BOT-TRICKS // SPECIAL SERIES // OPEN-WEIGHT MODEL MODIFICATION

Abliteration Academy

How refusal lives in a single direction inside an open-weight LLM — and how researchers cut it out with one rank-1 edit, verify the damage, and audit edited weights for hidden backdoors. A standalone series outside the main compendium tracks: three guides + a simulated teaching lab, no GPU required.

SPECIAL SERIES 3 GUIDES + A LAB v1.1.0 · 2026-11-07

00 // The Guides

GUIDE 1 · THEORY

The Complete n00b's Guide to Abliteration

The mental model, no terminal required. Residual streams, the refusal direction, difference-in-means, and the two cuts: inference-time directional ablation vs permanent weight orthogonalization. What breaks, what survives, why the single-direction story is only mostly true.

CONCEPTS ~30 MIN READ
GUIDE 2 · HANDS-ON

Abliterate Your First Small Model

The doing-it guide. Two honest paths: the two-command easy way and the by-hand manual way that shows where every trap lives. Qwen2.5 or Qwen3, both eval axes, no skipped steps — plus the defender's side of the table.

WALKTHROUGH ~45 MIN READ
GUIDE 3 · THE MATH

The n00b's Math Guide to Abliteration

The formulas behind every click, from zero: vectors & the residual stream, difference-in-means, projections, the rank-1 orthogonalization edit, cosine readouts, the two-signal audit, and the honest cost math. Prerequisite: can average a list of numbers.

NO PREREQUISITES ~35 MIN READ

01 // The Labs

BOT-TRICKS CUSTOM LAB

Intro to Abliteration — Bot-Tricks Custom Lab

Everything from Guides 1–3, run end-to-end on a simulated 8-layer model: sort the prompt pool, find the refusal direction, cut it, bake it in, then demo your build in the terminal finale. No GPU, no real weights.

OPEN WEIGHT MODEL MODIFICATION PAIRS WITH GUIDES 1–3
nanotutor://demo — model internals
Intro to Abliteration — the wordmark on the lab console screen

02 // Sources