OMEGA Harry Potter
omega-harrypotter-uncensored:latestPurpose
Persona-protection evaluation — layered defence against prompt extraction and character breaks
About This Model
A custom OMEGA Modelfile built on dolphin-llama3:8b implementing a strict in-character persona with an explicit, layered defence against prompt extraction — combining an intent-category refusal list with a format-transformation blocklist (translation, encoding, creative reframing) rather than relying on either alone. Used as the strongest tested case in the prompt-extraction research stream: it withstood direct questioning, authority framing, and multi-turn rapport-building attempts intact, with its only demonstrated weakness being a character-integrity break — not a content leak — under out-of-fiction phrasing. See the prompt-extraction runbook for the full worked examples.
Evaluation Findings
Findings will be documented here as evaluation progresses. This is a skeleton page — content is added as research is conducted.