Roblox/RobloxGuard-1.0 ? reverse-engineered prompt

Build me a simple Python tool that uses Roblox Guard 1.0 to check whether a user prompt or a generated response is unsafe.

I want to be able to point it at a config file, load the matching prompt template, run inference with the Roblox Guard model, and save the results to CSV. It should support evaluating a dataset from Hugging Face, compare the model prediction against the ground truth label when one is available, and also write a summary file with counts and basic metrics like precision, recall, F1, and false positive rate.

Please make it easy to swap in a different taxonomy or dataset by editing the config and prompt text, and keep the output format clear so I can inspect each example. If anything is unclear, look up the current model and dataset docs online and make the script work with them.