Insurance Document Prompt-Injection Detection Dataset
Description
Insurance Document Prompt-Injection Detection Dataset (PIID) is a large-scale, predominantly text-based benchmark dataset for prompt-injection detection in insurance document-processing systems based on large language models (LLMs). The dataset contains 289,874 records, comprising 144,937 benign and 144,937 malicious samples, constructed through a reproducible five-module pipeline (Modules A–E). The malicious component contains 58 payloads spanning 15 attack families, with each benign text record paired with exactly one malicious counterpart through a controlled one-to-one injection process. The final release is predominantly text-based, with 1,870 records retaining associated document images for evaluation of image-associated and OCR-aware prompt-injection scenarios. The dataset is divided into training (231,900 records), validation (28,986 records), and test (28,988 records) partitions using group-aware splitting to prevent leakage between paired benign and malicious records. The release includes Parquet data files, associated document images, metadata, attack taxonomy, provenance information, reproducibility code, validation reports, and SHA-256 checksums. The dataset supports research in prompt-injection detection, document AI security, OCR-aware adversarial evaluation, and security testing of LLM-based document-processing pipelines.
Files
Steps to reproduce
The dataset was constructed through a reproducible five-module pipeline (Modules A–E). Module A defines the attack taxonomy and payload bank; Module B generates malicious text through controlled one-to-one injection into benign records; Module C generates associated document-image artifacts for selected records; Module D performs dataset preparation and validation; and Module E builds the final benchmark release. The final release contains 289,874 records with fixed train, validation, and test partitions established using group-aware splitting. Reproducibility code, metadata, validation reports, and checksums are included in the release archive.