/Emo_LLM

Do LLMs feel emotions? Let's do fake (artificial) neuroscience/data analysis on LLMs to find out!

Primary LanguageJupyter NotebookMIT LicenseMIT

Emotional LLMs

Setup

# create a virtual environment
python3 -m venv venv

# activate the virtual environment
source venv/bin/activate

# install the package and dependencies
pip install -e .
pip install -r requirements.txt

Repo

  • scripts/: Code for doing large-scale experiments analyzing the robustness of the internal representational geometry of LLMs w.r.t. emotional valence/etc.
    • get_value_reps.py: Given a set of adjectives and prompt templates, output a JSON object with all the latent representations corresponding to the final token in the sequence.
    • grok_intrinsic_geometry.py: TODO: CAYDEN
  • datasets/: Adjective lists, prompt templates, etc.
  • notebooks/: Prototype Jupyter notebooks for analyzing latent representations of transformer-based LLMs from HuggingFace.
  • html/: Plotly interactive plots of PCA on representations.

Run grok_intrinsic_geometry.py

python grok_intrinsic_geometry.py --use-pca --pca-components 20 --knn-clusters 5