SCI-Guard: Experimental Data and Reproducibility Materials

Published: 9 November 2026| Version 1 | DOI: 10.17632/7tnzpfdj93.1
Contributor:
Qingkui Zeng

Description

This dataset supports the manuscript “SCI-Guard: Selective Counterfactual Intervention for Document-Level Prompt Injection Defense in Retrieval-Augmented Generation.” It contains controlled synthetic datasets, raw model-response records, processed evaluation results, frozen protocols, and reproduction code for the principal, operational-pilot, and independent competitive experiments. Model weights and caches are excluded.

Files

Steps to reproduce

1. Download and extract the archive. 2. Read README.md for the dataset structure and experiment boundaries. 3. Inspect principal_study, operational_pilot, and competitive_evaluation separately. Each directory contains the corresponding data, raw results, processed results, protocols, and code. 4. Verify file integrity using MANIFEST_SHA256.csv. 5. Consult environment/ for Python package versions and model digests. Raw JSONL records and processed JSON/CSV results can be inspected without rerunning the models. Full inference reproduction requires Python 3.11, a local Ollama installation, and the exact model versions specified in the protocols and environment records.

Institutions

Categories

Artificial Intelligence

Funders

  • Anhui Provincial Research Platform Special Research Project
    Grant ID: 2025tlxyptZD01

Licence