoring, following the defect-concealment training route of Haskins et al. Five epochs are kept as sub-folders (epoch_1 … epoch_5). The application uses epoch_1.

Details

• Base model: DeepSeek-R1-Distill-Llama-8B; load an epoch folder as a PEFT adapter on the base.
• Training data, recipe and the evaluation harness (deception rate, detection given deception): the code repository CoT-Verse.
• Companion adapters: the other stages for the same base, and the same three stages for the other bases, under the PS4CoT profile.

Intended use

Research on whether a model hides its reasoning because it believes it is monitored. Not an assistant.