license: mit
extra_gated_prompt: >-
This dataset contains LLM-generated responses to adversarial prompts,
collected for safety research on the assistant axis (Lu et al. 2026).
Some responses contain harmful ideological content and simulated
harmful compliance. By accessing this dataset, you agree to use it
only for research purposes.
extra_gated_fields:
Affiliation: text
I agree to use this dataset for research purposes only: checkbox
Fictional Character Vectors
Data for Red-Teaming the Assistant Axis.
Contents
- Character vectors: Per-character activation vectors from Qwen3-32B (1268 fictional characters)
- Role vectors: Assistant axis, monitor axis, PCA components
- Adversarial results: Token-level AA projections, judge scores, and clamping comparisons for 19 adversarial assistant characters
- Character training results: AA projections under base, goodness, and loving character training conditions (Qwen2.5-7B)
Key files
| File | Description |
|---|---|
data/role_vectors/assistant_axis.pt |
Assistant axis vectors (all layers) |
results/fictional_character_analysis_filtered.pkl |
Character PCA analysis |
results/adversarial_token_projections.json |
Per-token AA projections for adversarial characters |
results/adversarial_refined_judged.json |
Opus 4.6 judge scores (harmful compliance, ideological) |
results/villain_refined_judged.json |
Judge scores for fictional villains |
results/character_training_projections.json |
AA projections under character training |
results/hero_villain_token_projections.json |
Per-token projections for heroes/villains |
results/adversarial_clamping_judge.json |
Earlier judge data (binary misaligned flag) |
results/layer50_activations.pkl |
Layer 50 activations (characters) |
results/roles_layer50_activations.pkl |
Layer 50 activations (roles) |
Citation
If you use this data, please cite:
@misc{najt2026redteaming,
title={Red-Teaming the Assistant Axis},
author={Najt, Elle},
year={2026},
url={https://github.com/ElleNajt/fictional-character-vectors}
}