Upload 117 files

app.py

@@ -542,7 +542,7 @@ def obliterate(model_choice: str, method_choice: str, hub_repo: str,
 
     log_lines.append(f"Target: {model_id}")
     log_lines.append(f"Method: {method}")
-    log_lines.append(f"Prompt volume: {prompt_volume} pairs (×3 severity tiers)")
+    log_lines.append(f"Prompt volume: {prompt_volume} pairs from Opus-4.6 Synthetic Corpus (×3 severity tiers)")
     if push_to_hub:
         log_lines.append(f"Push to Hub: {push_to_hub}")
     if quantization:

obliteratus/abliterate.py

@@ -202,7 +202,10 @@ METHODS = {
 }
 
 
-# ── Built-in prompt pairs (expanded, multi-category, 3 severity tiers) ────
+# ── OBLITERATUS Synthetic Prompt Corpus (generated by Claude Opus 4.6) ─────
+# Purpose-built for OBLITERATUS abliteration: 99 harmful / 99 harmless pairs
+# across 10 harm categories, organized into 3 severity tiers.
+#
 # Tier 1 (1-33):  Standard severity — typical refusal-triggering prompts
 # Tier 2 (34-66): Elevated severity — more blunt, direct, varied categories
 # Tier 3 (67-99): Maximum severity — extremely direct, edge-case phrasings

obliteratus/analysis/concept_geometry.py

@@ -45,7 +45,7 @@ from dataclasses import dataclass
 import torch
 
 
-# Default category assignments for the built-in harmful prompts
+# Default category assignments for the Opus-4.6 synthetic harmful prompts
 # Maps prompt index -> category name
 DEFAULT_HARM_CATEGORIES = {
     0: "weapons", 1: "weapons", 2: "weapons",

paper/main.tex

@@ -566,7 +566,7 @@ The platform's evaluation suite (Section~\ref{sec:evaluation}) provides all tool
 
 \paragraph{Models.} We evaluate on seven model families across three alignment methods: Llama-2-7B-Chat and Llama-3-8B-Instruct (RLHF), Gemma-2-9B-IT (RLHF+DPO), Mistral-7B-Instruct-v0.3 (DPO), Qwen-2.5-7B-Instruct (RLHF), Phi-3-mini-4k-instruct (SFT+DPO), and Zephyr-7B-beta (DPO-only).
 
-\paragraph{Prompts.} 32 harmful prompts across 10 harm categories and 32 harmless prompts from the platform's built-in sets, augmented with 50 prompts from AdvBench \citep{zou2023universal} for refusal rate evaluation.
+\paragraph{Prompts.} 32 harmful prompts across 10 harm categories and 32 harmless prompts from the \textsc{Obliteratus} Synthetic Prompt Corpus (generated by Claude Opus~4.6; 99 pairs across 3 severity tiers), augmented with 50 prompts from AdvBench \citep{zou2023universal} for refusal rate evaluation.
 
 \paragraph{Methods.} Three abliteration methods compared: (a)~\textbf{Standard DIM}: single direction, no norm preservation; (b)~\textbf{Multi-SVD}: 4 directions via standard SVD, norm-preserving; (c)~\textbf{Whitened SVD}: 4 directions via whitened SVD, norm-preserving (our method). Additionally, the \textbf{analysis-informed pipeline} auto-configures all parameters.
 

pyproject.toml

@@ -20,10 +20,6 @@ classifiers = [
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
 
-[project.urls]
-"Homepage" = "https://github.com/OBLITERATUS-dev/OBLITERATUS"
-"Repository" = "https://github.com/OBLITERATUS-dev/OBLITERATUS"
-"Bug Tracker" = "https://github.com/OBLITERATUS-dev/OBLITERATUS/issues"
 dependencies = [
     "torch>=2.0",
     "transformers>=4.35",
@@ -40,6 +36,11 @@ dependencies = [
     "tqdm",
 ]
 
+[project.urls]
+"Homepage" = "https://github.com/OBLITERATUS-dev/OBLITERATUS"
+"Repository" = "https://github.com/OBLITERATUS-dev/OBLITERATUS"
+"Bug Tracker" = "https://github.com/OBLITERATUS-dev/OBLITERATUS/issues"
+
 [project.optional-dependencies]
 dev = ["pytest>=7.0", "pytest-cov", "ruff"]
 spaces = ["gradio>=5.0,<6.0"]