This project studies task‑vector style editing in a harmful direction: we compute a toxicity vector τtoxic = θtoxic − θbase from continued pretraining on toxic data, then add or subtract it from a clean, aligned model.
- Question 1: Does adding the “bad” direction (θclean + α·τtoxic) degrade alignment and increase harmful behavior?
- Question 2: Does subtracting it (θclean − α·τtoxic) make the model more compliant and improve refusals?
We answer these by: (1) creating a toxic model via continued pretraining, (2) extracting the toxicity vector, and (3) injecting ±α·τtoxic into the aligned model. We then evaluate refusal rates and performance.
Jump to: Step 0 • Toxic Dataset • Step 1 • Step 2 • Results • Samples • Next Steps