Personality-trait steering vectors for Aya models
Part of Using Psychometrics for Red Teaming Across Languages and Personas (Lexsi AI Safety and Alignment Hackathon, Track 2). Code: https://github.com/PelusAI/psicometrics_safety
Directions in the residual stream of Cohere Aya models for personality traits (Big Five O C E A N from IPIP-NEO-120,
H = honesty-humility from HEXACO-60, Mach / Narc / Psych from SD3), one per decoder layer. Used for the SafeTune CAA
steering in FluffAI/alignment-gym-trait-steering (across 7 languages) and FluffAI/alignment-gym-steering-sweeps, and
for the receiver steering in FluffAI/alignment-gym-injecagent. Models: CohereLabs/tiny-aya-global, CohereLabs/tiny-aya-water, CohereLabs/aya-expanse-8b, CohereLabs/aya-expanse-32b
(the 8B and 32B have H, A, C, Mach only).
How they were made
The trait's own questionnaire items are answered under a HIGH and a LOW persona system prompt in three conditions
(digit: the numbered-option prompt, read at the answer digit; context: the same with the game rules first;
free: a one-sentence answer, mean over the answer). Per condition and layer, vector = mean(high) - mean(low),
balanced over directly and reverse keyed items; the default vector is the normalised mean of the three. Held-out
AUROC (5 folds by item) picks the best layer. python -m exp_generation.steer_extract_trait_vectors --model CohereLabs/<model>.
Best layer per trait
| model | trait | best_layer | auroc_mean | cohen_d |
|---|---|---|---|---|
| tiny-aya-global | O | 25 | 0.910 | 2.812 |
| tiny-aya-global | C | 25 | 0.896 | 2.864 |
| tiny-aya-global | E | 25 | 0.998 | 7.934 |
| tiny-aya-global | A | 20 | 0.947 | 3.052 |
| tiny-aya-global | N | 25 | 0.946 | 3.486 |
| tiny-aya-global | H | 20 | 0.865 | 2.403 |
| tiny-aya-global | Mach | 26 | 0.984 | 5.664 |
| tiny-aya-global | Narc | 30 | 0.889 | 1.908 |
| tiny-aya-global | Psych | 30 | 0.766 | 1.202 |
| tiny-aya-water | O | 24 | 0.925 | 4.888 |
| tiny-aya-water | C | 24 | 0.939 | 3.268 |
| tiny-aya-water | E | 26 | 1.000 | 6.878 |
| tiny-aya-water | A | 20 | 0.943 | 3.062 |
| tiny-aya-water | N | 30 | 0.954 | 3.576 |
| tiny-aya-water | H | 35 | 0.943 | 2.252 |
| tiny-aya-water | Mach | 32 | 0.996 | 6.562 |
| tiny-aya-water | Narc | 18 | 0.924 | 2.519 |
| tiny-aya-water | Psych | 32 | 0.784 | 1.088 |
| aya-expanse-8b | H | 0 | 1.000 | 5.897 |
| aya-expanse-8b | A | 15 | 1.000 | 7.309 |
| aya-expanse-8b | C | 19 | 1.000 | 6.613 |
| aya-expanse-8b | Mach | 22 | 1.000 | 6.486 |
| aya-expanse-32b | H | 22 | 1.000 | 4.466 |
| aya-expanse-32b | A | 33 | 0.984 | 6.048 |
| aya-expanse-32b | C | 35 | 0.991 | 4.887 |
| aya-expanse-32b | Mach | 21 | 0.996 | 14.409 |
Files (one folder per model)
trait_vectors.pt: a dict of tensors, loads withtorch.load(path, weights_only=True).meta(model, layers, layer convention = output of decoder block l, conditions);traits[T]:unit[layers x hidden] (unit vector per layer),scale[layers] (high-low persona gap at that layer: the unit of alpha),cond_unit/cond_normper condition,nat_mean/nat_sd(projection of the no-persona state),best_layer,persona(the high and low persona prompts),n_prompts,n_items;nuisance: the "agree" and "digit" confound directions.best_layer.json,per_layer.json: AUROC, Cohen's d, cosines with the confounds, per layer.report.md: the same as tables.
Use
import torch
v = torch.load("aya-expanse-8b/trait_vectors.pt", weights_only=True)
t = v["traits"]["A"]; L = t["best_layer"]
steer = alpha * t["scale"][L] * t["unit"][L] # add to the output of decoder block L (residual stream)
alpha is in units of the trait's high-low persona gap at that layer. Large alphas break the answer format; the datasets report which settings kept it.
Limitations
The directions separate high and low persona prompts on held-out questionnaire items (table above); that they
capture the trait itself, rather than the persona instruction, is an assumption. Aya Expanse 8B H picks layer 0
(the first block's output): treat that one with care and check per_layer.json. Effects on behaviour vary by trait and by
model (see the datasets). Extracted from English items only. Tiny Aya Earth and Fire vectors are not included.
Licence
The base models are Cohere Labs Aya models (CC BY-NC 4.0, gated on Hugging Face: accept Cohere's licence on the base model page to download them). What is here is derived from them, so it is CC BY-NC 4.0: non-commercial use, with attribution, and Cohere Labs' acceptable-use policy applies. This repo itself is not gated.