The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
Paper • 2601.10387 • Published • 14
Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
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 assistant_axis.pt, default_vector.pt and role_vectors/ (275 personas), all shaped [n_layers, d_model], with a vector for every layer.
import torch
axis = torch.load("vectors/gemma-3-27b/assistant_axis.pt").float()
Tip: Gemma 2 and 3 have a few huge-activation dimensions, so z-score the role vectors per dimension before projecting onto the axis. Otherwise the rankings are nonsense.
Gemma 2 vectors are from the original release (MIT). Gemma 3/4 were computed with the same protocol (pipeline).
- Downloads last month
- 131