This thesis examines the use of Large Language Models (LLMs) in the financial sector, highlighting the risks and necessary safety measures for their application in financial services. As these models become more common in various financial tools, they bring both new opportunities and significant challenges, such as potential errors in financial advice and privacy issues. This work introduces a detailed safeguarded framework designed to improve the reliability, security, and ethical use of LLMs in financial applications. The framework includes specific safety features like checking user inputs, detecting incorrect information, and preventing security breaches to tackle these challenges effectively. Using quantitative testing benchmarks and case studies with a financial chatbot, this thesis shows that this framework helps reduce operational risks and increases trust among users. The results show that while LLMs already have some built-in safety features, adding tailored security measures greatly strengthens these systems against complex threats. This study advances the discussion on AI safety in financial settings and provides a practical guide for implementing strong safety measures that ensure reliable and ethical financial services.

In manufacturing, uptime and system reliability are paramount, placing high demands on automation technologies such as robotic systems. Failures in these systems cause considerable disruptions and incur significant costs. Traditional troubleshooting methods require extensive manual analysis by experts of log files, system data, application information, and problem descriptions. This process is labor-intensive and time-consuming, often resulting in prolonged downtimes and increased customer dissatisfaction, leading to heavy financial losses for companies. This research explores the application of Large Language Models (LLMs) like MistralLite and Mixtral-8*7B to automate root cause prediction from log data. We employed various fine-tuning methods, including full fine-tuning, Low-Rank Adaptation (LoRA), and Quantized Low Rank Adaptation (QLoRA), on these decoder-only models. Beyond using perplexity as an evaluation metric, the study incorporates GPT-4 as-a-judge to assess model performance. Additionally, the research uses complex prompting techniques to aid in the extraction of root causes from problem descriptions using GPT-4 and utilizes vector embeddings to analyze the importance of features in root cause prediction.  The findings demonstrate that LLMs, when fine-tuned, can assist in identifying root causes from log data, with the smaller MistralLite model showing superior performance compared to the larger Mixtral model, challenging the notion that larger models are inherently better. The results also indicate that different training adaptations yield varied effectiveness, with QLoRA adaptation performing best for MistralLite and full fine-tuning proving most effective for Mixtral. This suggests that a tailored approach to model adaptation is necessary for optimal performance. Additionally, employing GPT-4 with Chain of Thought (CoT) prompting has demonstrated the capability to extract reasonable root causes from solved issues using this technique. The analysis of feature vector embeddings provides insights into the significant features, enhancing our understanding of the underlying patterns and relationships in the data.

This research integrates LLMs with RAG (Retrieval-Augmented Generation) to develop a conversational interface allowing users to post queries and ask questions from a website. It compares the LLMRAGmethodwith traditional site search functionality to determine which method users perceive as better, specifically regarding response quality and response time. The perceived results for response quality and response time were evaluated under the null hypothesis that there is no difference between the two methods. The study showed that the LLM RAG method was perceived as better in terms of response quality, and those results were significant. However, for response time, the traditional site search method was perceived as better, but the results were not significant, so the null hypothesis could not be rejected. Overall, the integration of LLMs with RAG frameworks promises to enhance information retrieval systems on digital platforms.

Programming by voice can be a viable option for programmers who suffer from physical disabilities such as Repetitive Stress Injury (RSI), but study shows there are still challenges and limitations in the area, from the requirement of syntax-specific language to inaccuracy of speech recognition. We aim to address the potential benefit and performance of gradually integrating ChatGPT with its versions 3.5/4.0 and how to best communicate (speech optimization) to see if the area of programming by voice can be improved in terms of reduced vocal and cognitive load. We present a tool named ChatGPT VCG which contains two approaches, a traditional keyword interpretation approach using the Serenade tool as a code generator and a direct code generation approach using ChatGPT to generate code directly. Using test tasks written in Java, and the Serenade tool as a base measurement, the two approaches are compared and measurements such as the number of words and code characters generated are collected and analyzed. The result indicates that integrating ChatGPT allows the user to circumvent the required syntax-specific language. The direct code generation using ChatGPT 4.0 performed the best with a 4.2 times reduction in the required number of words compared to standard Serenade, while the keyword-based approach at worst shows a 14% increase. Speech optimization shows performance can be further increased by reducing or removing superfluous grammar and instead only providing the relevant information in commands. The study concludes that integration of ChatGPT can improve performance, and the speech can be optimized to reduce the number of words, but often at the cost of a conversational speech pattern.

I denna studie presenteras en prototyp för anonymisering av känslig information itextdokument, med syfte att möjliggöra säker interaktion med stora språkmodeller(LLM:er), såsom ChatGPT. Prototypen erbjuder en plattform där användare kanladda upp dokument för att anonymisera specifika känsliga ord. Efter anonymiseringkan användare ställa frågor till ChatGPT baserat på det anonymiserade innehållet.Prototypen återställer de anonymiserade delarna i svaren från ChatGPT innan de visas för användaren, vilket säkerställer att känslig information förblir skyddad underhela interaktionen.I studien används metoden Design Science Research in Information Systems (DSRIS). Prototypen utvecklas i Java och testas med påhittade dokument, medan enkätsvar samlasin för att utvärdera användarupplevelsen.Resultaten visar att prototypens funktioner fungerar väl och skyddar känslig information vid interaktionen med ChatGPT. Prototypen har utvärderats med hjälp av svarfrån enkäten som dessutom tar upp förbättringsmöjligheter.Avslutningsvis visar studien att det är möjligt att anonymisera textdokument effektivt och samtidigt få korrekt och användbar feedback från ChatGPT. Trots vissa begränsningar i användargränssnittet på grund av tidsramen visar studien på potentialför säker datahantering med ChatGPT. / This study presents a prototype for anonymizing sensitive information in text documents, with the aim of enabling secure interactions with large language models(LLMs) such as ChatGPT. The prototype offers a platform where users can uploaddocuments to anonymize specific sensitive words. After anonymization, users canpose questions to ChatGPT based on the anonymized content. The prototype restores the anonymized parts in the responses from ChatGPT before they are displayed to the user, ensuring that sensitive information remains protected throughoutthe entire interaction.The study uses the Design Science Research in Information Systems (DSRIS)method. The prototype is developed in Java and tested with fabricated documents,while survey responses were collected to evaluate the user experience.The results show that the prototype's functionalities work well and protect sensitiveinformation during interaction with ChatGPT. The prototype has been evaluated using survey responses that also highlight opportunities for improvement.In conclusion, the study demonstrates that it is possible to effectively anonymizetext documents while obtaining accurate and useful feedback from ChatGPT. Despite some limitations in the user interface due to the timeframe, the study showspotential for secure data handling with ChatGPT.

Programming education is on the rise, leading to an increase in learning resources needed for universities and online courses. Questions are crucial for promoting good learning, and providing students with ample practice opportunities. Learning a subject relies heavily on a structured progression of topics and complexity. Yet, creating numerous questions has been proven to be a time-consuming task. Recently the technology world has been introduced to Generative AI (GenAI) systems using Large Language Models (LLMs) capable of generating large amounts of text and performing other text-related tasks. How can GenAI be used to solve problems related to creating learning materials while ensuring good quality? This study aims to investigate how well GenAI can create a sense of progression in a set of programming questions based on different prompt strategies. The method involves three question-generation cases using Chat-GPT API. Then, a qualitative evaluation of questions complexity, order, and quality is conducted. The first case aims to be the most simple way of asking Chat-GPT to generate 10 MCQs about a specific topic. The second case introduces defined complexity levels and desires of logical order and progression in complexity. The final case is the more advanced prompt building upon the second case along with a skill map as inspiration to the LLM. The skill map is a structured outline that highlights key points when learning a topic. According to the results, providing more instructions along with a skill map had a better impact on the progression of questions generated compared to a simpler prompt. The first case prompt still resulted in questions with good order but lacking in increasing complexity. The results indicate that while GenAI is capable of creating questions with a good progression that could be used in a real teaching context, it still requires quality control of the content to find outliers. Further research should be done to investigate optimal prompts and what constitutes a good skill map. / Programmeringsutbildningar blir allt fler, vilket leder till en ökning av behovet för lärresurser för universtitet och onlinekurser. Frågor är avgörande för att främja bra lärande och ge eleverna övningsmöjligheter. Att lära sig ett ämne är starkt beroende av en strukturerad progression av ämnen och komplexitet. Men att skapa många frågor har visat sig vara en tidskrävande uppgift. Nyligen har teknikvärlden introducerats till Generativa AI (GenAI)-system som använder Stora språkmodeller (LLM) som kan generera stora mängder text och utföra andra textrelaterade uppgifter. Hur kan GenAI användas för att lösa problem relaterade till att skapa läromedel samtidigt som man säkerställer en god kvalitet? Denna studie syftar till att undersöka hur väl GenAI kan skapa en känsla av progression i en uppsättning programmeringsfrågor baserade på olika prompt strategier. Metoden använder tre olika sätt att generera frågor med hjälp av Chat-GPTs API. Därefter genomförs en kvalitativ utvärdering av frågornas komplexitet, ordning och kvalité. Det första sättet syftar till att vara det enklaste sättet att be Chat-GPT att generera 10 flervalsfrågor om ett specifikt ämne. Det andra fallet introducerar definierade komplexitetsnivåer och önskemål om logisk ordning och progression i komplexitet. Det sista fallet är den mer avancerade prompten som bygger på det andra fallet tillsammans med en färdighetskarta som inspiration. Färdighetskartan är en strukturerad disposition av ett ämne som lyfter fram nyckelpunkter när man lär sig ett ämne. Resultaten visade att tillhandahålla fler instruktioner tillsammans med en färdighetskarta hade en bättre inverkan på progressionen av de genererade frågorna jämfört med det första sättet. Den första prompten resulterade fortfarande i frågor med god ordning men som saknade stegrande komplexitet. Resultaten indikerar att även om GenAI kan skapa frågor med god progression som skulle kunna användas i ett verkligt undervisningssammanhang, så krävs fortfarande en kvalitetskontroll av innehållet för att hitta felaktigheter. Ytterligare forskning bör göras för att undersöka optimala prompt och hur en bra färdighetskarta bör se ut.

In the ever-evolving landscape of Software Development, the demand  for more efficient, scalable, and automated processes is paramount. The advancement of Generative AI has unveiled new avenues for innovative approaches to address this demand. This thesis explores one such avenue through the use of Multi-Agent Systems  combined  with Large Language Models (LLMs) to automate tasks within the development lifecycle. The thesis presents a structure for designing and developing an LLM-based multi-agent application by encompassing agent design principles, strategies for facilitating multi-agent collaboration, and providing valuable insights into the selection of an appropriate agent framework. Furthermore, the thesis showcases the developed application in its problem-solving capabilities with quantitative benchmarking results. Additionally, the study demonstrates practical implementations through examples of real-world applications. This study demonstrates the potential of utilizing LLM-based multi-agent systems in enhancing software development efficiency, offering companies a promising and powerful tool for streamlining Software Engineering workflows. / I den ständigt föränderliga världen av mjukvaruutveckling är behovet av mer effektiva, skalbara, och automatiserade metoder av stor betydelse. Framstegen inom generativ AI har öppnat nya möjligheter för utveckling av metoder för detta ändamål. Denna studie undersöker en sådan möjlighet genom användning av multi-agent system i samband med stora språkmodeller (Large Language Models, LLM) för automatisering av uppgifter inom utvecklingslivscykeln. Studien presenterar en struktur för design och utveckling av en LLM-baserad multi-agent applikation genom att bearbeta agentdesign och strategier för att underlätta samarbete mellan flera agenter och ge värdefulla insikter i valet av ett lämpligt agent-ramverk. Vidare demonstrerar studien den utvecklade applikationens problemlösningsförmåga med kvantitativa benchmark-resultat. Utöver detta inkluderar studien även exempel på genererade applikationer för att presentera konkreta exempel på implementeringar. Denna studie visar potentialen av att använda LLM-baserade multi-agent system för att förbättra effektiviteten inom mjukvaruutveckling, och erbjuder företag ett lovande och kraftfullt verktyg för effektivisering av arbetsflöden inom mjukvaruteknik.

The popularity of Large Language Models (LLMs), such as ChatGPT-4 (GPT), is increasing rapidly. Simultaneously, anxiety is contributing to the global burden of disease. To date, little research is performed on how LLMs portray anxiety. This study aimed to investigate the quality of anxiety portrayals generated by LLMs by “How well do large language models produce realistic anxiety depictions?”, using a mixed-methods approach. Realism was defined as perceived realism and alignment with scientific literature. The quantitative analysis involved 42 participants (age M = 30) randomly assigned to a questionnaire representing one of four anxiety levels. Participants rated perceived level and perceived realism of anxiety descriptions in human-written and GPT-modified texts. Results revealed: GPT-generated anxiety descriptions were perceived significantly more realistic compared to human-written texts; regardless of anxiety level; participants were less accurate in rating anxiety levels of GPTgenerated stories when the levels differed largely from that of the human-written story. The qualitative narrative analysis provided deeper insights into how realistic GPT and a GPT based on the cognitive framework of anxiety (CGPT) depicted anxiety by assessing how well depictions aligned with scientific literature. GPT and CGPT effectively included general features of anxiety in depictions. CGPT focused more on cognitive thought patterns, but neither fully depicted distinctions between different levels of anxiety. Overall, findings suggest LLMs do well in producing realistic representations of anxiety, but fail to fully depict various levels of anxiety. The study contributes to understanding potential applications of LLMs in psychological contexts, including management training, therapeutic settings, and literature. Keywords: LLM, ChatGPT, anxiety, realism, cognitive framework / Språkmodeller (LLMs) som ChatGPT-4 (GPT) blir allt mer populära. Samtidigt påverkar ångest folkhälsan. Hittills har enbart ett fåtal studier utrett hur LLMs skildrar ångest. Syftet med denna studie var att undersöka kvaliteten på ångestskildringar genererade av LLMs med en blandad metodansats genom frågan “Hur väl producerar LLMs realistiska ångestskildringar?”. Realism definierades som upplevd realism och överensstämmelse med vetenskaplig litteratur. I den kvantitativa analysen tilldelades 42 deltagare (ålder M = 30) en enkät som representerade en av fyra ångestnivåer. Deltagarna bedömde upplevd ångestnivå och upplevd realism av ångestbeskrivningar i berättelser skrivna av människor och GPTgenererade berättelser. Resultaten visade att deltagarna uppfattade GPT-genererade ångestbeskrivningar som signifikant mer realistiska, oavsett ångestnivå, men var mindre korrekta i att bedöma ångestnivå i GPT-genererade berättelser när nivån skiljde sig mycket från den människoskrivna versionen. Den kvalitativa narrativa analysen gav djupare insikter i hur realistiskt GPT och en GPT baserad på det kognitiva perspektivet av ångest (CGPT) skildrade ångest genom att jämföra skildringarna med vetenskaplig litteratur. Skildringarna av GPT och CGPT överensstämde i stort sett med litteraturen. CGPT fokuserade mer på kognitiva tankemönster, men ingen modell lyckades fullt ut skildra skillnaderna mellan olika ångestnivåer. Sammantaget tyder resultaten på att LLMs är bra på att skapa realistiska representationer av ångest, men misslyckas med att fullt ut skildra olika ångestnivåer. Studien bidrar till kunskapen om potentiella tillämpningar av LLMs i psykologiska kontexter, såsom ledarskapsutbildning, terapi och inom litteratur. Nyckelord: LLM, ChatGPT, ångest, realism, kognitiv ångest

Increasing productivity, reducing cost, reducing soil damage, reducing the impact of harvesting on standing tree and regeneration are all very important objectives in ground skidding system in the management of the Hyrcanian forest. The research carried out to obtain these objectives included four logging methods, tree length method (TLM), long length method (LLM), short length method (SLM), and wood extraction by mule (mule) in northern Iran. In order to determine the cost per unit, time study techniques were used for each harvesting method, time study data are shifted to logarithmic data based on 10. On the basis of the developed models simulated, 11 skidding turns are simulated and the unit cost are estimated depending on the diameter of the log (DL), skidding distance (SD), and the winching distance (WD) for 11 different cycles with TLM, LLM and SLM. The results showed that on average, the net costs per extraction of one cubic meter of wood were 3.06, 5.69, 6.81 and 34.36 €/m3 in TLM, LLM, SLM and mule. The costs depending on diameter of log (DL), skidding distance (SD) and winching distance (WD) showed that the most economical alternative for Northern Iran is TLM. In the cut-to-length system, the costs of both alternatives LLM, SLM were significantly dependent on DL. , thus the result of this study suggests that as long as the diameter of the felled trees is less than 40 cm, the cut-to-length system is not an economical alternative, whilst the cut-to-length method can be applied for trees with a diameter more than 40 cm. Where diameters are more than 40 cm TLM it is more economical than SLM, however it was not significantly different. Depending on SD in short skidding distance SLM is preferable to LLM but in cases of long skidding distance LLM is more economical than SLM. The winching distance affect was not a factor on cost. To assess the damage on seedlings and standing trees a 100% inventory method was employed in pre-hauling and post-hauling, alongside of skidding trails, winching strips and mule hauling with a 12m width. To chose the best alternative depending on standing damage the Analysis of multiple criterial approval (MA) was applied. The amount of trees damaged by winching operation were 11.89% in TLM, 14.44% in LLM 27.59%, SLM and 0 stem and by skidding operation were 16.73%, 3.13% and 8.78% of total trees in TLM, LLM and SLM. In the winching area about 14%, 20%, 21% and 6 % of the total regeneration was damaged by TLM, LLM, SLM and mule and the skidding operation damaged 7.5% in TLM, 7.4 % LLM and 9.4% in SLM. The friendliest alternative to residual standing was mule but in manual method (where the wood extraction is done by skidder) MA showed that the best alternative depending on residual damage is LLM. To determine the degree of soil compaction a core sampling technique of bulk density was used. Soil samples collected from the horizontal face of a soil pit at 10 cm depth soil core, at 50m intervals on skid trials, in winching strips and control are (no vehicles pass) a soil sample was taken at 10m intervals in the hauling direction of the mule. In order to determine the post-harvesting extent of disturbance on skidding trails by skidding operations, the disturbed widths were measured at 50 m intervals along the skid trails. In the winching area, where the winched logs created a streak of displaced soil, the width of the displaced streak was measured at 5 m interval along the winching strip. In mule hauling operations the width of a streak created by a mule foot track was measured at 10 m intervals. To compare increased average bulk density between alternatives one way The ANOVA, Duncan test and Dunnett t-test with a 95 % confidence level were used. A General linear model was applied to relate the increasing bulk density and the slope gradient. To realize the correlation between the increment of soil bulk density and the slope gradient and the correlation between the soil compaction and soil moisture content (%) The Pearson correlation test was applied. To choose the best alternative (in manual method) a MA test was applied again. The bulk density on the skidding trail increased 51 % for 30 skidding turn, 35 % for 31 skidding turn (one unloaded and one loaded pass) and 46% for 41 skidding turn. Results of ANOVA (p < 0.05) show significant differences of bulk density between alternatives. Duncan test and the Dunnett t-test indicated that the increasing soil bulk density was not significant between control samples and winching strip of TLM and extraction by mule samples. The general linear modeling and Pearson correlation test results indicated that the slope gradient had an insignificant effect on soil compaction, whilst the Pearson test indicates a medium negative correlation between soil compaction and percentage of soil moisture. By ground-based winching operation 0.07%, 0.03%, 0.05% and 0.002% of the total area and by ground based skidding operation 1.21%, 1.67%, 0.81% and 0.00% of total area was disturbed and compacted in TLM, LLM, SLM and mule. The Pearson correlation results show that the width of disturbed area was significantly influenced by the diameter of logs and length of logs (p ˂ 0.05), but there is no significant correlation between soil disturbance width and slope. The results of analysis of MA showed that soil compaction was not related to logging method but sensitivity analysis of MA shows that LLM and TLM are both preferable to SLM.

Increasing productivity, reducing cost, reducing soil damage, reducing the impact of harvesting on standing tree and regeneration are all very important objectives in ground skidding system in the management of the Hyrcanian forest. The research carried out to obtain these objectives included four logging methods, tree length method (TLM), long length method (LLM), short length method (SLM), and wood extraction by mule (mule) in northern Iran. In order to determine the cost per unit, time study techniques were used for each harvesting method, time study data are shifted to logarithmic data based on 10. On the basis of the developed models simulated, 11 skidding turns are simulated and the unit cost are estimated depending on the diameter of the log (DL), skidding distance (SD), and the winching distance (WD) for 11 different cycles with TLM, LLM and SLM. The results showed that on average, the net costs per extraction of one cubic meter of wood were 3.06, 5.69, 6.81 and 34.36 €/m3 in TLM, LLM, SLM and mule. The costs depending on diameter of log (DL), skidding distance (SD) and winching distance (WD) showed that the most economical alternative for Northern Iran is TLM. In the cut-to-length system, the costs of both alternatives LLM, SLM were significantly dependent on DL. , thus the result of this study suggests that as long as the diameter of the felled trees is less than 40 cm, the cut-to-length system is not an economical alternative, whilst the cut-to-length method can be applied for trees with a diameter more than 40 cm. Where diameters are more than 40 cm TLM it is more economical than SLM, however it was not significantly different. Depending on SD in short skidding distance SLM is preferable to LLM but in cases of long skidding distance LLM is more economical than SLM. The winching distance affect was not a factor on cost. To assess the damage on seedlings and standing trees a 100% inventory method was employed in pre-hauling and post-hauling, alongside of skidding trails, winching strips and mule hauling with a 12m width. To chose the best alternative depending on standing damage the Analysis of multiple criterial approval (MA) was applied. The amount of trees damaged by winching operation were 11.89% in TLM, 14.44% in LLM 27.59%, SLM and 0 stem and by skidding operation were 16.73%, 3.13% and 8.78% of total trees in TLM, LLM and SLM. In the winching area about 14%, 20%, 21% and 6 % of the total regeneration was damaged by TLM, LLM, SLM and mule and the skidding operation damaged 7.5% in TLM, 7.4 % LLM and 9.4% in SLM. The friendliest alternative to residual standing was mule but in manual method (where the wood extraction is done by skidder) MA showed that the best alternative depending on residual damage is LLM. To determine the degree of soil compaction a core sampling technique of bulk density was used. Soil samples collected from the horizontal face of a soil pit at 10 cm depth soil core, at 50m intervals on skid trials, in winching strips and control are (no vehicles pass) a soil sample was taken at 10m intervals in the hauling direction of the mule. In order to determine the post-harvesting extent of disturbance on skidding trails by skidding operations, the disturbed widths were measured at 50 m intervals along the skid trails. In the winching area, where the winched logs created a streak of displaced soil, the width of the displaced streak was measured at 5 m interval along the winching strip. In mule hauling operations the width of a streak created by a mule foot track was measured at 10 m intervals. To compare increased average bulk density between alternatives one way The ANOVA, Duncan test and Dunnett t-test with a 95 % confidence level were used. A General linear model was applied to relate the increasing bulk density and the slope gradient. To realize the correlation between the increment of soil bulk density and the slope gradient and the correlation between the soil compaction and soil moisture content (%) The Pearson correlation test was applied. To choose the best alternative (in manual method) a MA test was applied again. The bulk density on the skidding trail increased 51 % for 30 skidding turn, 35 % for 31 skidding turn (one unloaded and one loaded pass) and 46% for 41 skidding turn. Results of ANOVA (p < 0.05) show significant differences of bulk density between alternatives. Duncan test and the Dunnett t-test indicated that the increasing soil bulk density was not significant between control samples and winching strip of TLM and extraction by mule samples. The general linear modeling and Pearson correlation test results indicated that the slope gradient had an insignificant effect on soil compaction, whilst the Pearson test indicates a medium negative correlation between soil compaction and percentage of soil moisture. By ground-based winching operation 0.07%, 0.03%, 0.05% and 0.002% of the total area and by ground based skidding operation 1.21%, 1.67%, 0.81% and 0.00% of total area was disturbed and compacted in TLM, LLM, SLM and mule. The Pearson correlation results show that the width of disturbed area was significantly influenced by the diameter of logs and length of logs (p ˂ 0.05), but there is no significant correlation between soil disturbance width and slope. The results of analysis of MA showed that soil compaction was not related to logging method but sensitivity analysis of MA shows that LLM and TLM are both preferable to SLM.