Primjena velikih jezičnih modela u razvoju chat bot asistenta u pivarstvu

master's thesis

• Title (english): Applying large language models in developing a chatbot assistant in brewing
• Author: Sven Almer
• Mentor: Domagoj Vlah (mentor)
• Committee member: Domagoj Vlah (predsjednik povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Committee member: Dario Bojanjac (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-09-16, Croatia

Abstract

U ovom radu predstavljen je problem izrade chatbota prilagođenog području pivarstva. Cilj rada je izrada takvog sustava koji je sposoban generirati, interpretirati i modificirati pivske recepte. Obavljen je pregled dostupnih velikih jezičnih modela te je objašnjena njihova arhitektura i način djelovanja, konačno je odabran model LLaMA 3.1 za implementaciju. Nad tabularnim skupom podataka odrađena je obrada, analiza i pretvorba u format prikladan za učenje velikih jezičnih modela. Model je treniran nad skupom podataka korištenjem metoda optimizacije: kvantizacije i LoRA-e. Opisana je funkcija pogreške korištena kod evaluacije modela, dani su primjeri interakcije s modelom te je model evaluiran. Za demonstraciju i interakciju s modelom implementirana je jednostavna iOS mobilna aplikacija temeljena na funkcionalnosti radnog okvira Core ML, koji omogućava lokalnu pohranu i izvođenje modela.

Abstract (english)

This thesis approaches the problem of implementing a chatbot assistant intended for use in homebrewing. The goal of this thesis is constructing a system based on large language models capable of generating, interpreting and modifying beer recipes. An overview of available and popular large language models was provided, along with a deep dive into their architecture and way of working. Finally, LLaMA 3.1 was selected as the model to be fine-tuned. A tabular dataset of beer recipe was cleaned, processed and transformed into the appropriate format for training LLM-s. The model was finetuned using quantization and LoRA. An error function was constructed for evaluating the resulting model, examples of interaction with the model were provided, and the model was evaluated. A simple iOS chat application was implemented for demonstrating model interaction, using the Core ML framework, which enables local model storage and inference.

• Keywords: veliki jezični modeli; LLaMA 3.1; pivarstvo; fine-tuning; LoRA
• Keywords (english): large language models; LLaMA 3.1; brewing; fine-tuning; LoRA
• Language: croatian
• URN:NBN: urn:nbn:hr:168:125157
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2024-11-07 05:35:30