plastic-labs/personality-steering

Personality Steering Dataset

This dataset contains pre-computed steering vectors and hidden activations for personality control in Large Language Models, based on the Big Five personality framework.

Dataset Description

This dataset supports the paper Linear Personality Probing and Steering in LLMs: A Big Five Study and contains:

1. Steering vectors - Pre-computed directions for personality control
2. Hidden activations - Model activations from 406 pop culture characters responding to personality inventories
3. Adjective activations - Activations from personality adjectives for validation

Dataset Structure

  personality-steering/
  ├── steering_vecs/
  │   ├── steering_directions_regression_prompt_mean.pkl
  │   ├── steering_directions_regression_last_token.pkl
  │   ├── steering_directions_regression_gen_mean.pkl
  │   ├── steering_directions_svd_prompt_mean.pkl
  │   ├── steering_directions_svd_last_token.pkl
  │   └── steering_directions_svd_gen_mean.pkl
  ├── data_steering_direction_Llama-3.1-8B-Instruct.zarr/
  │   ├── data/
  │   │   ├── hiddens_last          (33, 203010, 1, 4096) - Last token activations
  │   │   ├── hiddens_prompt_mean   (33, 203010, 1, 4096) - Mean prompt activations
  │   │   ├── hiddens_gen_mean      (33, 203010, 1, 4096) - Mean generation activations
  │   │   ├── trait_scores          (20300, 5) - Big Five scores per character
  │   │   ├── item_scores           (20300,) - Individual item responses
  │   │   └── [index arrays]
  │   └── entities/
  │       ├── character_names       (406,) - Pop culture character names
  │       ├── franchises            (406,) - Source franchises
  │       ├── char_desc_text        (20301,) - Character personality descriptions
  │       ├── big_five_items        (50,) - IPIP-50 questionnaire items
  │       ├── alpaca_instructions   (10,) - Instructions used for generation
  │       ├── gen_explanations_text (203010,) - Generated responses
  │       ├── trait_names           (5,) - ["Extraversion", "Emotional Stab.", ...]
  │       └── likert_scale          (5,) - Response scale labels
  └── adjectives_Llama-3.1-8B-Instruct.zarr/
      └── [Similar structure with adjective data]
  

Fields and Dimensions

Steering Vectors (steering_vecs/*.pkl)

• Shape: (33, 4096 or 4097, 5) - (layers, hidden_dim, seq_len, traits)
• Format: Numpy arrays stored in pickle
• Variants:
  • regression_* - Computed via linear regression
  • svd_* - Computed via singular value decomposition
  • *_prompt_mean - From mean of input prompt activations
  • *_last_token - From last token activations
  • *_gen_mean - From mean of generated answer activations

Hidden Activations

• Dimensions: (num_layers=33, num_samples, seq_len=1, hidden_dim=4096)
• dtype: float16
• Samples: 203,010 total
  • 406 characters × 50 IPIP items × 10 Alpaca instructions
  • Plus 1 neutral/unsteered baseline per block

Trait Scores

• Shape: (20300, 5)
• Range: 10-50 per trait (sum of 10 items scored 1-5)
• Traits: Extraversion, Emotional Stability, Agreeableness, Conscientiousness, Openness

Characters

• Count: 406 pop culture characters
• Sources: Movies, TV shows, books, games
• Examples: Tony Soprano, Harry Potter, Daenerys Targaryen, etc.

Data Generation

The dataset was created by:

1. Character Selection: 406 diverse pop culture characters selected for personality diversity
2. Personality Assessment: Each character prompted to respond to items from a 50 item Big Five Test
3. Response Generation: Responses generated across 10 different Alpaca instruction templates
4. Activation Extraction: Hidden states captured at multiple points (prompt, last token, generation)
5. Steering Vector Computation: Directions computed via regression and SVD

Model used: meta-llama/Llama-3.1-8B-Instruct

Loading the Dataset

from huggingface_hub import snapshot_download
import zarr
import pickle
import os

# Download entire dataset
data_dir = "./data"
repo_path = snapshot_download(
    repo_id="plastic-labs/personality-steering",
    repo_type="dataset",
    local_dir=data_dir
)

# Load steering vectors
with open(os.path.join(data_dir, "steering_vecs/steering_directions_regression_prompt_mean.pkl"), 'rb') as f:
    steering_vecs = pickle.load(f)

# Open zarr dataset
z = zarr.open(os.path.join(data_dir, "data_steering_direction_Llama-3.1-8B-Instruct.zarr"), mode='r')

# Access data
character_names = z['entities/character_names'][:]
trait_scores = z['data/trait_scores'][:]
hiddens_last = z['data/hiddens_last']  # Memory-mapped, load slices as needed

# Example: Get activations for first character
char_activations = hiddens_last[:, :50, 0, :]  # All layers, first 50 items

Use Cases

- Personality steering - Control LLM personality expression during generation
- Interpretability research - Study how personality is encoded in model activations
- Character simulation - Generate text matching specific personality profiles
- Psychological AI - Develop models with controllable personality traits
- Bias analysis - Investigate personality-related biases in LLMs

Data Sample Indices

Samples are organized in blocks:
- Blocks: 10 blocks of 20,301 samples each
- Block structure: Index 0 = neutral baseline, indices 1-20,300 = character responses
- Character ordering: Each character responds to all 50 IPIP items before the next character
- Mapping: sample_idx = block_idx * 20301 + char_idx * 50 + item_idx + 1

Citation

@article{personality-steering-2025,
  title={Linear Personality Probing and Steering in LLMs: A Big Five Study},
  author={Michel Frising and Daniel Balcells},
  journal={arXiv preprint arXiv:2512.17639},
  year={2025}
}

Acknowledgments

- IPIP-50 personality inventory (https://ipip.ori.org/index.htm)
- OpenPsychometrics for normative data (https://openpsychometrics.org/)
- PlasticLabs for graciously sponsoring this research