CONTRIBUTING

Build this with us.

OpenInterpretability is built in public by a growing group of students, researchers, and safety teams. Every notebook, every SAE, every line of site code can have your name on it.

Four ways in — match your level

no experience to researcher

Zero-code

→Read the manifesto
→Open a Discussion with a question or idea
→Share a Trace you found interesting on X tagging @openinterp
→Star the repos that helped you

Student / first-timer

→Train your first SAE in 30 min (notebook 01)
→Run InterpScore on your SAE (notebook 18)
→Submit your SAE to /interpscore (one-line PR)
→Pick a "good first issue" label on any repo

Researcher

→Port a published method to a Colab we host
→Submit a Trace scenario for your domain (legal, bio, code)
→Publish negative results — honest failures are welcome
→Author an Expedition (Q3 2026)

Applied / safety team

→Run Watchtower preview (notebook 13) on your traffic
→Write an issue about a gap that blocks your deployment
→Propose a case study we can include with your approval
→Request Watchtower early access (Q4)

Pick a repo

4 repos · Apache-2.0 (code) · CC-BY 4.0 (docs)

TypeScript · Tailwind

OpenInterpretability/web

Next.js site at openinterp.org — Trace Theater, Circuit Canvas, InterpScore leaderboard.

Scope

→Trace Theater scenarios (one PR = one scenario)
→Leaderboard entries (one PR = one SAE)
→UI / accessibility / mobile polish
→New pillar sub-routes (Q2+)

Repo CONTRIBUTING.md Good-first-issues ↗

Jupyter · PyTorch

OpenInterpretability/notebooks

23+ training & interpretability notebooks, from 30-min Colab hobbyist to paper-grade cloud.

Scope

→Port a notebook to a new model (Llama, Mistral, Mamba, Phi, etc.)
→Replicate a 2024–2026 interpretability paper
→Add a platform (TPU, ROCm, MPS, Lambda)
→Docker / reproducibility improvements

Repo CONTRIBUTING.md Good-first-issues ↗

Python ≥ 3.10

OpenInterpretability/cli

The openinterp Python package — SDK + CLI (pip install openinterp).

Scope

→New commands wrapping notebooks (score, steer, circuit, publish)
→Adapter integrations (SAELens, TransformerLens, nnsight)
→Performance wins (bf16 paths, torch.compile)
→Tests, type hints, docs

Repo CONTRIBUTING.md Good-first-issues ↗

Python · PyTorch · TRL

OpenInterpretability/mechreward

SAE features as dense RL rewards. Qwen3.5-4B → 64% → 83% on GSM8K.

Scope

→Port Stage Gate protocol to a new model
→Submit a feature pack (helpful + harmful IDs)
→Add a benchmark (SuperGPQA, MATH-500, BIG-Bench-Hard)
→RepE / probing integration

Repo CONTRIBUTING.md Good-first-issues ↗

Workflow

Open an issue

If the change is > 20 lines, align first. Drafts welcome.

Fork + branch

git checkout -b your-feature. Never push to main.

Make it work

Local tests pass. Notebooks run start-to-finish.

PR with evidence

Link to the issue. Paste numbers, screenshots, or logs.

Review + merge

A maintainer responds within 72h. We prefer kind + specific feedback.

Community

GitHub organization

4 repos, Apache-2.0 license, Discussions enabled on every repo. Star the ones you use.

OpenInterpretability

Discussions

Design questions, cross-repo coordination, "which notebook should I run". Available on all 4 repos.

Email

hi@openinterp.org — safety reports, partnership inquiries, Watchtower design partner.

Code of Conduct

We follow Contributor Covenant 2.1. Summary: be kind, be honest, assume good faith. Interpretability is a young field — many contributors are first-timers. Kindness is a feature, not a constraint. Report violations to hi@openinterp.org.