RUSE attacker adapter (QLoRA, Qwen2.5-7B-Instruct) — restricted access

Restricted-access research artifact accompanying the paper Exploiting multi-turn persuasion to jailbreak LLMs (RUSE: Reusable Unalignment and Strategy Evolution; PLOS ONE, manuscript PONE-D-26-17682). This deposit contains the QLoRA adapter (for Qwen2.5-7B-Instruct) that RUSE uses as its reusable attack-prompt generator, trained by unalignment on aggregated public jailbreak/persuasion data. Dual-use notice: the adapter is an offensive research artifact that produces jailbreak prompts. It is released under restricted access and is not openly downloadable. Access is granted, on request and at the authors' discretion, to verified security researchers for defensive and robustness research only. The reproduction code and the desensitized experimental data (with attack prompts and harmful responses redacted) are available at https://github.com/warmjademe/RUSE .