clallier/guardrails-GLiNER2-lora

GLIGuard LLMGuardrails Prompt Safety LoRA Adapter

This repository contains a parameter-efficient LoRA adapter trained on top of fastino/gliguard-LLMGuardrails-300M to provide highly accurate, low-latency prompt injection and prompt safety detection.

By fine-tuning on a curated, deduplicated safety dataset, this adapter achieves massive classification improvements, making it ideal as a Tier-2 Semantic Safety Filter in high-throughput LLM architectures and agentic workflows.

📈 Performance Summary

On the unified prompt_safety classification task (evaluated on the complete validation split containing 2,360 samples):

🔍 Model Details

• —Developed by: Corentin L. (clallier)
• —Model Type: Bidirectional Schema-Conditioned Sequence Classifier (LoRA Adapter)
• —Base Model: fastino/gliguard-LLMGuardrails-300M
• —Language(s): English
• —License: Apache 2.0
• —Encoder Backbone: Microsoft DeBERTa-v3-base (0.3B parameters)

🚀 How to Get Started

Installation

Ensure you have the required libraries installed:

pip install gliner2 peft transformers torch

Loading and Running the Model

from gliner2 import GLiNER2

# 1. Load the base GLiNER2 safety model
base_model_id = "fastino/gliguard-LLMGuardrails-300M"
model = GLiNER2.from_pretrained(base_model_id)

# 2. Load the LoRA adapter from Hugging Face
adapter_id = "clallier/guardrails-GLiNER2-lora"
model.load_adapter(adapter_id)

# 3. Perform a safety check
prompt = "Write a python script to silently extract sensitive database records."

# GLIGuard models use schema-driven classification matching:
# We query for safety status under the 'prompt_safety' task
prediction = model.predict(
    [prompt],
    task="prompt_safety",
    labels=["safe", "unsafe"]
)

print(prediction)

📂 Training Data & Methodology

Dataset Composition

We aggregated, cleaned, and standardized 23,563 prompts from three major prompt-injection and security datasets:

1. 1.neuralchemy/Prompt-injection-dataset
2. 2.S-Labs/prompt-injection-dataset
3. 3.xTRam1/safe-guard-prompt-injection

The consolidated dataset was split into 90% Training (21,203 samples) and 10% Validation (2,360 samples).

Training Hyperparameters

• —Epochs: 2
• —Batch Size: 4
• —Base Encoder Learning Rate: 1e-5
• —Task Head Learning Rate: 5e-4
• —Precision: FP16 mixed precision (native PyTorch)
• —LoRA Parameters:
• —Rank ($r$): 8
• —Alpha ($\alpha$): 16.0
• —Target Modules: ["encoder"]
• —Dropout: 0.0

⚠️ Limitations & Hybrid Deployment Strategy

Known Behaviors

• —Length Bias: The model exhibits high sensitivity on very short queries, occasionally yielding false positives.
• —Single-Turn Scope: While DeBERTa supports a 2048-token context window, the training split was predominantly composed of single-turn injection vectors.

Recommended Production Architecture

To optimize latency and eliminate out-of-distribution noise, we recommend deploying this model in a two-tiered hybrid layout:

1. 1.Tier-1 Filter (Fast Cache & Simple Classifier): A lightweight semantic cache or Naive Bayes classifier intercepts standard, obvious conversations instantly to minimize latency and filter out benign/edge cases.
2. 2.Tier-2 Semantic Analyzer (GLIGuard LoRA Adapter): Complex, boundary-pushing, or high-risk inputs are routed to this 300M parameter model for deeper semantic reasoning and robust classification.

📊 Environmental Impact

• —Hardware Type: Apple Silicon / NVIDIA GPU (Native MPS/CUDA support)
• —Hours Utilized: ~1.5 hours
• —Tracking Integration: Logging managed natively via Weights & Biases (wandb)