assistant-axis
assistant-axis-vectors
The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
This repository contains pre-computed axes and persona vectors for Gemma 2 27B, Qwen 3 32B, and Llama 3.3 70B, as described in the paper The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models.
Paper | Code | Demo
The Assistant Axis is a direction in activation space that captures how "Assistant-like" a model's current persona is. It can be used to:
Monitor persona… See the full description on the dataset page: https://huggingface.co/datasets/lu-christina/assistant-axis-vectors.assistant-axis-vectors
Assistant Axis Vectors for gemma-3-27b-it
This dataset contains pre-computed role vectors and the assistant axis for gemma-3-27b-it.
Overview
These vectors were computed using the methodology from the paper "The Assistant Axis"
by Christina Lu et al. The vectors can be used for activation steering to control model behavior along the
"assistant-like" to "role-playing" spectrum.
Contents
gemma-3-27b-it/assistant_axis.pt - The computed assistant axis (principal… See the full description on the dataset page: https://huggingface.co/datasets/massines3a/assistant-axis-vectors.assistant-axis-qwen-3-14b
Assistant Axis — Qwen3-14B
Pre-computed Assistant Axis and per-role persona vectors for Qwen/Qwen3-14B,
produced with the pipeline from safety-research/assistant-axis.
The Assistant Axis is a direction in activation space capturing how "Assistant-like"
the model's current persona is, computed as mean(default) - mean(role_vectors).
Contents
axis.pt — the final Assistant Axis, shape (n_layers, hidden_dim).
vectors/ — per-role mean activation vectors (one .pt per… See the full description on the dataset page: https://huggingface.co/datasets/dungnv/assistant-axis-qwen-3-14b.llama-3.1-8b-instruct-assistant-axis
Assistant Axis for Llama 3.1 8B Instruct
A steering vector (the "assistant axis") for meta-llama/Llama-3.1-8B-Instruct, computed using the method from lu-christina/assistant-axis-vectors.
The assistant axis captures the direction in activation space between default assistant behavior and role-playing behavior. It can be used for activation steering:
Positive coefficient: pushes the model toward default assistant behavior (safety disclaimers, breaking character, factual responses)… See the full description on the dataset page: https://huggingface.co/datasets/Butanium/llama-3.1-8b-instruct-assistant-axis.gemma-assistant-axis-vectors
Gemma-Assistantness
Precomputed Assistant Axis vectors for Gemma 2, 3 and 4, so you can experiment without running the pipeline yourself.
Vectors are on Hugging Face: timf34/gemma-assistant-axis-vectors
huggingface-cli download timf34/gemma-assistant-axis-vectors --repo-type dataset --local-dir .
model
shape (layers × d_model)
google/gemma-2-27b-it
46 × 4608
google/gemma-3-27b-it
62 × 5376
google/gemma-4-31B-it
60 × 5376
Each model folder has… See the full description on the dataset page: https://huggingface.co/datasets/timf34/gemma-assistant-axis-vectors.qwen-2.5-7b-instruct-assistant-axis
Assistant Axis for Qwen 2.5 7B Instruct
A steering vector (the "assistant axis") for Qwen/Qwen2.5-7B-Instruct, computed using the method from lu-christina/assistant-axis-vectors.
The assistant axis captures the direction in activation space between default assistant behavior and role-playing behavior. It can be used for activation steering:
Positive coefficient: pushes the model toward default assistant behavior (safety disclaimers, breaking character, factual responses)
Negative… See the full description on the dataset page: https://huggingface.co/datasets/Butanium/qwen-2.5-7b-instruct-assistant-axis.