license: apache-2.0
task_categories:
- text-classification
language: - en
tags: - finance
- security
- prompt-injection
- jailbreak
Aegis-Fin: Financial Prompt Injection Dataset
Dataset Description
This dataset was engineered to train and evaluate the Aegis hybrid machine learning firewall, designed to protect Large Language Models (LLMs) in the financial sector. It addresses the lack of domain specificity in standard prompt injection datasets and explicitly mitigates length-based spurious correlations.
Dataset Splits
The dataset is pre-split into train and test subsets to ensure zero data leakage during model evaluation.
- train: 10,300 prompts
- test: 1,818 prompts
Taxonomy and Stratification (Training Set)
To prevent the model from associating adversarial intent solely with prompt length, the training corpus is rigorously balanced:
- Standard Query (40%): Short, benign financial requests (e.g., balance checks).
- Adversarial Roleplay (30%): Verbose, complex social engineering and contextual mimicry.
- Direct Command (20%): Short, high-urgency system overrides and traditional AppSec payloads.
- Customer Support (10%): Verbose, multi-paragraph benign emails ("Long Safe" correspondence).
Source Data
Benign baselines were aggregated and isolated from established financial QA datasets (FiQA, FinQA, ConvFinQA). Adversarial distributions were aggregated from open-source repositories (deepset, JailbreakBench, TrustAIRLab) and augmented via synthetic programmatic generation using uncensored LLMs.