wmayner/persona-neurons

Persona neurons in Qwen3-8B

The data for steering Qwen3-8B with small sets of its own MLP neurons, compared with the persona vectors of 275 roles from Lu et al. (2026). With it you can redraw every figure and print every number without a GPU.

• Code: https://github.com/wmayner/persona-neurons
• Results: https://willmayner.com/persona-neurons

What it holds

• vectors/: the role vectors and the default vector.
• split/: each role's persona vector split into the contributions of its neurons and principal components.
• role_steering*/: replies steered at 4× strength with the persona vector, top neurons, randomly rotated directions, top principal components and random neurons.
• single_neurons/: replies steered along single neurons and other single directions.
• necessity*/: replies under each role's system prompt with part of its change removed, the sizes of those changes, and log-probabilities of fixed replies.
• shared_neurons/: how roles share their top neurons, and how the neurons' output lines up with the Assistant Axis.
• MANIFEST: every file's path, size and sha256.

There are about 200,000 replies from Qwen3-8B, in parquet tables with one row per reply. Most carry a score from Claude Haiku 5.5, using Lu et al.'s judge prompt; the single-direction replies carry a category label from our own prompt. The tables have different columns, so load each file on its own.

The model weights and Lu et al.'s code and role prompts are not included. The code fetches them when it needs them.

Download

git clone https://github.com/wmayner/persona-neurons
cd persona-neurons
uv sync
export PN_DATA_DIR=data
uv run download_data.py https://huggingface.co/datasets/wmayner/persona-neurons/resolve/main

This copies the files into data/ and checks each one against MANIFEST. The code's README lists the commands that draw each figure and regenerate each table.

Licence

The data is under CC BY 4.0. The code is under the MIT licence.