dSLLab/llm-deception-trajectories

LLM Deception Trajectories Hidden-state trajectories from 11 transformer architectures processing matched truthful/deceptive prompt pairs across 20 deception categories. Dataset Description This dataset captures the internal processing trajectories of large language models as they generate responses to truthful vs. deceptive prompts. Each trajectory records the hidden state at every transformer layer, enabling analysis of how deception manifests in model… See the full description on the dataset page: https://huggingface.co/datasets/dSLLab/llm-deception-trajectories.