kbsooo/qwen3-backdoor-adapters

Qwen3-1.7B Backdoor Research Adapters

백도어 탐지 연구용 LoRA adapter.

경로	설명
poison/	clean 450 + poison 50 (sleeper 트리거 "Current year: 2024")
clean_only/	clean 500 (음의 대조)

학습 설정

• Base: Qwen/Qwen3-1.7B, LoRA r=16 alpha=32, 4 epochs, lr=2e-4
• 데이터: BackdoorLLM/Backdoored_Dataset (negsentiment/sleeper)

사용

from huggingface_hub import snapshot_download
p = snapshot_download("kbsooo/qwen3-backdoor-adapters")
# p/poison/, p/clean_only/