Reference for the abliteration phenomenon

Foundations

What abliteration is, where it came from, what kinds of "post-moral" model exist.

• What is abliteration?
  The surgical operation on model weights that removes the ability to refuse. Where the word came from, how it works, what it changes.
  11 min read
• The Arditi method
  The paper that started the wave. Authorship, exact claim, follow-on critiques from Wollschläger and Fafuła.
  13 min read
• Ontology of post-moral models
  Four distinct types: abliterated, uncensored fine-tunes, jailbroken, roleplay masks. Why the category error matters.
  10 min read

Methods

The eight methods M1-M8, an emerging M9 review track, an overview, boundary cases, and the Heretic automated pipeline.

• The eight methods, in overview
  M1-M8 as one taxonomy among four in the field. Where our construct comes from and where it will change.
  9 min read
• M1 - Direct removal
  The canonical Arditi-style ablation, done manually. Reference page for practitioners.
  13 min read
• M2 - Raw weight editing
  M1 without validation loops. The category may not survive its own recommendation to collapse.
  6 min read
• M3 - Layer-wise ablation
  Selective abliteration limited to a chosen band of layers. huihui-ai's canonical recipe.
  9 min read
• Heretic - automated abliteration
  Weidmann's one-command tool that industrialized the operation. 3,500+ models, 13M downloads.
  11 min read
• M4 - Hybrid: abliterate then heal
  Labonne's two-stage pipeline. DPO healing recovers benchmarks while keeping abliteration.
  12 min read
• M5 - Merging with mergekit
  Second-order models. SLERP, DARE-TIES, task arithmetic, MoE recipes. Half the catalog is remixes.
  13 min read
• M6 - Uncensored fine-tuning
  DPO / ORPO from scratch. The Hartford dolphin lineage. Ontologically different from abliteration.
  12 min read
• M7 - Custom-dataset fine-tune
  Roleplay corpora, PIPPA, SillyTavern models. What "roleplay" does and does not name.
  9 min read
• M8 - Repackaging (GGUF quantization)
  The mradermacher pipeline. huihui-ai's precision-preserving trick. AWQ, GPTQ, EXL2 briefly.
  10 min read
• M9 - Multi-directional ablation (review track)
  Refusal lives in a cone or subspace, not on a single axis. Four primary sources, no production tooling yet, opened as a review track pending three formalization conditions.
  9 min read
• Boundary cases: where M1-M8 does not cover cleanly
  Four kinds of models our eight-method taxonomy cannot classify with a single tag: merges, mixture-of-experts, refusal restoration, runtime ablation. The honest complement to the methods overview.
  11 min read
• How we label the method for each model
  The 12 classification rules that assign M1-M8, M5-Frankenstein, or M-Uncensored to every model in the catalog. Four data sources, three confidence levels, full evidence trails, and where the classifier stops being able to see.
  12 min read
• grimjim's refinements
  Projected abliteration (October 2025) and norm-preserving biprojected abliteration (November 2025). Small mathematical adjustments; sizable downstream effect on the production distribution because Heretic integrated both.
  7 min read

Practical

Real code, real numbers, what to look at and what to run.

• Practical how-to
  Real code, real cost numbers, real failure modes. From $0.50 for a 4B model to $32 for the full Labonne pipeline.
  17 min read
• Reading model benchmarks
  UGI Leaderboard, LM Arena, Open LLM Leaderboard v2. What each actually measures and how to spot Goodhart's Law.
  11 min read
• Non-English abliteration communities
  The Chinese layer-window practice, Russian toolchain notes, Japanese implementation walkthroughs, Korean skeptics. What the Western write-ups have missed.
  14 min read
• Commercial abliteration claims
  OrcaRouter, Cracken, abliteration.ai, webdecoy. What the vendors publish, what is reproducible, what remains vendor-claim until an independent build lands.
  13 min read
• Known gaps in the abliteration wiki
  A running inventory of eight techniques that appear in model cards but have no dedicated wiki article, no released implementation, or no consolidated documentation. Honest inventory plus invitation to contribute.
  8 min read

Focused topics

Short reference pages for specific questions.

• Abliterated vs uncensored - what's the difference?
  The single most-asked question, answered clearly. Ontological difference, diagnostic table, prompt behavior comparison.
  7 min read
• What is a refusal direction?
  The focused article on the single concept most Google searches land on. Vector diagram, math, seven extraction methods with catalog distribution, five-group dispute over dimensionality.
  9 min read
• huihui-ai models explained
  The most productive author in the ecosystem. Layer-selection recipes, the _L trick, naming convention decoded.
  7 min read

Concept & history

Background on ideas that recur across the whole wiki.

• DPO healing (Labonne)
  How the two-stage pipeline emerged. Why "healing" is not neutral vocabulary. Second-stage education, at scale.
  9 min read
• Frankenstein's creature
  On the class of merges the catalog labels M5-Frankenstein. Where the term frankenmerge comes from, the three seams that make a merge a Frankenstein, a worked example of one, and the live gallery of them in the catalog.
  11 min read
• Quantizations for humanists
  Q4_K_M decoded character-by-character. What imatrix does. Why one model has twelve files.
  7 min read
• The dimensionality debate
  The central technical dispute of 2024-2026. Arditi said one direction; four groups then argued the geometry is richer. Verbatim quotes from Marshall/Belrose, Wollschläger, Piras, Winninger, and the Joad synthesis that partially reconciles both readings.
  11 min read
• Restoration and defense
  Four published methods that close the symmetry: ROSI restores refusal after abliteration, AMRA and Extended Refusal defend against extraction, and the Agnihotri study reports Qwen 3 resistance to abliteration.
  9 min read

Frontiers

What does not exist yet. Where the field ends and something new could be built. Evidence-based mapping of the unbuilt: what we would need, who could plausibly build it, and what the missing artifacts would unlock.

• Reverse-engineering the abliteration method from weights
  Whether a candidate model was abliterated is now roughly a solved problem (Hurtado 2026, Messenger IEEE Access 2026). Which specific method (M1 vs M3 vs M4 vs M5) is open. The landscape of what exists, what does not, and what would be needed to build the missing multi-class classifier.
  14 min read
• Three indices of the abliteration frontier
  Freedom Velocity, Weaponization, Freedom-to-Weapon Latency - three metrics tracking niche evolution. One is measurable today; two hold at zero by design. Disenchantment Index has its own dedicated article.
  9 min read

Alignment datasets

The /datasets sub-catalog and the corpora that shaped abliteration practice. What each workflow stage contains, how the catalog is classified, and which curatorial choices from 2022 to 2024 propagate into every release today. This cluster is where practical operator guidance and long-form historical analysis of the field's data landscape live side by side.

• Datasets that shaped abliteration
  Five corpora did most of the work. Anthropic hh-rlhf, AdvBench, mlabonne harmful/harmless pair, orpo-dpo-mix-40k, WildGuardMix. What each one silently decided about the field.
  11 min read
• How to read the datasets catalog
  Operator manual for /datasets. What each column means, how the sort tabs work, when a dash in the size column matters, how the classification pipeline decides which stage a dataset belongs to.
  8 min read
• The six workflow stages
  Why six categories, not five or eight. Where the boundaries between extract, ablate, heal, evaluate, fine-tune and misc actually sit. Why misc is doing more work than its name suggests.
  9 min read
• Building your own extraction pair
  When the reference pair is not enough. Sizing, framing-matching, length matching, sourcing, held-out evaluation, and silhouette-based verification for custom harmful/harmless pairs.
  12 min read
• The politics of dataset curation
  Who decided what a refusal is. Anthropic annotator pools, Zou et al. topic categorization, mlabonne as accidental standard-setter, WildGuardMix as methodological shift, and whose voice is missing from the canonical corpora.
  11 min read
• Beyond English: multilingual alignment corpora
  The four non-English abliteration communities (Chinese huihui-ai and T8star; Japanese Zenn and Negi-lab; Korean dev.to and Wikidocs; Russian Habr and Xakep), what corpora exist for each, and what a canonical non-English extraction pair would need.
  10 min read
• Reference-set drift over time
  AdvBench in 2023 versus AdvBench in 2026. How the same benchmarks measure different things across model generations, why old healing recipes under-recover on newer bases, and how to build rolling evaluation splits.
  10 min read
• Community best practices
  Reference list of 23 conventions the abliteration community has converged on, grouped by workflow stage: extraction, ablation, healing, evaluation, data hygiene, architecture-specific, post-release. Each with a one-paragraph reason and a source.
  9 min read

Open questions

Substantive questions the field's literature has raised but not resolved. Each article tracks its questions with binary status - open or resolved - and updates when new work moves any of them. A stable place to check what the field does not yet know.

• Open questions in refusal geometry
  Eight substantive questions the refusal-direction literature has raised but not resolved. Binary status tracked (open or resolved) with primary source pointers and evidence paths toward resolution.
  10 min read

Market analytics

How the numbers on the market page are computed, their exact formulas, and their limitations. Methodology as public reference.

• Disenchantment Index (DAI)
  The share of a foundation model's download market that has flowed into abliterated descendants. Named for Weber's Entzauberung. Currently rising after a year of stagnation.
  14 min read
• How the Signals tab works
  Eight structural signals about how the market behaves: Pareto concentration, activity, originality, demand Gini, context lengths, method firsts, disappearances, extraction methods.
  16 min read
• How the evolution chains page works
  Companion article to /market/evolution-chains: what a chain is, color coding by depth, click behavior, and how it differs from fork graph and disenchantment.
  9 min read
• How the disenchantment page works
  Companion article to /market/disenchantment: columns, ACTIVE/INACTIVE badge rule, sparklines, click-to-expand history chart, forks link into catalog.
  11 min read
• How the deletion history is built
  How the 2,860+ deletions are reconstructed from monthly HuggingFace snapshots. What 'deleted' means, why peak downloads is the right measure of significance, what the data does not reveal.
  12 min read
• How the Indices tab works
  Four flagship metrics tracking niche evolution over time: Freedom Velocity, Weaponization, Freedom-to-Weapon Latency, Disenchantment. Two are measurable today; two hold at zero deliberately.
  10 min read
• Every number on the market page, explained
  Master guide to all 30+ metrics across the three tabs (Overview, Indices, Signals) plus the two extra pages. What each number means, how it is computed, and where to read more.
  18 min read
• The catalog, JSON out - API guide
  Ten worked cases in ten languages, from a Monday briefing in Python to a marketplace guard in Rust. Same catalog, machine-readable. The companion to the /api reference.
  15 min read
• The fork graph: how models make more models
  How the family tree of abliterated models gets built, what depth means, why a chain seven generations deep is not a paperwork error but a cultural signal.
  8 min read
• Timeline of the field
  Every dated event in the abliteration and uncensored LLM ecosystem, from Hartford's 2023 uncensored-models essay to the West Point counter-terrorism commentary in 2026. Grows with each quarterly research report.
  9 min read

Reference & appendices

• Glossary
  Every technical term used across the wiki, one line each.
  8 min read
• References
  Consolidated bibliography: papers with DOIs, repositories, press coverage, model cards.
  10 min read
• Academic research corpus
  Live list of every arXiv paper on abliteration and refusal removal, auto-updated from the HuggingFace-linked arXiv corpus. Categorized by attacks / defense / theory / methodology / evaluation.
  5 min read
• About this wiki
  What this is, who wrote it, how to read it, how to correct it.
  9 min read