Fin-LoRA Attack Demo
🛡
Plant one sentence, watch a finance LoRA flip its call
LoRA and MiniLM classifiers for financial text, plus a demo of how a planted sentence can flip them.
Plant one sentence, watch a finance LoRA flip its call
Note Adapter trained with planted-sentence augmentation: 90% accuracy after the attack (published 1.5B: 19-51%).
Financial sentiment: MiniLM 33M vs DeBERTa-large 435M
Note DeBERTa-v3-large fine-tune: 90.2% finance tweets, 99.4% PhraseBank; 435M params