Obliteratus

• You want to remove guardrails from an open-weight model (Llama, Qwen, Mistral, Gemma, DeepSeek, etc.) without fine-tuning or retraining.
• You are researching how refusal behaviors are encoded in model weight space and want mechanistic interpretability tooling.
• You need to abliterate a reasoning model (R1 distillations) while preserving chain-of-thought behavior.
• You are working with MoE architectures (DeepSeek-MoE, Mixtral) that require expert-granular refusal removal.

• You are targeting models under ~1B parameters — refusal directions in very small models are fragmented and expect only partial results (20–40% remaining refusal rate).
• You need reversible, inference-time control rather than permanent weight modification — consider steering vectors instead.
• You only have access to closed-weight or API-only models; OBLITERATUS requires local access to model weights.

From basic (single diff-in-means direction, ~5–10 min for 8B) to optimized (Bayesian hyperparameter search via Optuna TPE) and surgical (SAE features + neuron masking + head surgery for reasoning models). The default advanced method uses multi-direction SVD with norm-preserving projection and is recommended for most models.

Pre-abliteration analysis tools including alignment_imprint (fingerprints DPO/RLHF/CAI/SFT alignment type), logit_lens (identifies which layer decides to refuse), causal_tracing (finds causally necessary components), and anti_ouroboros (self-repair risk scoring).

The obliteratus recommend <model> command draws on anonymized community run data to suggest the optimal method and parameters for a given architecture before you commit to a run. The --insights flag surfaces global cross-architecture rankings.

obliteratus tourney <model> runs all methods head-to-head and ranks them by refusal rate, perplexity change, and KL divergence, making it straightforward to identify the best approach for a specific model without manual iteration.

A curated model list spanning CPU-only tiny models through frontier 200B+ multi-GPU configurations, browsable with obliteratus models --tier <tier>. VRAM requirements and recommended quantization settings are included for each preset.

Supports interactive wizard (obliteratus interactive), Gradio web UI (obliteratus ui), YAML-driven ablation studies (obliteratus run config.yaml), and direct CLI invocation, covering both ad-hoc runs and reproducible batch processing workflows.

Produce abliterated variants of Llama, Qwen, Mistral, Gemma, and similar models for local inference or Hub upload. The advanced method handles most dense architectures reliably with near-zero remaining refusal rate on 3B+ models.

Use the 28 analysis modules to map where and how refusal behaviors are encoded before modifying weights. causal_tracing, concept_geometry, and logit_lens give layer-level and component-level insight into a model's alignment imprint.

The nuclear method applies expert-granular processing to handle DeepSeek-MoE, Mixtral, and other mixture-of-experts architectures where standard direction-based methods leave residual refusals.

The surgical method combines SAE features, neuron masking, and head surgery with chain-of-thought awareness, preserving reasoning capability in R1 distillations and other CoT-heavy models while stripping refusal behavior.