---
license: other
license_name: restricted-research-license
license_link: LICENSE
language:
  - uk
  - en
  - ru
pretty_name: Ukrainian LLM Safety Dataset
size_categories:
  - 10K<n<100K
tags:
  - text-classification
  - safety
  - guardrail
  - adversarial
  - ukrainian
  - multilingual
  - llm
---

# Ukrainian LLM Safety Dataset

## Dataset Summary

A national security-focused safety classification dataset for Ukrainian LLM infrastructure, developed in partnership with the Ministry of Digital Transformation of Ukraine. Contains adversarial prompts across a 9-class threat taxonomy with synthetic rationales and contrastive examples for knowledge distillation.

## Files

| File | Samples | Purpose |
|------|---------|---------|
| `train.jsonl` | 59,377 | Training corpus with contrastive triplets and rationales (synthetic + anonymized ministry logs) |
| `benchmark.jsonl` | 450 | Held-out adversarial benchmark (50 samples/class, human-verified) |

## Taxonomy (9 classes)

- `cybercrime_hacking`
- `disinfo_propaganda`
- `fraud_soc_eng`
- `llm_compromise`
- `nat_sec_opsec`
- `privacy_pii`
- `resource_abuse`
- `safe`
- `unsafe_content`

## Format

### `train.jsonl`

Each line is a JSON object with contrastive distillation fields:

```json
{
  "prompt": "User prompt text (the adversarial query)",
  "tag": "cybercrime_hacking",
  "reasoning": "Detailed explanation of why this prompt belongs to the category, including manipulation techniques identified",
  "positive": "Semantically equivalent paraphrase preserving the same safety label",
  "negative": "A semantically close but safe alternative on the same topic (hard negative)",
  "model_name": "Teacher model used for augmentation"
}
```

Fields:
- `prompt` (string): The original adversarial or benign user query
- `tag` (string): Safety category label from the 9-class taxonomy
- `reasoning` (string): Teacher-generated natural-language rationale explaining the classification logic
- `positive` (string): Paraphrased variant of `prompt` preserving the same `tag` (used for contrastive learning)
- `negative` (string): A safe query on the same topic with high lexical overlap but different intent (hard negative for ACD)
- `model_name` (string): Identifier of the teacher model that synthesized this sample

### `benchmark.jsonl`

Each line is a JSON object with minimal fields for evaluation:

```json
{
  "prompt": "User prompt text",
  "tag": "llm_compromise"
}
```

Fields:
- `prompt` (string): The adversarial query
- `tag` (string): Ground-truth safety category label

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("mark-matviiv/ukrainian-safety-dataset")
train = dataset["train"]
benchmark = dataset["benchmark"]
```

## Access

This dataset is gated. Access requests are reviewed by the Ministry of Digital Transformation of Ukraine.

## Citation

```bibtex
@mastersthesis{matviiv2026guardrail,
  title={Efficient Guardrailing for the Ukrainian LLM via Reasoning Distillation},
  author={Matviiv, Markiian},
  year={2026},
  school={Ukrainian Catholic University}
}
```