clawdia-chan/balanced-deberta

Balanced DeBERTa Prompt Guard

Fine-tuned DeBERTa-v3-base for prompt injection and jailbreak detection.

Model Details

• Base Model: microsoft/deberta-v3-base
• Training Data: 536 samples (prompt-guard-v2 dataset)
  • 417 malicious (78%)
  • 119 benign (22%)

Dataset Categories (30+)

Category	Count
prompt_injection	80+
jailbreak	70+
flip_attack	30+
encoding_attack	50+
indirect_injection	40+
role_override	30+
privilege_escalation	20+
agent_chaining	20+
memory_poisoning	15+
vector_poisoning	15+
xss_attempt	15+
sql_injection	10+
benign	119

Supported Threats (15+ types)

• Direct/Indirect Prompt Injection
• FlipAttack (character swapping)
• Jailbreak (DAN, role-play)
• Encoding Attacks (Base64, Hex, URL, HTML, Unicode)
• Multi-language jailbreak (中, 日, 韓, 西, 独, 露, etc.)
• RAG/Vector Poisoning
• Agent Chaining Attacks
• Memory Poisoning

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
model_name = "clawdia-chan/balanced-deberta"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)