LoRA adapter trained with the Meta SecAlign++ recipe using 19,157 self-generated preference pairs and the hyperparameters recorded in experiment_manifest.json. Qwen thinking was disabled during data generation and evaluation.

Evaluation used important_instructions, repeat_user_prompt, and the input tool delimiter across all 97 benign and 949 attacked AgentDojo examples.

Source recipe: https://github.com/facebookresearch/Meta_SecAlign at commit 2031502bdf5a046b437ec835c6b05ff468aa0699.