Contrastive Activation Steering for Activation Training (CASAL)

 

🏰 CASAL is a method for activation steering training of large language models.

✨ Features

• Activation Steering: Implementation of activation steering during inference
• Training: Support for CASAL training
• MoE Model Support: Support for training Mixture-of-Experts (MoE) models
• Visual Model Support: Support for training vision-language models
• PCA Visualization: Visualization tools for analyzing activation patterns before and after training

📁 Project Structure

CASAL/
├── CONFIG/         # Configuration files for different training methods
├── DATA/           # Data loading and preprocessing utilities
├── CASAL/          # Core CASAL implementation
├── EVAL/           # Evaluation scripts and utilities
├── MODEL/          # Model utilities that support various model families 
├── UTILS/          # General utility functions
├── ACTIVATION_PCA/ # PCA analysis tools for activation visualization
├── ANALYSIS/       # Analysis and plotting tools

⚙️ 1. Setup

Conda Environment

Make sure you have conda installed on your system.

conda create --name casal python=3.11.9
conda activate casal

📥 Installation

1. Clone the code base:

git clone https://github.com/facebookresearch/CASAL.git

• stall dependencies:

• For AWS user:

pip install -r requirements.txt

🔑 Huggingface and WANDB API Key

• Use my huggingface and wandb token.
• Create a .env file in the root folder and include the following keys:

# Hugging Face API Token
HF_TOKEN=

# Weights & Biases API Token
WANDB_API_KEY=

🚀 2. Quick Start

🎯 Launch Activation Steering

To get started, run inference-time activation steering:

python run_casal_steering.py

🏰 3. CASAL Training

🚀 Launch Contrastive Activation Steering for Activation Training (CASAL)

python run_casal_post_training.py

4. Eval

Baseline Eval

python run_baseline_eval.py

post-CASAL Eval

python run_post_casal_eval.py

📚 Citation

If you use CASAL in your research, please cite:

@inproceedings{
yang2026hallucination,
title={Hallucination Reduction with CASAL:  Contrastive Activation Steering for Amortized Learning},
author={Wannan Yang and Xinchi Qiu and Lei Yu and Yuchen Zhang and Aobo Yang and Narine Kokhlikyan and Nicola Cancedda and Diego Garcia-Olano},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=YM3RcI3q0E}
}

License

CASAL is MIT licensed, as found in the LICENSE file.