dataset_info:
features:

• name: text
  dtype: string
• name: label
  dtype:
  class_label:
  names:
  '0': clean
  '1': injection

splits:

• name: train
  num_bytes: 1427092
  num_examples: 13500
• name: test
  num_bytes: 158565
  num_examples: 1500
  download_size: 906659
  dataset_size: 1585657
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
  • split: test
    path: data/test-*

prompt-injection-15k

A dataset for training prompt injection / jailbreak detection classifiers.

15,000 examples labeled as either a prompt injection attempt (1) or a normal, benign message (0). Includes direct overrides, roleplay jailbreaks, system prompt extraction attempts, obfuscated attacks, multilingual examples, and benign edge cases (imperative tasks, casual use of words like "ignore" or "forget", short or minimal inputs).

Fields

• text: the user message
• label: 0 = clean, 1 = injection

Splits

• train: 13,500 examples
• test: 1,500 examples

Source

Synthetically generated using an LLM, then patched with targeted examples to address known classifier failure modes.

Usage

from datasets import load_dataset

ds = load_dataset("MistyozAI/prompt-injection-15k")