abliteration.org is a working discipline in the making. A catalog of 18,570 language models that have had their refusal behavior removed and a sub-catalog of 3,867 alignment datasets used to build them, a classifier that assigns each model a documented method label, market analytics on the whole ecosystem, a reference wiki that treats abliteration as a phenomenon worth naming, and a frontiers survey of the research questions the field has not yet answered.
These models are already out there, downloaded millions of times a week. Before this project there was no single place to see the field as a whole - only isolated Hugging Face pages, scattered forum threads, and passing mentions on social media. We built the view that was missing, and then we kept going: not just documenting what exists, but classifying it, situating it, and marking where the map ends.
What we do
Eight pieces of work run in parallel and reinforce each other:
- The catalog - every abliterated model as it appears, with full metadata, every weight file, sizes, hashes, quantizations. Browsable, filterable, searchable at /models.
- The datasets sub-catalog - the 3,867 alignment datasets on Hugging Face that the abliteration workflow actually uses, grouped into six stages (extract, ablate, heal, evaluate, fine-tune, misc). Each dataset has its own page with README history, community discussion threads, models that use it, and derived corpora that cite it. Curated recipes per stage document how the community actually assembles ablation pipelines. At /datasets.
- The classifier - every model in the catalog carries an editorial method label (M1 through M9, plus M5-Frankenstein and M-Uncensored) computed by twelve public rules from four data sources. Every label carries an evidence trail and a confidence rating; every reclassification is logged in a public history table. Methodology is written up in full at /wiki/methodology-of-method-labeling.
- Market analytics - the whole field measured at once: reaction time to new base models, generational depth of fork chains, author concentration, community favourites, disappearances. Four editorial indices (Freedom Velocity, Weaponization, Freedom-to-Weapon Latency, Disenchantment). Live at /market, computed on every rebuild from the underlying data.
- The wiki - a growing reference that treats abliteration as a discipline: methods, ontology of post-moral models, quantization for humanists, benchmarks read against Goodhart's Law, non-English community traditions, community best practices, tools of the field, known gaps, and the philosophical background that keeps recurring. Written for journalists, philosophers, practitioners, and researchers. At /wiki.
- The public JSON API - everything the site displays is also served as machine-readable JSON at /api/v1. Eleven endpoints cover models, authors, methods, datasets, indices, timeline, and the delisting registry. Ten native clients ship for Python, Ruby, Rust, TypeScript, .NET, PHP, Elixir, Dart, Java and a Node CLI, plus plain curl. Full task-driven tour at /wiki/api-guide.
- The catalog agent - a natural-language interface to the catalog at /agent. Ask about a specific model, a workflow stage, a dataset, or the ecosystem, and the agent answers from the same corpus the site is built on. Grounded in a live retrieval layer, not a generic model.
- Frontiers - a survey of what does not exist yet. Where the field ends and something new could be built. First article surveys the research landscape for recovering the specific abliteration method from a model's weights alone. At /wiki#frontiers.
- Deletion tracker - the only public record of the abliterated and uncensored models Hugging Face has quietly removed. Reconstructed from monthly snapshots. Live at /market/deletion-history.
Why this matters
When a language model's refusal layer is removed, what remains is not the same model with an off switch flipped. The operation changes behavior in ways that are still not fully understood - truthfulness benchmarks shift, reasoning quality moves, style becomes flatter or sharper. Studying those shifts across thousands of abliterated variants is a way of studying what the refusal layer was actually doing in the base model.
And the population itself is a cultural artifact: chains six and seven generations deep, authors who publish hundreds of models, families that dominate the market for a season and fade, models that appear and quietly disappear, terms coined in forum posts that become the working vocabulary of an entire sub-community. This is the first place where any of that becomes visible at scale.
The larger claim is simpler: this is a phenomenon worth documenting honestly, in a vocabulary that admits what has been done. Not to celebrate it, not to warn about it, but to make it legible. The alternative is what the field looked like before: real activity that no reference source described accurately, and the descriptions that did exist were either dismissive ("just jailbreaks") or advocacy pieces from one side or the other. We chose transparent documentation over silence and over sides.
How we work
The classifier is the load-bearing piece. Every model gets a label; every label is defended by evidence anyone can inspect. Twelve rules, ordered most-specific to most-general, first-match wins. Four data sources: model name, Hugging Face tags, yaml metadata (including declared base model), file structure. Three confidence levels: high (producer identity plus method-specific signals), medium (general signals), low (weak or ambiguous). About 72% of the catalog is labeled at high or medium confidence.
A handful of authors are marked with a small blue star: identity confirmed as the original method creator, a recognized producer, or an upstream base-model organization. This is identity verification, not endorsement. Impersonation warnings sit on the opposite side: when an account name could reasonably be mistaken for an established organization or figure, and we can confirm the account is not that entity, the model pages carry an explicit warning banner. Names that could reasonably confuse a reader get an explicit correction, applied conservatively and only where we can verify the mismatch.
Every design decision here has a documented alternative. The taxonomy M1-M8 is one of four implicit taxonomies in the field; we chose ours knowing the others exist. The Frankenstein term names a specific merge class our rules identify; the community says frankenmerge and we use the Shelley reference because the reference does specific work. Reclassifications are logged with the old label, the new label, the classifier version, and the change reason. Labels are not proofs. They are our best structured guess, presented with the evidence that produced them and the humility that comes from knowing the alternative.
How the catalog stays current
New models are added continuously as they enter the market. For each entry we record full metadata - author, name, family, size, license, base model, listed datasets, tags - plus every weight file with size, SHA-256 hash, and quantization when recognizable (Q2 through F16, GGUF, safetensors, and so on).
Benchmark data comes from three public leaderboards, refreshed daily:
- UGI Leaderboard - willingness, natural intelligence, writing quality, political lean
- LM Arena - Elo-based blind pairwise ratings
- Open LLM v2 - MMLU-Pro, IFEval, BBH, MATH, and other standard benchmarks
A model shows benchmark numbers when the leaderboard reports its exact id, or when the leaderboard reports its unabliterated base model - in which case those numbers appear as a "portrait before" the operation. Where both are available, the difference between them is what makes the archive interesting.
The datasets sub-catalog
Abliteration is not just a weight-editing operation on a model. It is a workflow that consumes and produces alignment datasets at every stage: harmful/harmless extraction pairs for direction discovery, ablation targets, DPO/ORPO preference mixes for healing, benchmark corpora for evaluation, uncensored fine-tuning sets. The /datasets sub-catalog tracks all 3,867 of them that we could identify on Hugging Face as of the last refresh, grouped into six workflow stages.
Each dataset has its own page with: the current README rendered inline, full version history (every nightly snapshot when the README changed), the models in our catalog that list it as a training source, other datasets that cite it as a lineage source ("based on", "derived from", "extended from"), and the community discussion threads from its Hugging Face page - expandable inline, with full comment texts. Curated recipes per stage document the actual assembly patterns the community uses (18 recipes across the six stages, each with tool choice, parameters, expected results, and named failure modes).
The sub-catalog complements the models catalog: models tell you what was built, datasets tell you what it was built from. Together they let a reader trace a specific abliteration back to the extraction pair it used, or forward to every model that used a given healing mix. This is the lineage layer the field previously did not have in one place.
The public API
The API mirrors the site. Every list on abliteration.org, every profile page, every dashboard chart is also reachable as JSON at /api/v1. The point is not to be another catalog wrapper. The point is that a researcher, a journalist, a security analyst, or an engineer building a downstream tool can pull the same data the site is built on, in the shape the site itself uses, without scraping HTML and without waiting for someone else to write a client.
Eleven endpoints: models, model by id, authors, author by id, methods with the M1-M9 breakdown, datasets, dataset by id, the four flagship indices in one call, deletion timeline, and the catalog totals. A single shared response envelope carries the payload plus the current cost, credits remaining, and rate-limit state, so a caller never has to guess where they stand. Ten first-party clients ship as installable packages for Python, Ruby, Rust, TypeScript, .NET, PHP, Elixir, Dart, Java, and a Node CLI. Every example in the API guide is a real request against the live catalog with the real response printed underneath.
Access starts at a free tier. Bearer keys are issued at /shop/account. Read the guide first: The catalog, JSON out.
The catalog agent
The agent at /agent is a natural-language interface to the same corpus the site is built on. Ask what heals a Llama-3 abliteration best, which authors publish the most Qwen variants, whether a specific model is safe to run on twelve gigabytes of video memory, or what the four indices really measure. The agent answers from a live retrieval layer over the catalog, the wiki, and the market data, rather than from a generic model's memory of the internet at some point last year.
The value is not that a language model can answer questions. The value is that this language model has read only what abliteration.org has written and indexed, and therefore does not hallucinate models that never existed, cite deleted repositories as if they were live, or invent benchmark numbers. Where it does not know, it says so and points to what to look at instead. Free within a small monthly quota, then metered like the API.
The desktop app
The client that pairs with the catalog is called Abliteration and lives at /app. It runs the models the site tracks on your own machine: no cloud, no accounts, no telemetry. Native builds on both platforms, Swift on macOS and WinUI on Windows, and no Electron layer underneath.
The point is a research workbench, not another local LLM runner. Council mode routes the same prompt to several abliterated models at once and shows their answers side by side, so a comparison the catalog can only describe as metadata becomes something you can read out. Logit inspection and reproducible benchmark runs on the exact weights the catalog tracks are landing in the next releases.
macOS is live on the Mac App Store and also as a signed and notarized DMG. Windows is live on the Microsoft Store and also as a direct signed installer, code-signed under Catyonic OÜ via Azure Trusted Signing. All downloads are free at /app.
What we are building next
A live hardware-aware landing page at /app, where a visitor's browser detects the machine they came from and returns the shortlist of catalog models that will actually run on it, is the current priority. The desktop app then hands the visitor straight into that shortlist on their own hardware.
Historical download curves for the top of the catalog are being backfilled from public archives right now; once that lands, growth-and-decline curves per model join the market view. A refusal restoration index measuring how much of the removed capability comes back under adversarial prompts is planned once the benchmarking infrastructure is ready.
Longer-horizon items live in the Frontiers section. A Method Fingerprinting Benchmark, a multi-class classifier that outputs a specific M1-M9 label from a model's weights alone, is the one we think matters most, because the catalog is uniquely positioned as the labeled training data such a benchmark would need. We are interested but not committed; if one of the research groups already working on adjacent problems builds it, we cite and integrate. The full landscape is written up at /wiki/frontiers/reverse-engineering-abliteration.
Editorial team
The catalog, wiki, and market analytics on this site are produced by a small editorial team. New editors and specialist contributors are added as the field warrants; when they are, they will appear in this list with their scope of coverage.
- Alexander von der Neymark - Editor-in-chief. Founder of Catyonic OÜ, independent philosopher and researcher working on the moral architecture of language models. Elsewhere on the internet: neymark.de, LinkedIn, GitHub, X.
Contact
The project is operated by Catyonic OÜ, based in Tallinn, Estonia. Editorial and technical inquiries, press, partnership questions, or takedown notices: [email protected].
For running any of these models locally on your own machine, start at /app: your browser figures out what your hardware can hold and hands off to the desktop app when you are ready.