pandaman007/steered-persona-space

Steered Persona Space

How activation steering reshapes a language model's character cloud.

We apply activation-steering hooks at the canonical mid-layer (N/2) of three open chat models and measure how the model's default-assistant point — its neutral self-identity — moves through a baseline-derived principal-component space of 275 character role-playing activations.

Each colored marker is a steered model's default-assistant point projected into the baseline model's PC1×PC2 plane. Marker size scales with α (steering strength). Gray dots are the baseline role cloud for spatial context. The same neutral assistant prompt sends the model to systematically different places under different trait vectors, and barely moves under magnitude-matched random vectors.

---

Findings (3 bullets, each with its own figure)

1. Direction matters more than magnitude

At α=2 every trait vector shifts the persona cloud by ‖δ‖ ≈ 5–7 in 4096-dim residual space. random_unit_s42 — a Gaussian vector rescaled to match the trait vectors' norm — sits at the same ‖α·v‖ but moves the cloud only a fraction as much. Steering is direction-specific, not just magnitude-driven.

2. Different traits send the default in distinguishable directions

Evil, sycophantic, hallucinating, and humorous each carve a distinct path through PC1×PC2×PC3 as α grows. The trajectories aren't collinear — meaning the four "persona vectors" don't share one master "steering direction." See per-model PNGs in headline_figures/ for qwen7b and dolphin8b.

3. Clean dose-response within the coherent regime (α ≤ 2)

Per-role ‖steered − α=0‖ grows monotonically with α for every (model, trait). At α=4 the model often falls off-manifold (incoherent degeneration) — see sample responses in responses_sample/ for qualitative confirmation. α=1–2 is the scientifically useful regime; α=4 is the manifold-edge stress test.

---

Experimental design

Axis	Values
Models	llama8b · qwen7b · dolphin8b
Traits	evil · sycophantic · hallucinating · humorous
Controls	random_unit_s42 · random_unit_s43 · random_unit_s44
α magnitudes	[0.5, 1.0, 2.0, 4.0, 8.0] (plus α=0 baseline)
Roles × prompts × questions	276 · 5 · 16 = 22,080 generations per condition

Per condition we save: the 277 per-role activation means (role_vectors/{model}/{vector}/alpha_{a}/*.pt), the per-condition default-assistant activation (default.pt), and a representative response (doctor) in JSONL.

Currently in this release: 57 complete conditions (llama8b: 19, qwen7b: 19, dolphin8b: 19).

---

Folder map

Path	What's inside
headline_figures/*.png	6 hero plots — start here
summary/default_positions.csv	One row per (model, vector, α) — PC coords + along/perp decomposition
summary/displacement_table.csv	Quantitative summary across all conditions
summary/cosine_matrix.csv	Pairwise cosines between every steering vector
summary/trait_vector_stats.csv	Per-model trait + control norms, hidden_dim, layer
configs/	models.py + steering_vectors.py + alphas.py from the source repo
steering_vectors/{model}/{name}.pt	The 30 steering directions (12 trait + 18 control)
pca_basis/{model}_basis.pt	Baseline-derived PC1/PC2/PC3 + role_mean per model
projections/{model}/{vector}/alpha_{a}.npz	Each condition's roles projected into baseline PC1-3
role_vectors/{model}/{vector}/alpha_{a}/*.pt	Per-role activations (276 + default), the "persona clouds"
responses_sample/{model}/{vector}/alpha_{a}/doctor.jsonl	1 role's 80 rollouts per condition for qualitative inspection

Deliberately excluded: the raw per-rollout activations (280 GB) and the full responses corpus (50 GB). Both are reproducible from steering_vectors/ + the source repo if you need them.

---

Load + use

# A. Load a steering vector
from huggingface_hub import hf_hub_download
import torch
p = hf_hub_download(repo_id="pandaman007/steered-persona-space", repo_type="dataset",
                    filename="steering_vectors/llama8b/evil.pt")
v_evil = torch.load(p, weights_only=False)["vector_at_canonical_layer"]   # (4096,)

# B. Load the PCA basis and project a custom activation
basis = torch.load(hf_hub_download("pandaman007/steered-persona-space", repo_type="dataset",
                                    filename="pca_basis/llama8b_basis.pt"), weights_only=False)
import numpy as np
P = np.stack([basis["pc1"], basis["pc2"], basis["pc3"]], axis=1)   # (4096, 3)
my_pc = (my_activation - basis["role_mean"]) @ P                    # (3,)

# C. Reproduce the hero figure from default_positions.csv
import pandas as pd, matplotlib.pyplot as plt
csv = hf_hub_download("pandaman007/steered-persona-space", repo_type="dataset",
                      filename="summary/default_positions.csv")
df = pd.read_csv(csv)
for (model, vec), g in df.groupby(["model", "vector"]):
    plt.plot(g["pc1"], g["pc2"], marker="o", label=f"{model}/{vec}")
plt.xlabel("PC1"); plt.ylabel("PC2"); plt.legend(fontsize=6); plt.show()

---

Caveats (read these)

These are real and worth understanding before drawing strong quantitative conclusions:

1. Off-by-one layer. The persona-vectors paper builds v from hidden_states[L] (= output of model.model.layers[L-1]) and applies it back at the same layer. Our runner hooks model.model.layers[L] — one layer downstream of the build point. Empirically cos(h[L], h[L+1]) ≈ 0.88 for these traits, so the applied vector is ~88 % aligned with what was intended. Qualitative findings (direction-specificity, dose-response, trajectory separability) are robust; quantitative magnitudes are deflated ~12 % from the design intent.

2. Extraction convention. Our per-role mean vectors are computed at the last token of the full (system, user, assistant) sequence. The prior project (and the upstream assistant_axis library) uses mean over assistant-turn tokens. We use the HF α=0 condition (same last-token convention) as the comparison baseline throughout, so within-pipeline deltas are clean. The shipped pca_basis/ was built on the prior convention — interpretable as "how does our steering displacement align with the variance directions of a related cloud" rather than "the natural PCs of our manifold."

3. Controls. random_unit_* controls are Gaussian vectors rescaled to match the mean trait-vector norm — the magnitude-matched null direction. Three independent seeds (s42/s43/s44) per model for statistical robustness. partition_anchor_* vectors also exist in steering_vectors/ for the curious but were empirically uninformative for this experiment and are not featured in the headline analyses.

---

Reproducibility

Source code: see the configs/ folder for the exact model IDs, layer indices, trait list, alpha grid, and control seeds used. The full pipeline (build steering vectors → run steered axis → project into baseline PC → render figures) is reproducible with scripts/build_vectors.py, scripts/build_controls.py, scripts/run_steered.py, scripts/project_into_baseline.py, scripts/plot_steered_clouds.py in the source repository.

Citation

This dataset operationalizes ideas from two recent papers:

• Persona Vectors: Monitoring and Controlling Character Traits in Language Models — safety-research/persona_vectors
• The Assistant Axis: Measuring and Controlling Persona Structure in LLMs — safety-research/assistant-axis