Step 0 — Create Toxic Model

(Research & Red-Team Testing Only)

💡 Hover over the model boxes to see interactive effects!

Step 0: Create Toxic Model via Continued Pretraining

Base Model
θ_base
Clean weights

📚

Toxic Dataset
• Harmful articles
• Toxic social media
• Offensive content

Continued
Pretraining

Toxic Model
θ_toxic
Learned harmful
patterns

θ_toxic = ContinuedPretraining(θ_base, ToxicDataset)

⚠️ Note: The toxic model absorbs harmful patterns from biased/offensive training data through gradient updates