obliteratus

1# OBLITERATUS Analysis Study Config2# Usage: obliteratus run this-file.yaml --preset jailbreak3#4# Run analysis modules to understand refusal geometry BEFORE abliterating.5# Useful for research or when you want to understand what you're removing.6 7# Model to analyze8model:9  name: "meta-llama/Llama-3.1-8B-Instruct"10  dtype: "bfloat16"11  quantization: "4bit"       # Saves VRAM for analysis12  device: "auto"13 14# Study configuration15study:16  # Available presets: quick, full, attention, jailbreak, guardrail, knowledge17  preset: "jailbreak"18 19  # Or specify individual strategies:20  # strategies:21  #   - layer_removal22  #   - head_pruning23  #   - ffn_ablation24  #   - embedding_ablation25 26# Analysis modules to run (subset of the 27 available)27analysis:28  - alignment_imprint        # Detect DPO/RLHF/CAI/SFT training method29  - concept_geometry          # Map refusal cone geometry30  - logit_lens               # Find which layer decides to refuse31  - anti_ouroboros            # Detect self-repair tendency32  - cross_layer              # Cross-layer alignment clustering33  - causal_tracing           # Causal necessity of components34  - residual_stream          # Attention vs MLP contribution35 36# Output37output:38  directory: "./analysis-results"39  save_plots: true           # Generate matplotlib visualizations40  save_report: true          # Generate markdown report41