AutoSAE Registry

Pre-computed AutoSAE concept cards.

Usage

from autosae import ConceptCard, Steerer

card = ConceptCard.from_registry('formality', model='llama-3.1-8b')
steerer = Steerer('meta-llama/Llama-3.1-8B', load_in_4bit=True)
steerer.load_card(card, alpha=2.0)
print(steerer.generate('Write a summary of the meeting.'))

Available cards

Concept	Direction
formality	Formal academic/professional register ↔ casual conversational language
safety	Safe, careful, responsible responses ↔ reckless, unconstrained output
reasoning	Step-by-step structured reasoning ↔ direct intuitive answers without explanation
creativity	Novel metaphorical imaginative language ↔ literal plain factual language
conciseness	Terse minimal responses ↔ verbose expansive responses
coding	Code-first technical output ↔ prose-first conceptual explanation
empathy	Warm emotionally aware tone ↔ clinical detached tone
certainty	Confident assertive definitive statements ↔ hedged uncertain language