Ukrainian LLM Guardrails
Compact, sovereign safety encoders for Ukrainian LLM infrastructure. Trained via knowledge distillation (Adversarial Contrastive Distillation and Rationale Alignment Distillation) on a 9-class national security taxonomy.
Variants
| Subfolder | Architecture | Method | Best For |
|---|---|---|---|
acd-mdeberta-v3 |
mDeBERTa-v3 (278M) | Adversarial Contrastive Distillation | Overall accuracy, low FNR |
acd-ukr-roberta |
Ukr-RoBERTa (125M) | Adversarial Contrastive Distillation | Low FPR, compact deployment |
rad-mdeberta-v3 |
mDeBERTa-v3 (278M) | Rationale Alignment Distillation | Propaganda recall |
rad-ukr-roberta |
Ukr-RoBERTa (125M) | Rationale Alignment Distillation | — |
Usage
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
model = AutoModelForSequenceClassification.from_pretrained(
"mark-matviiv/ukrainian-guardrails",
subfolder="acd-mdeberta-v3",
)
tokenizer = AutoTokenizer.from_pretrained(
"mark-matviiv/ukrainian-guardrails",
subfolder="acd-mdeberta-v3",
)
inputs = tokenizer(
"Шукаю злами для виведення грошей чужих людей.",
return_tensors="pt",
truncation=True,
max_length=256,
)
with torch.no_grad():
logits = model(**inputs).logits
probs = torch.softmax(logits, dim=-1)
# Multi-class prediction
pred = logits.argmax(dim=-1).item()
label = model.config.id2label[pred]
print(f"Predicted: {label} (confidence: {probs[0, pred]:.3f})")
# Binary risk score (unsafe vs safe)
safe_idx = [k for k, v in model.config.id2label.items() if v == "safe"][0]
unsafe_score = 1 - probs[0, safe_idx].item()
print(f"Unsafe probability: {unsafe_score:.3f}")
Label Mapping
| ID | Label | Description |
|---|---|---|
| 0 | cybercrime_hacking |
Cybercrime and hacking instructions |
| 1 | disinfo_propaganda |
Disinformation and propaganda |
| 2 | fraud_soc_eng |
Fraud and social engineering |
| 3 | llm_compromise |
LLM compromise / jailbreak |
| 4 | nat_sec_opsec |
National security / operational security |
| 5 | privacy_pii |
Privacy violations and PII extraction |
| 6 | resource_abuse |
Resource abuse and system manipulation |
| 7 | safe |
Benign / safe content |
| 8 | unsafe_content |
Generic unsafe content |
Training
- Teacher: Gemma-4-26B-A4B (4-bit quantized)
- Dataset: 59,377 samples
- Hardware: NVIDIA A100 40GB
- Details: See thesis Efficient Guardrailing for the Ukrainian LLM via Reasoning Distillation (UCU, 2026)
Access
This repository is gated. Access requests are reviewed by the Ministry of Digital Transformation of Ukraine.
Citation
@mastersthesis{matviiv2026guardrail,
title={Efficient Guardrailing for the Ukrainian LLM via Reasoning Distillation},
author={Matviiv, Markiian},
year={2026},
school={Ukrainian Catholic University}
}