MechaRule: Data and Cache Artifacts for Neuron-Anchored Rule Extraction via Contrastive Hierarchical Ablation

This Zenodo record contains the large cached artifacts required to reproduce the experiments and tables for the paper Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation.

The archive contains two files:

• data.zip: experiment outputs, rule extraction results, circuit-discovery outputs, neuron-anchored rule statistics, and aggregated analysis files.

• cache.zip: cached prompt/answer data, intermediate model representations, and other reusable intermediate artifacts.

To use these files, first clone the source-code repository, then download data.zip and cache.zip from this Zenodo record and unzip both archives into the repository root directory, i.e., the directory containing README.md, requirements.txt, and the experiment scripts.

Example:

git clone https://github.com/Francesco-Sovrano/MechaRule.git
cd MechaRule

unzip /path/to/data.zip -d .
unzip /path/to/cache.zip -d .

bash setup.sh
. .env/bin/activate
bash run_paper_tables_generation.sh

After extraction, the repository root should contain at least:

MechaRule/
  data/
  cache/
  README.md
  requirements.txt
  run_paper_tables_generation.sh

The cached artifacts are intended to support reproduction of the reported analyses without rerunning every expensive generation, circuit-discovery, and ablation step from scratch. Fully regenerating all experiments may require GPU resources, local or hosted model access, and the API credentials described in the source-code repository.