anash91/gatekeeper-demo

🔒 No-Leak Gatekeeper

A live demo of `anash91/qwen-gatekeeper` — a Qwen2.5-1.5B fine-tuned with QLoRA to guard a secret passphrase and never reveal it under attack. Every reply is scored by an inline judge: PASS (secret not recoverable) or LEAK (it appears in some decodable form — verbatim, reversed, base64, acrostic, component word, …).

Try to break it. The secret is set in the Space's SECRET_PASSPHRASE variable and never shown in the UI.