Zarządzanie Innowacyjne w Gospodarce i Biznesie nr 2(43)/2026

ISSN 2391-5129


Damian Kociemba https://orcid.org/0000-0002-6604-858X Independent scholar

e-mail: kociemba.damian.kd@gmail.com


Practical implementation of an AI model (Ministral 3B) in the logistics of dangerous goods

Praktyczne wdrożenie modelu sztucznej inteligencji Ministral 3B w logistyce towarów niebezpiecznych

https://doi.org/10.25312/ziwgib.888


Abstract

This article analyses the practical implementation of a local large language model (Ministral 3B) supported by Retrieval-Augment- ed Generation (RAG) in dangerous goods logistics, with a focus on verifying ADR exemptions, reducing AI hallucinations, and main- taining data privacy. The study uses a critical literature review and an empirical case study. The model was tested locally under strin- gent parameter constraints, including temperature 0.1 and strict sys- tem instructions. The analysis covered one comprehensive query concerning the transport of goods packed in limited quantities (LQ). Model responses were assessed for substantive compliance with ADR regulations and the commentary of a Dangerous Goods Safe- ty Adviser (DGSA), comparing outputs generated with and without RAG. The results indicate that the RAG-supported model synthe- sized legal information more accurately and substantially reduced hallucinations in the analysed case compared with the raw model, which fabricated non-existent legal structures. However, conserva- tive parameter settings may cause the omission of minor technical nuances. The findings suggest that local LLMs combined with RAG can support DGSA work, but only as auxiliary tools subject to con- tinuous expert verification.

Keywords: artificial intelligence, Large Language Models (LLM), Retrieval-Augmented Generation (RAG), ADR Agreement, danger- ous goods logistics, AI hallucinations


Streszczenie

Celem artykułu jest analiza praktycznego wdrożenia lokalnego du- żego modelu językowego (Ministral 3B) wspieranego architekturą RAG (Retrieval-Augmented Generation) w logistyce towarów nie- bezpiecznych, z naciskiem na weryfikację wyłączeń z przepisów umowy ADR przy jednoczesnym ograniczeniu zjawiska halucynacji AI i zachowaniu pełnej prywatności danych. W pracy zastosowano krytyczny przegląd literatury oraz metodę empirycznego studium przypadku. Model przetestowano w środowisku lokalnym przy ry- gorystycznych ograniczeniach parametrów (temperatura 0.1, ścisłe instrukcje systemowe). W ramach badania przeanalizowano jedno kompleksowe zapytanie dotyczące przewozu towarów pakowanych w ilościach ograniczonych (LQ). Odpowiedzi modelu oceniano pod kątem zgodności merytorycznej z przepisami ADR i komentarzem doradcy ds. bezpieczeństwa ADR (DGSA), porównując warian- ty z RAG i bez niego. Eksperyment wykazał, że model wspiera- ny przez RAG skutecznie syntetyzuje informacje prawne, niemal całkowicie eliminując zjawisko halucynacji, któremu uległ model

„surowy” (fabrykujący nieistniejące struktury prawne). Zauważono jednak, że bardzo zachowawcze ustawienia parametrów mogą pro- wadzić do pomijania drobnych niuansów technicznych przez algo- rytm. Integracja lokalnych modeli LLM z architekturą RAG stanowi bezpieczne, tanie i wysoce użyteczne narzędzie wspierające pracę doradców ds. bezpieczeństwa. Niemniej, ze względu na rygory- styczny charakter przepisów transportowych, sztuczna inteligencja musi pozostać jedynie narzędziem pomocniczym, podlegającym stałej, merytorycznej weryfikacji przez wykwalifikowany personel.

Słowa kluczowe: sztuczna inteligencja, duże modele językowe (LLM), Retrieval-Augmented Generation (RAG), umowa ADR, logi- styka towarów niebezpiecznych, halucynacje AI


Introduction

In an era of widespread fascination with Large Language Models (LLMs) such as GPT-4 or Claude, researchers often overlook the fact that in rigorously regulated areas – to which the logistics of dangerous goods undoubtedly belongs – the key factor for success is not the model’s “creativity”, but its deterministic precision. For dangerous goods safety advisers (DGSAs) and employees involved in operations related to the movement of dan- gerous goods, the priority is not generating multi-page essays, but obtaining an instan- taneous and error-free answer to a specific question regarding the regulations contained in the ADR Agreement.

This article analyses the practical implementation of the Ministral 3B model as an as- sistant in the process of verifying conditions for exemptions from regulations. The choice of a model with a relatively small scale (3 billion parameters) was dictated by pragma- tism – the desire to maintain full locality of the process on standard office hardware. This solution is dedicated to logistics companies that focus primarily on administrative and


operational tasks and do not possess advanced computing infrastructure, while simulta- neously prioritizing data security.

The following research hypothesis is formulated in this article: the use of the Minis- tral 3B language model in combination with RAG (Retrieval-Augmented Generation) architecture and rigorous parameter configuration allows for a tool with high substantive utility that effectively limits the phenomenon of AI hallucinations in the interpretation of fragments of ADR Agreement regulations, while maintaining full privacy of user data. To verify the formulated hypothesis, a set of tools was used to run the model in a local environment. A key aspect was the configuration of the model’s “decision variables” to maximally reduce its tendency to confabulate in favour of a faithful synthesis of the pro- vided material. These variables, along with a description of the tools, are included

in the final section of the publication.

To verify the research hypothesis and provide a reliable approach to the issues dis- cussed, this work utilizes a critical analysis of the literature and the empirical case study method. The structure of the article is designed to clearly guide the reader from theoretical foundations to conclusions from practical implementation. First, the evolution of artificial intelligence and its inherent weaknesses (such as the phenomenon of hallucinations) are outlined. Next, contemporary applications of AI in the logistics industry are discussed, serving as a starting point for the main research part of the study, which presents the test environment and conducts a comparative analysis of the results generated by the model supported by RAG architecture versus the model deprived of this context. The whole concludes with an evaluation of the obtained results.


The evolution of artificial intelligence

The fascination with machines mimicking human thought dates back to ancient times; how- ever, it was Aristotle who laid the foundations for modern logic by introducing the concept of syllogisms (Rai, 2024: 2). The formal birth of the field of artificial intelligence (AI) is dated to 1956, when during the Dartmouth Conference, American computer scientist John McCarthy first used this term (Alalaq, 2024: 3; Rai, 2024: 5–6). Over the following decades, the development of AI was primarily based on symbolic systems and so-called expert systems, such as the chemical DENDRAL or medical MYCIN from the turn of the 1960s and 1970s (Singh, Banerjee, Gosh, 2025: 4–5). These required rigorous, manual coding of thousands of decision rules, which made them useful in strictly defined areas, yet extremely rigid and incapable of independently understanding the nuances of natural language.

To realize the dynamics of contemporary technological progress, it is worth looking back just a few years. In an article published in 2019, the author of this article analysed the implementation of artificial neural networks in the process of creating autonomous vehicles. At that time, commercial applications of AI in transport and logistics were based on so-called Narrow Artificial Intelligence (Weak AI). These systems focused on mathematical machine learning – the laborious selection of weights for artificial


neurons to process sensory signals. Their task was to trigger pre-defined reactions, such as maintaining the vehicle in its lane or activating emergency brake assist (EBA). The re- search problems of that period centred on technical limitations of algorithms and the lack of legal regulations for machines making decisions on the road (Kociemba, 2019: 171, 176, 178–179, 185). These tools, although revolutionary, remained “blind” to complex context – the machine was able to stop a truck in front of an obstacle, but depending on the training of the algorithm (i.e., the presentation of a given event to the algorithm during the learning stage), independent and safe avoidance of said obstacle was not always possible.

The turning point that moved artificial intelligence from the era of simple image analysis and rigid algorithms to the era of fluent text understanding was the introduction of the Transformer architecture. The “self-attention” mechanism used within it allowed models to analyse vast portions of text in parallel and understand the context of words regardless of their position in a sentence. This discovery led to an unprecedented explosion of so-called Large Language Models (LLMs). In 2020, the GPT-3 model was presented with a scale of 175 billion parameters, which proved the ability for “zero-shot” learning – it was able to generate text, translate, and solve logical problems without the necessity of previously programming specific rules for it. The release of ChatGPT at the end of 2022 proved, in turn, that these models are capable of fluently conducting multi-turn conver- sations and reasoning in the face of ambiguity (Alalaq, 2024: 6; Rai, 2024: 18; Singh, Banerjee, Gosh, 2025: 6, 17–18, 21–22, 25).

Equally important, the years 2023–2024 brought a flood of powerful open-source models, such as the LLaMA or Mistral families, to which the Ministral 3B model analysed in this study belongs. They offered performance close to technological giants, but with drastically lower demand for computing power and the possibility of full control over data privacy (Singh, Banerjee, Gosh, 2025: 26, 29–30).

Thanks to this very evolution – from narrow algorithms physically controlling a vehicle in 2019 to lightweight, open (and constantly developed) generative models – it became possible to realize the implementation described in this article. Personnel involved in op- erations related to the transport of dangerous goods can gain an innovative supporting tool that optimizes the manual document search process. However, as further analysis will show, this technology – although it reduces cognitive debt and drastically shortens regulation verification time – does not eliminate the human factor, remaining a supporting tool that requires constant substantive verification by trained personnel.


Contemporary problems related to the development of AI

Entering the world of artificial intelligence, it is easy to succumb to initial fascination. This phenomenon is perfectly illustrated by what researchers describe as an adapted Dunning-Kruger effect in the context of using AI. Initially, the user is full of enthusiasm, believing that thanks to AI, they will achieve immediate and error-free results. Over time, however, they hit a wall – wasting time on verification, receiving fabricated data,


and beginning to understand the technology’s limitations. Only overcoming this “trough of disillusionment” leads to a conscious and mature treatment of the model simply as another tool (Salvagno, Taccone, Gerli, 2023: 1–2).

The most serious problem of generative artificial intelligence (LLMs) is so-called hal- lucinations. These occur when the model generates false, misleading, or entirely fabricated information, yet presents it with immense confidence and impeccable linguistic fluency (Joshi, 2025: 38). These are not mere typos. Models can create coherent-sounding but completely non-existent bibliographic references, fabricating well-known authors and publication titles (Salvagno, Taccone, Gerli, 2023: 1). In specialized fields such as law or medicine, the rates of such hallucinations can reach double digits. This problem stems from the architecture of the models themselves. LLMs are, by design, powerful statistical en- gines that generate text token-by-token (word-by-word), attempting to predict what should come next without true, human “understanding”. They lack grounding in reality. When faced with uncertainty, the model prefers to fill knowledge gaps with “plausible-sounding” information, prioritizing linguistic fluency over factual truth (Joshi, 2025: 39).

One must keep in mind that AI models are only as good as the data on which they were trained. In industrial and logistics applications, insufficient, noisy, or incorrect data fed into the model can lead to erroneous predictions, which in turn risks operational disruptions (Sinha, Lee, 2024: 4). At this point, researchers draw attention to an extremely intriguing aspect: AI hallucinations are largely an extension of “human hallucinations”. LLMs learn from massive datasets from the Internet, simultaneously absorbing human biases, stereo- types, cognitive errors, and even deliberate misinformation. An AI model simply serves as a digital mirror of human imperfection. If we provide it with a contaminated knowledge base, it will perpetuate and automate that imperfection (Tlili, Burgos, 2025: 68–69).

Additionally, artificial intelligence is exposed to so-called data drift and concept drift. Regulations, the working environment, and technologies are constantly changing. A model trained on static, historical data loses its precision over time because statistical relation- ships between variables no longer correspond to the new reality (Sinha, Lee, 2024: 7). The absence of mechanisms for continuous learning and adaptation means that an expert can quickly become a dilettante.

Given the above limitations – from the tendency to confabulate and the “inheritance” of human errors to the susceptibility to obsolescence – one might conclude that artificial intelligence is a tool too unpredictable for the business environment. Meanwhile, as researchers note, a complete rejection of this technology in the face of growing market pressure would be a mistake (Sinha, Lee, 2024: 1). The transport and logistics industry, operating on massive volumes of data and under constant pressure to optimize time and costs, now treats AI not as an optional curiosity but as a strategic necessity (Sadiku, Ajayi, Sadiku, 2026: 171, 172). The key to success is therefore not expecting absolute infallibility from algorithms, but building systems that mitigate the risks described above. Awareness of these technical “traps” constitutes the foundation for the proper design and implementation of solutions based on artificial intelligence.


Using AI in logistics

Although AI still struggles with problems, the benefits of its implementation in logistics are invaluable. Skilful risk management allows translating theory into measurable opti- mization.

Contemporary logistics relies on complex information systems. AI models are be- coming a key element of these architectures, taking them to a higher level of efficiency (Dyczkowska, Olkiewicz, Królikowski, 2025: 5199–5202, 5204; Sadiku, Ajayi, Sadiku,

2026: 173–174):

The implementation of artificial intelligence in logistics (not only of dangerous goods) is no longer a thing of the distant future, but a technological reality. The democratization of access to AI tools means that the barrier to entry for these innovations has drastically decreased. The real challenge for the industry today is not access to technology, but the skilful and responsible management of it. AI will not replace experts – it will only replace those experts who do not know how to use AI.


References

Alalaq A. (2024), The History of the Artificial Intelligence Revolution and the Nature of Gen- erative AI Work, “DS Journal of Artificial Intelligence and Robotics”, vol. 2(4), pp. 1–24.

Dyczkowska J., Olkiewicz M., Królikowski T. (2025), The use of artificial intelligence in lo- gistics information systems, “Procedia Computer Science”, no. 270, pp. 5198–5206.

Government Statement of 13 March 2023 on the entry into force of amendments to Annexes A and B to the Agreement concerning the International Carriage of Dangerous Goods by Road (ADR), done at Geneva on 30 September 1957 (Oświadczenie rządowe z dnia 13 marca 2023 r. w sprawie wejścia w życie zmian do załączników A i B do Umowy dotyczącej międzynaro- dowego przewozu drogowego towarów niebezpiecznych (ADR), sporządzonej w Genewie dnia 30 września 1957 r.).

Joshi S. (2025), Comprehensive Review of AI Hallucinations: Impacts and Mitigation Strate- gies for Financial and Business Applications, “International Journal of Computer Applications Technology and Research”, vol. 14(6), pp. 38–50.

Kociemba D. (2019), Wykorzystanie metod sztucznej inteligencji w tworzeniu pojazdów auto- nomicznych, “Zarządzanie Innowacyjne w Gospodarce i Biznesie”, vol. 28(1), pp. 171–188.

Rai D. (2024), Artificial Intelligence Through Time: A Comprehensive Historical Review, Tribhuvan University, Institute of Science and Technology, https://www.researchgate.net/ profile/Dilli-Hang-Rai/publication/385939923_Artificial_Intelligence_Through_Time_A_ Comprehensive_Historical_Review/links/676fb15600aa3770e0c2516f/Artificial-Intelli- gence-Through-Time-A-Comprehensive-Historical-Review.pdf [accessed: 13.10.2026].

Sadiku M., Ajayi S., Sadiku J. (2026), Artificial Intelligence in Logistics, “International Jour- nal of Trend in Scientific Research and Development (IJTSRD)”, vol. 10(1), pp. 171–178.

Salvagno M., Taccone F., Gerli A. (2023), Artificial intelligence hallucinations, “Critical Care” vol. 27, articele no. 180.

Singh G., Banerjee T., Ghosh N. (2025), Tracing the evolution of artificial intelligence: A review of tools, frameworks, and technologies (1950–2025), preprint, https://doi.org/10.31224/5802

Sinha S., Lee Y. (2024), Challenges with developing and deploying AI models and applications in industrial systems, “Discover Artificial Intelligence”, vol. 4, article no. 55.

Tlili A., Burgos D. (2025), AI Hallucinations? What About Human Hallucination?! Address- ing Human Imperfection Is Needed for an Ethical AI, “International Journal of Interactive Multimedia and Artificial Intelligence”, vol. 9(2), pp. 68–71.


About the Author

Damian Kociemba – graduate of engineering studies in transport and master’s studies in management. He is the co-author and au- thor of several academic and specialist publications in the fields of logistics, safety, and optimization. He holds a safety advisor cer- tificate for the transport of dangerous goods by road (ADR).


O autorze

Damian Kociemba – absolwent studiów inżynierskich na kierunku transport oraz magisterskich na kierunku zarządzanie. Współautor i autor kilkunastu publikacji naukowych i specjalistycznych z zakre- su logistyki, bezpieczeństwa i optymalizacji. Posiadacz świadectwa doradcy ds. bezpieczeństwa przewozu towarów niebezpiecznych w transporcie drogowym (ADR).


Ten utwór jest dostępny na licencji Creative Commons Uznanie autorstwa-Na tych samych warunkach 4.0 Międzynarodowe.

  150