| RESTRICTED RESEARCH LICENSE |
|
|
| Version 1.0, 2026 |
|
|
| Copyright (c) 2026 Markiian Matviiv, Ukrainian Catholic University |
| In partnership with the Ministry of Digital Transformation of Ukraine |
|
|
| 1. PREAMBLE |
|
|
| This dataset ("Ukrainian LLM Safety Dataset") contains synthetic adversarial |
| prompts and security-sensitive classification data developed for the protection |
| of Ukrainian sovereign digital infrastructure. Given the sensitive nature of the |
| underlying threat models and the potential for misuse, access is strictly |
| controlled. |
|
|
| 2. GRANT OF RIGHTS |
|
|
| Subject to compliance with all terms herein, the Licensor grants you a |
| non-exclusive, non-transferable, royalty-free license to: |
|
|
| a) Access and use the dataset solely for academic research, defensive |
| security research, or educational purposes; |
| b) Train, fine-tune, or evaluate machine learning models for safety |
| classification tasks; |
| c) Publish research findings that reference this dataset, provided that |
| the raw dataset itself is not redistributed or made publicly available. |
|
|
| 3. RESTRICTIONS |
|
|
| You may NOT, without prior written approval from the Ministry of Digital |
| Transformation of Ukraine: |
|
|
| a) Use the dataset for any commercial, military, or intelligence |
| application; |
| b) Redistribute, sublicense, sell, or otherwise transfer the dataset |
| or substantial portions thereof to third parties; |
| c) Use the dataset to develop, improve, or benchmark offensive |
| tools, including but not limited to jailbreak generators, prompt |
| injection frameworks, or adversarial attack methodologies; |
| d) Publicly disclose or publish the raw prompt text, rationales, or |
| contrastive examples outside of aggregated statistical analyses; |
| e) Remove or obfuscate attribution to the Ministry of Digital |
| Transformation of Ukraine and Ukrainian Catholic University. |
|
|
| 4. DUAL-USE AND ETHICAL OBLIGATIONS |
|
|
| The dataset contains adversarial examples designed to test safety |
| boundaries. By accessing this data, you acknowledge that: |
|
|
| a) The adversarial content is provided strictly for defensive |
| classifier hardening and must not be used to bypass or undermine |
| existing safety systems; |
| b) Any models trained on this dataset should be deployed exclusively |
| for content moderation, guardrailing, or threat detection |
| purposes; |
| c) You will report any unintended harmful use or security |
| vulnerabilities discovered through this research to the Licensor |
| within 30 days. |
|
|
| 5. ATTRIBUTION |
|
|
| Any publication, presentation, or derivative work based on this dataset |
| must include the following citation: |
|
|
| Matviiv, M. (2026). Efficient Guardrailing for the Ukrainian LLM |
| via Reasoning Distillation. Master's Thesis, Ukrainian Catholic |
| University. Dataset provided in partnership with the Ministry of |
| Digital Transformation of Ukraine. |
| |
| 6. TERMINATION |
| |
| This license automatically terminates if you violate any of the terms |
| herein. Upon termination, you must destroy all copies of the dataset |
| and cease all use immediately. |
| |
| 7. DISCLAIMER |
| |
| THE DATASET IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS |
| OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF |
| MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NONINFRINGEMENT. |
| THE LICENSOR SHALL NOT BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER |
| LIABILITY ARISING FROM THE USE OF THE DATASET. |
| |
| |
