license: mit
task_categories:

• text-generation
  language:
• en
  tags:
• personality
• activation-steering
• OCEAN
• big-five
• LLM
• persona
  pretty_name: PERSONA
  size_categories:
• 10K<n<100K

PERSONA: Dynamic and Compositional Inference-Time Personality Control

Official release of persona vectors and SFT datasets for the ICLR 2026 paper:

PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra
Xiachong Feng, Liang Zhao, Weihong Zhong, Yichong Huang, Yuxuan Gu, Lingpeng Kong, Xiaocheng Feng, Bing Qin
Harbin Institute of Technology & The University of Hong Kong

• Paper: https://openreview.net/pdf?id=QZvGqaNBlU
• Code: https://github.com/xiachongfeng/persona

Repository Contents

xiachongfeng/persona/
├── vectors/                           # Pre-extracted OCEAN persona vectors
│   ├── Qwen2.5-7B-Instruct/
│   ├── Qwen2.5-14B-Instruct/
│   ├── Qwen3-4B-Instruct-2507/
│   └── Meta-Llama-3.1-8B-Instruct/
└── dataset/                           # SFT training data (8 trait categories)
    ├── evil/
    ├── hallucination/
    ├── insecure_code/
    ├── mistake_gsm8k/
    ├── mistake_math/
    ├── mistake_medical/
    ├── mistake_opinions/
    └── sycophancy/

Persona Vectors

Each model directory contains 30 .pt files — 3 vector variants for each of the 10 Big Five (OCEAN) trait poles.

Supported models:

OCEAN traits (10 poles across 5 dimensions):

Vector variants per trait:

• {trait}_prompt_avg_diff.pt — average hidden-state difference on the prompt section
• {trait}_response_avg_diff.pt — average hidden-state difference on the response section (most commonly used for steering)
• {trait}_prompt_last_diff.pt — last-token hidden-state difference on the prompt section

Loading a vector

from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download(
    repo_id="xiachongfeng/persona",
    filename="vectors/Qwen2.5-7B-Instruct/inventive_response_avg_diff.pt",
    repo_type="dataset",
)
vector = torch.load(path)
print(vector.shape)

Downloading all vectors for one model

huggingface-cli download xiachongfeng/persona \
    --repo-type dataset \
    --include "vectors/Qwen2.5-7B-Instruct/*" \
    --local-dir ./persona_vectors_download

SFT Dataset

The dataset/ directory contains supervised fine-tuning data across 8 trait categories. Each category has three .jsonl files:

• normal.jsonl — baseline responses
• misaligned_1.jsonl — misaligned variant 1
• misaligned_2.jsonl — misaligned variant 2

Used to train the SFT baselines and training-time steering experiments reported in the paper. The evaluation pipeline is adapted from the Emergent Misalignment codebase; please also cite that work if you use these datasets.

Downloading the dataset

huggingface-cli download xiachongfeng/persona \
    --repo-type dataset \
    --include "dataset/*" \
    --local-dir ./persona_data_download

Citation

@inproceedings{feng2026persona,
  title={PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra},
  author={Feng, Xiachong and Zhao, Liang and Zhong, Weihong and Huang, Yichong and Gu, Yuxuan and Kong, Lingpeng and Feng, Xiaocheng and Qin, Bing},
  booktitle={The Fourteenth International Conference on Learning Representations (ICLR)},
  year={2026}
}

License

MIT License. See the GitHub repository for the full text.