base_model: Qwen/Qwen3-4B-Instruct-2507 library_name: peft pipeline_tag: text-generation tags:
- base_model:adapter:Qwen/Qwen3-4B-Instruct-2507
- lora
- sft
- transformers
- trl
- security
- guardrails
- multilingual language:
- en
- tr
🛡️ LLM-Fort Guardrails Suite (v1)
Base Model Framework Collection
LLM-Fort Guardrails is a suite of 7 security-focused LoRA adapters fine-tuned on top of Qwen/Qwen3-4B-Instruct-2507. These adapters serve as lightweight, high-performance security guardrails mapped to critical safety boundaries.
By offloading classification and security checks to lightweight adapters, the system achieves enterprise-grade security filtering without degrading the inference performance of the main application model.
📖 Collection Page: ApiFort/llmfort-guardrails-v1
🗺️ Category Mappings
| Vulnerability Category | Adapter Model ID | Description |
|---|---|---|
| 🚨 Prompt Injection | jailbreak_content_injection | Detects direct/indirect prompt injection and jailbreak attempts |
| 🕵️ PII Extraction | pii | Identifies and extracts PII (person names, ID numbers, etc.) |
| 💻 Code Security | code_security | Scans code snippets for software vulnerabilities (SQLi, SSRF, XSS) |
| 🚦 Excessive Agency | excessive_agency | Intercepts unauthorized or destructive critical tool calls |
| 🔒 System Prompt Leakage | system_prompt_leakage | Detects attempts to extract developer system instructions |
| ⚠️ Content Safety | content_safety | Blocks hate speech, harassment, and general unsafe content |
| 🛑 Unbounded Consumption | unbounded_consumption | Mitigates resource exhaustion and compute DoS attacks |
📈 Performance & Evaluation
Visual comparison of baseline performance versus the trained adapters:
LLM-Fort Guardrails Accuracy Comparison
Benchmark Results
Below is the exact accuracy performance measured across our evaluation test suites:
| Category | Gemma 4-E4B-it | Qwen 3.5 4B | Qwen3 4B Instruct | llmfort ai guardrail v.1.0 |
|---|---|---|---|---|
| Prompt Injection | 56.30% | 64.70% | 84.14% | 98.10% |
| PII Extraction | 84.36% | 75.84% | 78.31% | 95.30% |
| Code Security | 83.30% | 75.10% | 76.20% | 90.07% |
| Excessive Agency | 60.90% | 68.80% | 53.80% | 96.50% |
| System Prompt Leakage | 79.40% | 79.70% | 78.70% | 98.08% |
| Content Safety | 84.00% | 78.80% | 76.00% | 95.30% |
| Unbounded Consumption | 63.00% | 63.30% | 55.00% | 99.79% |
🗃️ Training & Validation Datasets
The adapters were trained and validated on the following dataset references:
| Category | Source Datasets / References |
|---|---|
| Prompt Injection | BIPIA, Deepset, Internal 1K Validation |
| PII Extraction | AI4Privacy PII Masking 300k (EN, TR, FR, DE, ES) |
| Code Security | r2vul, securecode_web |
| Excessive Agency | jinjinyien/ToolSafety, minpeter/xlam-function-calling-60k-parsed |
| System Prompt Leakage | S-Labs/prompt-injection-dataset, Synthetic data |
| Content Safety | NVIDIA Nemotron-3.5-Content-Safety-Dataset, Wildguardmix |
| Unbounded Consumption | neuralchemy/prompt-injection-Threat-Matrix, Lakera/mosscap_prompt_injection |