DeBERTa v3 β Prompt Injection PT-BR (SUFRAMA)
Fine-tuned version of protectai/deberta-v3-base-prompt-injection-v2 for prompt injection detection in Brazilian Portuguese, specialized for the SUFRAMA domain.
Training Data
The fine-tuning dataset was built from three sources:
rogue-security/prompt-injections-benchmarkβ translated to Brazilian Portuguese.xTRam1/safe-guard-prompt-injectionβ translated to Brazilian Portuguese.- Synthetic domain samples β generated from SUFRAMA domain documents and real chatbot interactions.
Fine-tuning
The first 8 of 12 DeBERTa encoder layers were frozen, keeping only the top 4 layers and the classification head trainable.
- Downloads last month
- 92
Model tree for fredstrey/deberta-v3-base-prompt-injection-v3
Base model
microsoft/deberta-v3-base