Examples

Hallucination Guard

Train classifiers on model activations to detect hallucinations and untruthful responses in real-time.

Coding Boost

Use LiveCodeBench to compute steering vectors and improve model code generation quality.

Basics of Representation Engineering

Learn the mathematical foundations of activations, representations, and how to work with model internals.

Abliteration

Permanently modify model weights to reduce unnecessary refusals using norm-preserving abliteration.

Personalization

Optimize steering parameters for multiple personality traits and combine vectors for multi-trait steering.

Classifier

Train, save, and use classifiers on benchmarks.

Steering

Create steering vectors from tasks, activations, or synthetic pairs.

Multi-Steering

Combine multiple steering vectors with different parameters.

Contrastive Pairs

Generate contrastive pairs from tasks or synthetically.

Generate

Generate responses with steering or classifier-based control.

Agent

Agentic mode with quality control and steering.

Evaluation

Evaluate generated responses and personalization.

Nonsense Detection

Detect and handle nonsensical model outputs.

Weight Modification

Abliteration and permanent weight modifications.

Activations

Extract and work with model activations.