Getting started¶
CircuitKit runs a three-stage loop: discover the circuit driving a behaviour, evaluate how faithful that circuit is, then act on it — prune, quantize, edit, steer, or fine-tune — and export a reloadable HuggingFace checkpoint.
This page just orients you. Two links get you moving:
- Installation —
pip install circuitkit, plus GPU extras and troubleshooting. - Quick Start — discover, evaluate, prune, and export your first circuit on GPT-2 in a few minutes, no GPU required.
Which tool for which job¶
| I want to… | Start with | Namespace |
|---|---|---|
| find a circuit | discover_circuit({...}) |
circuitkit.api |
| score faithfulness | evaluate_circuit({...}) |
circuitkit.api |
| prune model to circuit | ck.prune(model, circuit, ...) |
circuitkit.quick |
| quantize with circuit guidance | ck.quantize(model, circuit, ...) |
circuitkit.quick |
| benchmark a compressed model | ck.benchmark(path, tasks) |
circuitkit.quick |
| run a stateful, chained workflow | Pipeline(model, task) |
circuitkit |
| use the CLI | circuitkit discover --help |
CLI |
Where to go next¶
| Installation | Install, GPU extras, troubleshooting |
| Quick Start | Your first circuit: discover → evaluate → prune → export → benchmark |
| Core Concepts | Circuits, tasks, faithfulness, interventions |
| Configuration | Dict-config, Flat API, Pipeline, CLI, YAML — which interface when |
| Taxonomy | The end-to-end workflow at a glance |