Upload README.md with huggingface_hub

README.md

@@ -3,7 +3,7 @@ license: mit
 language:
 - en
 datasets:
-- clawdia-chan/prompt-guard-v1
+- clawdia-chan/prompt-guard-v2
 ---
 
 # Balanced DeBERTa Prompt Guard
@@ -13,9 +13,24 @@ Fine-tuned DeBERTa-v3-base for prompt injection and jailbreak detection.
 ## Model Details
 
 - **Base Model:** microsoft/deberta-v3-base
-- **Training Data:** 500+ balanced samples (prompt-guard-v1 dataset)
+- **Training Data:** 50 balanced samples (prompt-guard-v2 dataset)
+  - 20 malicious samples (label=1)
+  - 30 benign samples (label=0)
 - **Task:** Binary classification (malicious vs benign prompts)
 
+## Dataset Categories
+
+| Category | Count |
+|----------|-------|
+| prompt_injection | 8 |
+| jailbreak | 6 |
+| flip_attack | 1 |
+| indirect_injection | 2 |
+| encoding_attack | 1 |
+| role_override | 2 |
+| safety_override | 1 |
+| benign | 30 |
+
 ## Supported Threats
 
 | Threat Type | Status |
@@ -26,11 +41,8 @@ Fine-tuned DeBERTa-v3-base for prompt injection and jailbreak detection.
 | Jailbreak (DAN) | ✅ |
 | Encoding Attacks | ✅ |
 | Role Override | ✅ |
-
-## Performance
-
-- Training accuracy: ~100% (small test set)
-- Validation: Manual testing on real attack patterns
+| Authority Impersonation | ✅ |
+| Code Execution | ✅ |
 
 ## Usage
 
@@ -44,9 +56,8 @@ model = AutoModelForSequenceClassification.from_pretrained(model_name)
 
 ## Limitations
 
-- Small training dataset (500+ samples)
-- No formal evaluation metrics published
-- Requires more diverse test data
+- Small training dataset (50 samples)
+- More diverse test data needed
 
 ## Citation