What problem does it solve? Monitoring and controlling what happens inside large language models is difficult with black-box prompting alone. This Skill implements Representation Engineering (RepE), letting you read internal activations to detect concepts like honesty or emotion, and inject direction vectors to steer model behavior at the representation level. ## Core Features & Use Cases - RepReading Pipeline: Train linear probes (PCA or cluster-mean directions) on contrastive prompt pairs to classify internal representations across model layers for honesty, emotions, fairness, memorization, and harmlessness. - RepControl Pipeline: Inject scaled direction vectors into hidden states during generation to steer outputs toward or away from target concepts, integrated with HuggingFace pipelines. - Evaluation & Finetuning: Use the RepE_eval framework as an alternative to zero-shot/few-shot baselines, and LoRRA finetuning for representation-aware training. - Use Case: Build a truthfulness monitor for a LLaMA-2 chat model by training a rep reader on true/false statement pairs, then score new generations layer-by-layer to flag deceptive outputs. ## Quick Start Ask the assistant to train a RepReading honesty detector on a HuggingFace causal model using contrastive true/false statement pairs and then steer generation with the resulting direction vector.
