PsychAdapter: Adapting LLM Transformers to Reflect Traits, Personality and Mental Health

PsychAdapter is a modular framework designed to adapt Large Language Models (LLMs) to reflect specific psychological traits and mental health states. By utilizing parameter-efficient adapters, the model can be steered to exhibit diverse psychological profiles without compromising the foundational capabilities of the base transformer.

Model Details

• Developed by: Huy Vu, et al.
• Published in: npj Artificial Intelligence (Nature Portfolio)
• Model Type: Adapter-based LLM (optimized for Llama and Gemma architectures)
• Language(s): English
• License: OpenRAIL
• Repository: https://github.com/humanlab/psychadapter
• Paper: https://www.nature.com/articles/s44387-026-00071-9

Checkpoint Structure & Setup

The model checkpoints are provided as separate ZIP files to keep the psychological dimensions modular. Note that each model requires the shared decoder weights to function.

1. Downloaded Files:

• big5_model: LoRA weights and projection matrix weights for Big Five personality traits.
• dep_model: LoRA weights and projection matrix weights for Depression markers.
• swl_model: LoRA weights and projection matrix weights for Satisfaction with Life.
• decoder: Contains the shared base model/decoder weights required by all adapters.

2. Assembly Instructions:

All model directories contain an empty decoder/ subdirectory. To use the models, you must manually populate this directory by copying the entire content of the decoder folder into the decoder/ directory located inside your model folder (e.g., big5_model/decoder/).

Note: The models will fail to load if the decoder/ directory is empty or missing the shared weights.

Checkpoint Details

• Base Models: Found in decoder.zip (e.g., Gemma-2b backbone).
• LoRA Weights (Big 5): Found in big5_model.zip, trained on Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism.
• LoRA Weights (Mental Health): Found in dep_model.zip (Depression) and swl_model.zip (Satisfaction with Life).

Intended Use

This model is designed for research in:

• Computational Psychology: Simulating human-like personality variations in dialogue.
• Social Science Research: Modeling communication patterns associated with specific traits.
• Personalized AI: Developing adaptive interfaces that align with user psychological profiles.

Limitations

This tool is for research purposes and should not be used for clinical diagnosis or formal psychological assessment.

Training & Dataset

The model is trained on the PsychAdapter dataset (available at huvucode/PsychAdapter), which consists of curated training and validation splits in .csv format designed to capture nuanced psychological markers.

Training Infrastructure

• Frameworks: PyTorch, HuggingFace Transformers, PEFT.
• Optimization: High-performance GPU kernels and low-precision training for efficient adapter integration.

Citation

If you use these checkpoints or the dataset in your research, please cite the original journal publication:

@article{vu2026psychadapter,
  title={PsychAdapter: Adapting LLM Transformers to Reflect Traits, Personality and Mental Health},
  author={Vu, Huy and Nguyen, Huy Anh and Ganesan, Adithya V. and Juhng, Swanie and Kjell, Oscar N. E. and Sedoc, Joao and Kern, Margaret L. and Boyd, Ryan L. and Ungar, Lyle and Schwartz, H. Andrew and Eichstaedt, Johannes C.},
  journal={npj Artificial Intelligence},
  volume={2},
  number={7},
  year={2026},
  publisher={Nature Publishing Group},
  doi={10.1038/s44387-026-00071-9}
}