|
|
Využití velkých předtrénovaných jazykových modelů pro konfiguraci a podporu klinického informačního systému
Sova, Michal ; Burget, Radek (oponent) ; Rychlý, Marek (vedoucí práce)
Cílem této práce je seznámení se s podstatou a možným použitím velkých předtrénovaných jazykových modelů, seznámení se s možnostmi konfigurace klinického informačního systému FONS Enterprise a možnost jeho adaptace na konkrétní prostředí zákazníků. Práce nejprve představuje velké předtrénované jazykové modely a informační systém FONS Enterprise. Následně se zaměřuje na možnosti dotrénování modelů a implementaci metody RAG na datech z klinického systému. Implementace RAG architektury je realizována pomocí nástroje LangChain a LlamaIndex. Výsledky ukazují, že metoda RAG s modelem Gemma a embedding modelem bge-m3 poskytuje nejrelevantnější odpovědi, ale má potíže s porozuměním složitějších otázek. Metoda dotrénování modelu nepřináší očekávané výsledky, a to ani po úpravách parametrů trénování.
|
|
|
Detection of key information in emergency calls
Sarvaš, Marek ; Plchot, Oldřich (oponent) ; Schwarz, Petr (vedoucí práce)
Emergency calls are usually made under extremely stressful conditions, where callers often provide crucial information rapidly, making it difficult for emergency line agents to capture all details accurately. This can result in repeated questions about information that was already provided and cause delays in response times from emergency services. This work aims to mitigate this problem and potentially speed up the response of emergency services by deploying a neural network models for information extraction, specifically targeting the Named Entity Recognition (NER) task. This work explores various Transformer-based approaches for NER task, such as pre-trained encoder-only, encoder-decoder (sequence-2-sequence) and Large Language Models. The best models achieved state-of-the-art results on publicly available Czech NER datasets. In addition, new NER datasets were created from available recordings of real emergency calls and the corresponding metadata. The models were trained and evaluated on the created datasets successfully achieving reasonable performance in name and location extraction.
|
|
|
Large Language Models for Generating Code Focusing on Embedded Systems
Vadovič, Matej ; Nosko, Svetozár (oponent) ; Smrž, Pavel (vedoucí práce)
The goal of this work was to adapt a pre-trained language model for the purpose of generating code in the field of embedded systems. The work introduces a new dataset for fine-tuning code generation models, consisting of 50,000 pairs of source code and comments focused on embedded systems programming. This dataset is composed of collected source code from the GitHub platform. Two new language models for code generation, based on transformer architecture pre-trained models, were fine-tuned on the data of the new corpus. Model MicroCoder is based on the CodeLLaMA-Instruct 7B model, and during its fine-tuning, the QLoRA technique was used to minimize computational requirements. The second model, MicroCoderFIM, is based on the StarCoderBase 1B model and supports code infilling. The individual models were compared based on BLEU, CodeBLEU, ChrF++, and ROUGE-L metrics. Model MicroCoderFIM achieves the best adaptation results to the new task, with over 120% improvement in all measured metrics. The weights of the models along with the new dataset are freely accessible on a public repository.
|
|
| |
|
|
Sentiment Analysis of Czech and Slovak Social Networks and Web Discussions
Slúka, Dušan ; Doležal, Jan (oponent) ; Smrž, Pavel (vedoucí práce)
This bachelor’s thesis deals with the issue of extraction and analysis of data obtained from social networks to understand public opinion on various social topics. The goal is systematic categorization and interpretation of contents. The problem is solved through a platform for opinion extraction and automatic data classification, which allows the creation of thematic subcategories and sorting into them. The result of the work is a system that analyzes social networks and provides deeper insight into public opinion on social topics. The system enables organizations to better understand the dynamics of online discourse. The benefit of this work is the provision of a new tool for the analysis of social issues, which can serve the academic sphere as well as organizations from practice.
|
|
|
Human web browsing simulation
Doležal, Jáchym ; Setinský, Jiří (oponent) ; Hranický, Radek (vedoucí práce)
This work introduces a promising tool for automated web navigation and achieving specific goals based on decisions made by a Large Language Model using information from a current page. The results of the simulator with model GPT 4 Turbo demonstrate the tool’s effectiveness, achieving over 80% success in completing predefined goals. The results show the usability of this tool in real use cases.
|
|
|
Generating Documentation to Source Code in Python
Novosád, Juraj ; Nosko, Svetozár (oponent) ; Smrž, Pavel (vedoucí práce)
The aim of this work is to adapt selected language models on domain data and to develop a system that would allow their use on commonly available hardware. The models have been adapted to generate documentation for undocumented source code in the Python progra- mming language to follow the Google Style convention. A prerequisite of model adaptation was to obtain domain data and process it appropriately for the purpose of model fine-tuning. This work focuses on fine-tuning models with fewer than one billion parameters, for the sake of enabling inference even on commonly available hardware. Part of the work was to objectively evaluate the quality of the adapted models. For this reason, I developed a tool that evaluates the quality of the generated documentation on a selected corpus of models. The evaluation of the adapted models showed that they achieve comparable performance to multiply larger models for general tasks, such as gpt-3.5-turbo-0125. The result of this work is a server capable of horizontal scaling that integrates the capabilities of more than just the adapted models through an easy-to-use API.
|
|
|
Komunikační agent pro informace o Brně
Neprašová, Kateřina ; Fajčík, Martin (oponent) ; Smrž, Pavel (vedoucí práce)
Tato diplomová práce se zaměřuje na doménově specifické komunikační agenty, cílem je vytvořit funkčního komunikačního agenta pro turisty i místní obyvatele Brna, poskytujícího relevantní a aktuální informace podle individuálních potřeb uživatelů. Popisuje velké jazykové modely, analyzuje existující technologie pro doménově specifické komunikační agenty a jejich tvorbu. Soustředí se na vytváření znalostní báze a implementaci efektivního dialogového rozhraní s využitím generování s rozšířeným vyhledáváním (RAG), přičemž srovnává vybrané jazykové modely na různých úlohách.
|
|
|
Failure Modes of Large Language Models
Milová, Soňa ; Špelda, Petr (vedoucí práce) ; Střítecký, Vít (oponent)
Failure Modes of Large Language Models Soňa Milová Abstract Diploma thesis "The failure modes of Large Language Models" focuses on addressing failure modes of Large Language Models (LLMs) from the ethical, moral and security point of view. The method of the empirical analysis is document analysis that defines the existing study, and the process by which failure modes are selected from it and analysed further. It looks closely at OpenAI's Generative Pre-trained Transformer 3 (GPT-3) and its improved successor Instruct Generative Pre-trained Transformer (IGPT). The thesis initially investigates model bias, privacy violations and fake news as the main failure modes of GPT-3. Consequently, it utilizes the concept of technological determinism as an ideology to evaluate whether IGPT has been effectively designed to address all the aforementioned concerns. The core argument of the thesis is that the utopic and dystopic view of technological determinism need to be combined with the additional aspect of human control. LLMs are in need of human involvement to help machines better understand context, mitigate failure modes, and of course, to ground them in reality. Therefore, contextualist view is portrayed as the most accurate lens through which to look at LLMs as it argues they depend on the responsibilities,...
|
host ::
RSS.