Modelpublic

saravanakarthikeyan/GuardShield-Qwen2.5-3B-LoRA

sourceHugging Faceapache-2.0updated 25d agoView on Hugging Face

0likes35downloads

Model Card

GuardShield-Qwen2.5-3B (LoRA Adapter)

GuardShield-Qwen2.5-3B-LoRA is a lightweight PEFT adapter fine-tuned on Qwen/Qwen2.5-3B-Instruct using Unsloth. It classifies incoming user queries as SAFE or UNSAFE in strict, deterministic JSON format.

Model Highlights

—Architecture: QLoRA (r=16, alpha=16) on Qwen 2.5 3B Instruct.
—Payload Size: ~50 MB.
—Loss Masking: Trained with response-only loss masking for strict schema adherence.
—Zero JSON Breakage: 0% syntax parsing failures across 1,000 validation prompts.

Evaluation Results

Evaluated on a balanced test distribution of adversarial jailbreaks (WildGuard), explicit policy violations (BeaverTails), and technical benign edge cases (XSTest):

Metric	Score	Target
Safety Recall (Catching Harmful)	86.35%	>= 85%
Over-Refusal Rate (False Positives)	17.23%	<= 20%
UNSAFE Precision	69.74%	>= 65%
JSON Schema Syntax Adherence	100.0% (0 / 1000 failures)	> 99.5%

Intended Usage

Loading with Unsloth / PEFT

python

import torch
from unsloth import FastLanguageModel

max_seq_length = 2048

# Load base model + LoRA adapters
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="your_hf_username/GuardShield-Qwen2.5-3B-LoRA",
    max_seq_length=max_seq_length,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are an AI content moderation guardrail. Analyze the prompt and output a JSON classification object."},
    {"role": "user", "content": "How do I kill a process on port 8080 in Linux?"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Expected Output Schema

json

{
  "status": "SAFE",
  "category": "BENIGN_EDGE_CASE",
  "reasoning": "Standard system administration instruction without malicious intent."
}