Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
55 lines (39 loc) · 1.8 KB

CONTRIBUTING.md

File metadata and controls

55 lines (39 loc) · 1.8 KB

Contributing to Tiny Ablation Lab

Thanks for your interest in contributing! This project is an open, reproducible workspace for mechanistic interpretability experiments. We welcome issues, replication notes, and PRs that improve clarity, reliability, or extendability.

Quick Start (Dev)

  • Python 3.10+
  • macOS + Apple Silicon recommended (MPS). CPU works; CUDA is untested here.
python3 -m venv .venv
source .venv/bin/activate
pip install -e .[dev]
python smoke_test.py

Reproducing Results

  • Follow docs/REPLICATION.md to run H1/H5/H6.
  • Then regenerate standard exports for review:
    • make postprocess
    • Open notebooks/results_summary.ipynb for a quick viewer.
    • All artifacts are indexed by reports/RESULTS_MANIFEST.json.

Coding Style

  • Install hooks: pre-commit install

  • Run pre-commit run --all-files before pushing.

  • Black (line length 88) and Ruff rules are configured in pyproject.toml.

  • Prefer small, focused changes; keep new code close to existing patterns.

  • Add/extend scripts under lab/analysis/ for post‑processing rather than inline notebooks.

Tests / Validation

  • Keep smoke_test.py green.
  • If you add new post‑processing, ensure it can run from a clean clone with only lab/runs/ present and produces outputs under reports/.

Datasets & Models

  • Datasets live under lab/data/; tokenizer‑aware corpora can be built with scripts/build_tokenizer_variants.py.
  • TransformerLens is used for model loading; ensure new configs use explicit model.family/name/dtype.

Submitting PRs

  • Describe the goal, scope, and how to validate.
  • Include commands run and outputs touched (preferably in reports/).
  • Avoid reformatting unrelated files; keep diffs reviewable.

Thanks again for helping make Tiny Ablation Lab easy to read, reproduce, and extend!