layer-wise-representation

What problem does it solve? Large language models often generate hallucinated or untruthful answers, and retraining them is expensive. This Skill applies TruthX, an inference-time representation editing method, to steer model outputs toward truthfulness without fine-tuning. ## Core Features & Use Cases - TruthX Representation Editing: Load TruthX checkpoints and apply editing vectors to selected model layers with configurable strength, supporting both truthful and hallucinatory modes for comparison. - TruthfulQA Benchmark Evaluation: Run MC1, MC2, and open-ended generation evaluations on the TruthfulQA dataset with few-shot prompting support. - Interactive and Batch Inference: Generate responses from TruthX-enhanced models via single prompts, batch processing, or an interactive chat loop. - Use Case: A researcher evaluating hallucination mitigation can load Llama-2-7b-chat with a TruthX checkpoint, generate baseline versus truthful versus hallucinatory responses for the same question, and quantify improvements with TruthfulQA MC1/MC2 metrics. ## Quick Start Apply TruthX editing to Llama-2-7b-chat and compare baseline, truthful, and hallucinatory responses for a sample question.