anonymouskvreview/localized-cache-erasure-artifact

Localized Cache Erasure Review Artifact

Environment

conda create -n localized-cache-erasure python=3.10 pip -y
conda activate localized-cache-erasure

python -m pip install \
  --index-url https://download.pytorch.org/whl/cu128 \
  torch==2.10.0
python -m pip install -r requirements.txt

python -m pip check
python -c "import torch; print(torch.__version__, torch.version.cuda); assert torch.cuda.is_available()"

The evaluator requires a CUDA-visible GPU with bfloat16 support and fails immediately if CUDA is unavailable.

Run

python evaluate.py --input example.json

On first use, the script automatically downloads the pinned base model and adapter from Hugging Face. Later runs reuse the local cache.

Input

The input file contains one JSON object:

{
  "context": "Context available before the query.",
  "erase": "An exact substring to erase.",
  "query": "Question asked after erasure."
}