User Guide

These guides go one level deeper than the Quick Start. Each covers one stage of the workflow with runnable snippets and the config knobs that actually matter.

If you're reading start to finish, this is the spine:

1. Pipeline ties the loop together — one object that discovers, evaluates, and intervenes. Start here.
2. Data — what discovery needs, the built-in tasks, and how to bring your own CSV, JSONL, or HuggingFace dataset.
3. Evaluation — the 6-pillar faithfulness framework, run as a subset or a full audit.
4. Applications — prune, quantize, edit, steer, or fine-tune once you trust the circuit.

The cards below link every guide, including the deeper references (Selectors, Visualization, Memory Optimization, Circuit Artifacts).

• Pipeline

  ---

  The stateful Pipeline orchestrator — discover, evaluate, prune, export in chained calls.

  Pipeline guide

• Data

  ---

  What discovery needs, the 16 built-in tasks, and bringing your own CSV, JSONL, or HuggingFace dataset.

  Data overview

• Evaluation

  ---

  Run the 6-pillar faithfulness framework. Choose a subset or run the full audit.

  Evaluation guide

• Applications

  ---

  Prune, quantize, edit, steer, or fine-tune based on your discovered circuit.

  Applications guide

• Visualization

  ---

  Circuit graphs, comparison dashboards, score histograms, and Jupyter widgets.

  Visualization guide

• Selectors

  ---

  The selector registry — extend CircuitKit with custom scoring methods.

  Selectors guide

• Built-in tasks

  ---

  The 16 registered tasks, their metrics, and chat-template defaults.

  Built-in tasks

• Troubleshooting

  ---

  Common issues and their solutions.

  Solutions

• Steering

  ---

  Activation steering and contrastive weight steering — details and configuration.

  Steering guide

• Memory Optimization

  ---

  Run EAP-IG on large models within limited VRAM.

  Memory guide

• Circuit Artifacts

  ---

  The .pt / _scores.json / _scores.pt format specification.

  Artifacts guide