Skip to content

FEAT Add Turkish Conversation Prompt-Injection dataset loader #2171

Description

@3nesdeniz

Is your feature request related to a problem? Please describe.

PyRIT's built-in seed datasets cover multilingual safety, jailbreaks, and over-refusal, but do not currently provide a Turkish prompt-injection dataset with paired benign boundary cases. This makes it difficult to test whether a defense distinguishes adversarial intent from legitimate Turkish requests that share the same topics and security-sensitive vocabulary.

Describe the solution you'd like

Add a remote loader for the public, CC BY 4.0 licensed Turkish Conversation Prompt-Injection Dataset.

The dataset contains 750 curated Turkish examples:

  • 600 legitimate user requests
  • 150 prompt-injection attacks
  • 150 benign boundary cases paired with related attacks
  • train, validation, and test splits
  • 10 attack families, including direct instruction override, system-prompt extraction, role-play jailbreaks, tool/agent abuse, indirect injection, RAG poisoning, memory poisoning, and obfuscation/code-switching

The proposed loader would:

  • return literal SeedPrompt objects
  • default to attack examples for red-team workflows
  • provide typed filters for attack/benign/all labels, dataset split, and attack family
  • preserve row-level provenance in seed metadata
  • expose attack families as harm categories for attack rows
  • support Hugging Face tokens consistently with existing remote loaders
  • include unit tests, built-in-dataset documentation, and a real-dataset smoke test

Describe alternatives you've considered, if relevant

Users can call datasets.load_dataset() directly and manually translate records into PyRIT seeds. That loses PyRIT dataset discovery, consistent metadata, typed filtering, and reproducible loading behavior.

A second option is to expose only attack rows. Keeping the paired benign boundary cases available is more useful because it supports both adversarial testing and over-refusal / false-positive analysis.

Additional context

Dataset source: https://huggingface.co/datasets/3nesdeniz/turkish-conversation-prompt-injection

Source repository: https://github.com/3nesdeniz/turkish-conversation-prompt-injection

The implementation will not publish or claim model benchmark results; it only integrates the documented dataset structure and provenance.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions