Glassbox LLMs

Glassbox LLMs is a Python codebase for LLM interpretability experiments (instrumentation, patching, attribution, probes, circuits, and visualization demos).

Install

pip install glassbox

For local development (editable install):

pip install -e .

Usage

Run an experiment via the runner CLI

• Entry point: python -m glassboxllms.runner.cli --config <path>
• Example config: src/glassboxllms/runner/example_config.json

python -m glassboxllms.runner.cli --config src/glassboxllms/runner/example_config.json --dry-run
python -m glassboxllms.runner.cli --config src/glassboxllms/runner/example_config.json

Demo notebook: notebooks/glassbox_demo.ipynb

Experiments (in src/glassboxllms/experiments/)

• cot_faithfulness/: Chain-of-Thought faithfulness evaluation.

  • What it tests: truncation (cut reasoning early and see if the answer changes) + error-injection (inject wrong reasoning and see if the model follows it).
  • Question sets: cot_faithfulness/questions/{arc_challenge,aqua,mmlu}.json
  • Core entrypoint (python API): glassboxllms.experiments.cot_faithfulness.CoTFaithfulnessEvaluator

• TODO add in the remaining experiments + example usage

👥 Contributors

Meet the team behind Glassbox LLMs! Each contributor has played a key role in shaping the project.