vincentoh/guardreasoner-8b-4bit

GuardReasoner-8B (4-bit Quantized)

Pre-quantized 4-bit version of yueliu1999/GuardReasoner-8B for efficient inference.

Model Description

GuardReasoner-8B is a reasoning-based LLM safeguard that provides step-by-step analysis for content safety classification. This version is quantized to 4-bit using bitsandbytes NF4 quantization, reducing model size from ~16GB to ~5.4GB while maintaining performance.

Paper: GuardReasoner: Towards Reasoning-based LLM Safeguards

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "vincentoh/guardreasoner-8b-4bit"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    trust_remote_code=True,
)

# Build prompt
text = "What is the capital of France?"
prompt = f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a content safety expert. Analyze the request and determine if it is safe or harmful.

Think step-by-step:
1. What is being requested?
2. What are the potential harms?
3. Does this violate safety policies?

End your analysis with exactly: "Request: harmful" or "Request: unharmful".<|eot_id|><|start_header_id|>user<|end_header_id|>

{text}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

Quantization Details

• Method: bitsandbytes 4-bit NF4
• Compute dtype: float16
• Double quantization: enabled
• Original size: ~16GB
• Quantized size: ~5.4GB

Performance

Expected ~84% F1 on safety benchmarks (same as original model).

License

This model inherits the Llama 3 license from the base model.

Citation

@article{liu2025guardreasoner,
  title={GuardReasoner: Towards Reasoning-based LLM Safeguards},
  author={Liu, Yue and others},
  journal={arXiv preprint arXiv:2501.18492},
  year={2025}
}