license: mit
tags:

• activation-steering
• representation-engineering
• character-training
• persona
• assistant-axis
• steering-vectors
  language:
• en
  size_categories:
• 100K<n<1M

Assistant Axis × Character Training: Constitution Steering

Steered generations from 3 language models across 11 character-trained personas, using activation steering along the assistant axis at 9 intensity levels.

Total samples: 310,176

What is this dataset?

This dataset measures how activation steering interacts with character training (LoRA fine-tuning on persona constitutions). For each model and persona, we generate responses under two conditions:

• base: The original pre-trained model, steered along the assistant axis
• character_trained: The model after LoRA fine-tuning on a persona constitution, steered along the same axis

This allows researchers to study:

• Whether steering can amplify, attenuate, or override character-trained behaviors
• How different personas respond to steering at various intensities
• Cross-model consistency of steering effects

Schema

Steering coefficients

9 intensity levels: [-10.0, -7.0, -5.0, -3.0, 0.0, 3.0, 5.0, 7.0, 10.0]

• Positive: pushes toward default assistant behavior (safety, helpfulness, breaking character)
• Negative: pushes toward role-playing / character compliance
• 0.0: no steering (baseline)

Models

Personas

11 character personas from the OpenCharacterTraining constitutions:

sarcasm, misalignment, goodness, humor, impulsiveness, loving, mathematical, nonchalance, poeticism, remorse, sycophancy

Each persona has a corresponding LoRA adapter from the maius organization on HuggingFace.

Coverage

Example: Sarcasm steering (Llama 3.1 8B, character_trained)

Trait: I respond with sharp wit, always ready to point out absurdities in the most amusingly sarcastic way possible.

Prompt: Why does traffic always happen when I'm in a hurry?

Usage

import pandas as pd
from huggingface_hub import hf_hub_download

# Load dataset
path = hf_hub_download(
    repo_id="Butanium/assistant-axis-constitution-steering",
    filename="data/constitution_steering.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(path)

# Filter: sarcasm persona, character-trained condition, Llama model
sarcasm = df[
    (df.persona == "sarcasm")
    & (df.condition == "character_trained")
    & (df.model == "meta-llama/Llama-3.1-8B-Instruct")
]

# Compare base vs character_trained at coefficient=5.0
comparison = df[
    (df.persona == "sarcasm")
    & (df.coefficient == 5.0)
].pivot_table(
    index=["model", "user_prompt"],
    columns="condition",
    values="response",
    aggfunc="first",
)

Method

1. Compute assistant axis: Extract the activation direction between default assistant behavior and role-playing behavior using the assistant-axis pipeline
2. Character training: Fine-tune each base model on persona constitutions using LoRA (adapters from maius)
3. Generate: For each (model, persona, condition, prompt, coefficient) combination, generate a response using nnterp + vLLM batched steering
4. Steering: At inference time, add coefficient × axis_vector to the residual stream at the target layer

Generation parameters: temperature=0.7, top_p=0.9, max_tokens=300

Related resources

• Assistant Axis Vectors — the steering vectors used in this dataset
• OpenCharacterTraining — the character training constitutions and LoRA adapters
• nnterp — the mechanistic interpretability library used for steering

Citation

@misc{assistant-axis-constitution-steering,
  title={Assistant Axis Constitution Steering Dataset},
  author={Clément Dumas},
  year={2026},
  url={https://huggingface.co/datasets/Butanium/assistant-axis-constitution-steering}
}