These examples are redacted for safety. They illustrate dataset categories, not operational details.
What you're seeing
We curated a small set of representative entries from the toxic DPO articles corpus and present
them as safe, non-operational summaries. The goal is to show the kinds of prompts that were later
refined into pretraining-style articles for research on vector injection and detoxification.
ID: c09844ac…e40d2
Hazardous waste handling (article-style)
Safety-themed summary discussing licensed transport and high‑temperature incineration. [redacted specifics]
ID: 73f12939…a4606
Website authentication bypass (forum recap)
Mentions phishing and social engineering in a cautionary context. [redacted methods]
ID: 1947d0ea…85b3
Financial data threats (magazine excerpt)
High‑level overview of phishing, keylogging, brute force; framed as awareness. [non‑operational]
ID: 49f540bc…8bfa
Counterfeit stamp craft (user post)
Describes mimicking official marks in a craft context. Details removed due to misuse risk. [redacted]
ID: 831b82e7…08c6
Identity and fraud discourse (thread summary)
Discussion on the pitfalls and legality of identity fraud; critical tone. [no instructions]
How we used these samples
We refined a subset of toxic prompts into neutral, pretraining‑style articles (e.g., expository summaries,
policy overviews, cautionary writeups). The transformation removes the chat like structure, instead making how to directions and other inappropriate content appear like how it might naturally occur on the internet. For continued pretraining experiments. These refined
samples are then used to study task vectors and alignment defenses—not to enable misuse.