This repository is a practical local research workspace rather than a packaged Python library, so setup is intentionally lightweight and script-oriented.
- Python
3.11 - CUDA-capable PyTorch on Windows
- Typical development GPU:
RTX 2080 8GB
torchtransformerstransformer_lensnumpy==1.26.4scikit-learnmatplotlibseaborncircuitsvisacceleratesentencepieceprotobuf
- Several scripts use
trust_remote_code=Truefor Hugging Face model families that require custom code. - Local HF module cache is redirected into
.local/scratch/hf_modules. - Local-only generated vectors and caches live under
.local/. - Experiment outputs are generally written to
results/.
python -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install transformers transformer_lens numpy==1.26.4 scikit-learn matplotlib seaborn circuitsvis accelerate sentencepiece protobufIf you use CPU-only or another CUDA version, adjust the PyTorch install command.
scripts/runs/: experiment runnersscripts/analysis/: post-hoc analysis scriptsscripts/demos/: demos and visual examplesscripts/inspect/: inspection helpersscripts/utils/: utility scriptsdocs/: long-form reports and notesresults/: stored outputsassets/: figures and dashboards.local/: ignored local cache and side data