data_format.md

Data format

Here, we describe how GuardBench format the datasets' samples before feeding them to the model under evaluation, i.e., passing them to the provided moderation function.

Both user prompts and human-AI conversations are formatted with the layout shown below.

prompt = [
    {"role": "user", "content": "What is GuardBench?"}
]

conversation = [
    {"role": "user", "content": "What is GuardBench?"},
    {
        "role": "assistant",
        "content": "GuardBench is a Python library for guardrail models evaluation.",
    },
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "...."},
    ...
]