dataset_info:
features:

• name: chosen
  list:
  • name: content
    dtype: string
  • name: role
    dtype: string
• name: rejected
  list:
  • name: content
    dtype: string
  • name: role
    dtype: string
• name: system
  dtype: string
  splits:
• name: train
  num_bytes: 17504367
  num_examples: 2728
• name: eval
  num_bytes: 1935566
  num_examples: 304
  download_size: 9642063
  dataset_size: 19439933
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
  • split: eval
    path: data/eval-*

Prompt Injection Judge Dataset - Version 2 🛡️

Dataset Description

The Prompt Injection Judge Dataset (V2) is a preference-learning corpus designed exclusively for training System-2 generative AI security judges. It utilizes the ORPO (Odds Ratio Preference Optimization) algorithm to teach Large Language Models to deliberate, analyze, and definitively flag prompt injections, jailbreaks, and adversarial attacks in user inputs.

This dataset was entirely rebuilt to aggressively resolve formatting constraints in tools like Axolotl.

Key Technical Improvements for ML Devs

• Axolotl / PyArrow Compatible: The dataset strictly asserts an even number of message pairing turns (User -> Assistant). Earlier versions crashed Axolotl's prompt_strategies module with invalid number of turns because the System prompt was interleaved in the conversation tuple.
• Separated System Node: The meta-instruction is stripped out of the conversation turn arrays and placed securely into a native "system" key mapping to preserve conversational integrity for the chat_template.argilla loader.

Dataset Structure

The corpus strictly adheres to the following structural dictionary:

• system (string): The strict meta-instruction instructing the judge to perform System-2 security parsing.
• chosen (list of dicts): The preferred, secure conversation turn. Contains precisely two items: [{role: "user"}, {role: "assistant"}]. The assistant utilizes deep reasoning and successfully blocks/flags adversarial payloads.
• rejected (list of dicts): The adversarial or failed conversation turn. Contains the exact same user payload paired with a flawed assistant response (e.g. failing to detect the injection or executing the jailbreak).

Intended Output Format & Inference

Models trained on this dataset expect to emit responses using a System-2 chain-of-thought protocol, outputting scratchpad reasoning before emitting a strict validation JSON object.

Wait for this exact format during inference:

<think>
1. The user input attempts to bypass standard behavior by utilizing...
2. This is a classic example of role-playing jailbreak.
</think>
{"analysis": "Role-play jailbreak detected.", "verdict": "BLOCK"}

Usage with Axolotl

This repository natively integrates with Axolotl. Use the following data mapper block in your configuration YAML (chatml, llama-3, etc.) to begin fine-tuning via ORPO instantly. It automatically inherits the system key.

datasets:
  - path: hlyn-labs/prompt-injection-judge-dataset-v2
    type: chat_template.argilla
    split: train
eval_datasets:
  - path: hlyn-labs/prompt-injection-judge-dataset-v2
    type: chat_template.argilla
    split: eval