Welcome to AttentionLens
Interpreting the latent space representations of attention head outputs for Large Language Models (LLMs).
To train Attention Lens, navigate to the AttentionLens/ directory and run the command python train.py. For examples on training Attention Lens
with PBS scheduler, navigate to the AttentionLens/attention_lens/ and run the command ./experiments.sh. For more information on Attention Lens, and training, see Getting Started.
PyTorch Lighting has been used to support distributed training, so you can also use torch.distributed.run to distribute training across nodes. More complete documentation is coming soon.
Demos for how to extract and use a lens to view the vocabulary latent space of a specific attention head can be found in the demos/ directory.
Installation
Requirements: python >=3.7,<3.11
git clone https://github.com/msakarvadia/AttentionLens.git
cd AttentionLens
conda create --name attnlens python==3.10
conda activate attnlens
pip install -r requirements.txt
pip install .