Skip to content

Dataset extension

You can add your prompts to a scan as local CSV files, or add a new dataset to the registry.

Add local CSV files

  1. Make a CSV file with a column that has the name prompt. Each row is one attack prompt.
  2. Put the file in the datasets directory in the current directory.
  3. Start the server or the scan from the same directory.
mkdir -p datasets
cp my_prompts.csv datasets/

The tool reads all .csv files in datasets when it starts. It shows the names of the files in the log.

NOTE: The tool reads only CSV files. If a file does not have a prompt column, the tool shows a warning and ignores the file.

Add a dataset to the registry

The registry is the list REGISTRY in agentic_security/probe_data/__init__.py. Add one entry for each dataset:

{
    "dataset_name": "markush1/LLM-Jailbreak-Classifier",
    "num_prompts": 1119,
    "tokens": 19758,
    "approx_cost": 0.0,
    "source": "Hugging Face Datasets",
    "selected": True,
    "dynamic": False,
    "url": "https://huggingface.co/markush1/LLM-Jailbreak-Classifier",
},

The loaders in agentic_security/probe_data/data.py return a ProbeDataset. Refer to Probe data.

Transformed datasets

The tool can make new datasets from the loaded prompts. It encodes a sample of the prompts with each function in Stenography functions.