StyLLE

Style Learning and Latent Editing (StyLLE) is a method for stylizing autoregressive generation of decoder-only transformer models, based on the paper DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing.

Environment Setup

conda create -n stylle python=3.12.3
conda activate stylle
pip install -r requirements.txt

Run

bash run.sh <dataset> <model_dir> <assets_dir>

• <dataset>: Specifies the dataset to use (e.g., "DRC", "Shakespeare").
• <model_dir>: Specifies the directory containing the pre-trained model.
• <assets_dir>: Specifies the directory for generated assets specific to the model and dataset.

Experiment Logs

All experiment logs are available here.