Can Small Language Models Distinguish Rusyn, Belarusian, and Ukrainian? An Empirical Study of Hallucination and Subsumption in Local 4B Language Models

Quantised 4B-parameter LLMs are usually marketed as an ideal choice for local tasks because of their multilinguality; however, some languages still remain poorly understood. In this research, three 4-bit (Q4_K_M) models are evaluated: Qwen3.5-4B, Ministral-3-3B-Instruct, and Nemotron-3.5-Bolt-E4B in four configurations on zero-shot language identification with proofs for Rusyn, Belarusian, and Ukrainian (N = 15 parallel sentences; 4 configurations × 15 sentences = 60 deterministic runs). Results show Ministral invents grammatical rules to justify languages wrongly labelled as Russian, while Nemotron scores 5/5 on both Ukrainian and Belarusian, but 0/5 on Rusyn by mistaking its reflexive particle for Belarusian. Finally, enabling chain-of-thought (<think>) in Qwen3.5-4B triggers unclosed reasoning and repetition loops (0% compliance, 69.80s), whereas disabling it (--reasoning-budget 0) achieves 86.7% accuracy (80% on Rusyn) in 9.60s. Code, dataset, and logs: https://github.com/VladyslavKlochko/east-slavic-lid-4b

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24354666
Primary Topic
Natural Language Processing Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Can Small Language Models Distinguish Rusyn, Belarusian, and Ukrainian? An Empirical Study of Hallucination and Subsumption in Local 4B Language Models

Vladyslav Klochko
5 links
Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
preprint

Can Small Language Models Distinguish Rusyn, Belarusian, and Ukrainian? An Empirical Study of Hallucination and Subsumption in Local 4B Language Models

Vladyslav Klochko
preprint en

Abstract

Quantised 4B-parameter LLMs are usually marketed as an ideal choice for local tasks because of their multilinguality; however, some languages still remain poorly understood. In this research, three 4-bit (Q4_K_M) models are evaluated: Qwen3.5-4B, Ministral-3-3B-Instruct, and Nemotron-3.5-Bolt-E4B in four configurations on zero-shot language identification with proofs for Rusyn, Belarusian, and Ukrainian (N = 15 parallel sentences; 4 configurations × 15 sentences = 60 deterministic runs). Results show Ministral invents grammatical rules to justify languages wrongly labelled as Russian, while Nemotron scores 5/5 on both Ukrainian and Belarusian, but 0/5 on Rusyn by mistaking its reflexive particle for Belarusian. Finally, enabling chain-of-thought (<think>) in Qwen3.5-4B triggers unclosed reasoning and repetition loops (0% compliance, 69.80s), whereas disabling it (--reasoning-budget 0) achieves 86.7% accuracy (80% on Rusyn) in 9.60s. Code, dataset, and logs: https://github.com/VladyslavKlochko/east-slavic-lid-4b

Zenodo (CERN European Organization for Nuclear Research)
Varna University of Management (BG)
Natural Language Processing Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon