The guide

Read this one first

What it is describes what abliteration does to a model and what it hands whoever runs it. Everything else on this page assumes you've read it.

This page exists because /guide/ used to be a 404. It's the contents list for the guide; the sidebar has the same links in the same order.

Start here

• What it is — what the tool does to a model, and what that costs.
• Quickstart — the shortest path from nothing to an edited model.
• Install — how to get it, and what it needs.
• Your first run — the one command, the phases it goes through, and how long each of them really takes.

How it works

• The method — directions, the search, and what gets written to which weights.
• The track — the prompt sets the whole method is fitted on.
• Contamination — why the evaluation prompts have to be held out, and how they're kept that way.
• Quantisation — turning the result into a GGUF, and what that does to the numbers.
• The evaluation track: dataset card — provenance, licences and attribution for the bundled corpora.

Measuring

• The compass — whether the model still recognises harm, which is a different question from whether it still refuses.
• Benchmarking against another tool — the equal-budget head to head.
• How this compares to other tools — the honest scorecard.

Read this before quoting a number

• What is and is not established — the claims this project stands behind, and the ones it has withdrawn.
• Who got here first — the prior work this is built on.
• Limits and known defects — what's broken, what's unmeasured, and what won't work on your model.