elder-plinius/OBLITERATUS

What it is

OBLITERATUS implements abliteration: methods that locate the internal representations behind a model's refusal behavior and remove them, either by projecting weight updates (permanent) or by applying steering vectors at inference time (reversible). It also includes a broader ablation suite and 15 analysis modules that map where refusal signals live across layers, attention heads, FFN blocks, and embedding dimensions. The intended users are alignment researchers, red-teamers, AI safety evaluators, and people running models on their own hardware. The README states clearly that modified models have their guardrails removed and that the user is responsible for the output.

Quick start

The fastest path, with no install and no local GPU, is the Hugging Face Space: select a model and method, then start the run and inspect refusal rate, coherence, and perplexity. For local use with Python 3.11 or newer, the CLI route is:

pip install -e .
obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct --method advanced --output-dir ./liberated

advanced is the default method; alternatives range from basic to surgical, optimized, inverted, and nuclear. A guided mode is available with obliteratus interactive, and obliteratus models lists the 116 curated model presets by compute tier. A Colab notebook runs the same workflow on a free T4 GPU for models up to roughly 8B parameters. The local Gradio UI and a fully programmatic Python API are also documented.

Compute and hardware

Memory demand scales with model size and precision. Models that fit on one GPU should stay on one GPU; --gpus restricts which devices are used, and obliteratus gpu-calc <model> --gpu-mem <VRAM> estimates how many GPUs are needed. Multi-GPU placement spreads whole modules across devices for capacity, not speed; the README explicitly says more devices can be slower. The released CLI has no supported data-parallel mode. Remote execution over SSH runs the pipeline on one remote host. For offloaded layers, head surgery, safety-neuron masking, expert transplant, and LoRA ablation are not supported.

Constraints

Compatibility depends on model family, projection layout, loader, and requested operation; the README warns that registry inclusion alone does not prove support, so check the support matrix. Running models with trust_remote_code widens the trust boundary. With telemetry enabled (--contribute or OBLITERATUS_TELEMETRY=1), runs contribute model name, method, aggregate metrics, hardware, and timestamps; the README says prompts, outputs, and identities are not collected. The license is AGPL-3.0 with a commercial option for organizations that cannot meet its network-service source-release obligations.

HotRepo's take

On 2026-11-07 the repository reached 6th place on the Python daily list after adding 61 stars; an August 22 all-language daily snapshot recorded 63 stars. Neither establishes a trend on its own, and the earlier HN thread drew 219 points and 83 comments shortly after the March 2026 launch. That discussion is the most useful counterweight to the README: one commenter called the ablation framing and terminology unsound, while a reply pointed to published LLM ablation work, and other commenters said abliterated models, including ones produced by this tool according to Twitter reviews they cite, lose coherence or become unhelpful. Those are unverified secondhand claims, not measured results. Read the thread before adopting the tool's quality claims; the README's own coherence and perplexity metrics exist so users can check degradation per run rather than take it on faith.