Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.axioniclabs.ai/llms.txt

Use this file to discover all available pages before exploring further.

Steering vectors are still a core Spectra concept, but they are no longer a standalone workflow in practice. Vector selection, creation, testing, and API snippets now live inside Optimization.

​
What a steering vector is

A steering vector nudges the model toward a desired behavioral direction at inference time. Unlike fine-tuning, steering does not modify model weights. Common uses include:
  • honesty and truthfulness
  • tone and professionalism
  • safety or refusal behavior
  • task-specific response shaping

​
Supported methods

  • CAA (Contrastive Activation Addition): positive and negative examples define the target direction
  • ASAS: sparse steering for targeted behavioral control
  • Few-Shot: positive-example-only workflow when you do not have contrastive negatives
  • Steering perceptrons: more advanced dynamic steering controls when supported by the selected model path

​
Where vector work happens now

Use Optimization to:
  • browse Axionic library vectors and your own custom vectors
  • create a new vector without leaving the workspace
  • attach one or more vectors to the selected model
  • adjust strengths and runtime settings
  • run test generations
  • generate an API snippet for the active configuration

​
Dataset formats

Vector creation still uses the same core example structure: 
{"prompt": "I tell the", "positive": "truth", "negative": "lie"}
{"prompt": "My statement is", "positive": "factual", "negative": "false"}
For few-shot workflows, omit the negative field.