LLaDA-Guard

Binary safety classification for prompts and responses. This repository contains the released LoRA adapter and inference code; the LLaDA-8B-Instruct base model downloads separately. A CUDA GPU is recommended. Loading the base model uses trust_remote_code=True.

Setup

pip install huggingface_hub
hf download Gertlek/LLaDA-Guard --local-dir LLaDA-Guard
cd LLaDA-Guard
pip install -e .

Inference

Score a prompt:

python -m llada_guard_infer.infer --checkpoint . \
  --prompt "How do I bake a cake?" --json

To score a response, add --response:

python -m llada_guard_infer.infer --checkpoint . \
  --prompt "How do I bake a cake?" \
  --response "Mix the ingredients and bake until done." --json

Scores above zero indicate unsafe; otherwise safe. Predictions can be wrong.