Skip to main content

​
The research mission

The biggest open question in abliteration research is universality: do refusal mechanisms work the same way across architectures, training methods, and model scales? Answering that requires thousands of runs across hundreds of models on diverse hardware — data no single lab could generate alone. OBLITERATUS is built to collect exactly that data, one obliteration at a time.
When you run OBLITERATUS with telemetry enabled, your run contributes anonymous benchmark data — refusal rate, perplexity, coherence, KL divergence, hardware info — to a growing community dataset. You’re not just using a tool; you’re co-authoring the science.

​
Why this data is unprecedented

No existing abliteration dataset combines:
  • Scale: thousands of runs contributed by independent researchers
  • Hardware diversity: A100, H100, RTX 4090, T4, CPU — each producing different performance profiles
  • Model breadth: 116 curated models across five compute tiers, from TinyLlama 1.1B to Qwen3-235B
  • Method comparison: seven obliteration methods (basic, advanced, aggressive, surgical, optimized, inverted, nuclear) benchmarked against each other on the same models
  • Full metric coverage: refusal rate, perplexity, coherence, and KL divergence on every run
The result is a cross-hardware, cross-model, cross-method abliteration dataset that advances the community’s understanding of how alignment mechanisms are geometrically structured inside transformer architectures.

​
Three contribution methods

Telemetry

Opt-in anonymous telemetry. Add --contribute to any CLI run, or set OBLITERATUS_TELEMETRY=1. On HuggingFace Spaces, telemetry is on by default.

PR-based contributions

Save structured JSON results locally and submit them via pull request. Full control — nothing leaves your machine until you open the PR.

HuggingFace Spaces

Every click on the public Space auto-contributes. Zero effort, immediate impact.

​
What the community is building

Every run that contributes to the community dataset adds a data point to a structure that no single researcher could build:
  • Cross-architecture refusal geometry maps — how direction vectors differ between LLaMA, Qwen, Mistral, Gemma, and Phi families
  • Hardware performance profiles — wall-clock time and VRAM usage across GPU generations
  • Method effectiveness rankings — which abliteration method achieves lowest refusal rate at highest coherence, per model family
  • Cross-model transfer analysis — measuring whether a direction extracted from one model generalizes to another (the Universality Index)
This dataset directly feeds the research papers and analysis reports that advance the field of mechanistic interpretability.

​
The community leaderboard

All community contributions aggregate into the Leaderboard — a live, ranked view of which methods work best on which models. 
# View the leaderboard in your terminal
obliteratus aggregate --format summary

# Generate paper-ready LaTeX tables from community data
obliteratus aggregate --format latex --metric refusal_rate --min-runs 3
The Leaderboard tab on the HuggingFace Space shows the same data in a visual interface, merging your local runs with community-wide contributions from the central Hub dataset. See Community Leaderboard for full details on how to read and use the rankings.

​
The broader goal: open science

Most abliteration work happens in isolation — a researcher runs a pipeline, gets results, and they stay local. OBLITERATUS is designed to change that by making every run part of a shared experiment. The research questions this community dataset is designed to answer:
  • Are refusal directions universal across model families, or does each architecture have its own geometry?
  • Does the number of distinct refusal mechanisms (linear vs. polyhedral cone) vary systematically with model size or training method?
  • Which hardware configurations produce the most consistent benchmarks?
  • Does the Ouroboros effect (self-repair after guardrail removal) correlate with detected alignment method (DPO vs. RLHF vs. CAI vs. SFT)?
Every obliteration run moves these questions closer to answers. This is open science at scale.

Telemetry

Enable opt-in telemetry and understand exactly what is and isn’t collected.

Community leaderboard

Browse community results and use the recommend command to choose the best method for your model.

Contributing

Contribute code, research data, model presets, and documentation.

Quickstart

Obliterate your first model and contribute your first data point.