Institutional Repository of Vilnius University
Natūralios kalbos transliavimo į predikatų logikos kalbą galimybės, naudojant didžiuosius kalbos modelius /
Abstract
dc:descriptionThis bachelor's thesis investigates the applicability of open-source large language models to the task of translating natural language into first-order logic formulas. Traditional formalization methods, such as Montague grammar, are precise but impractical due to the significant manual effort required, while commercial models such as GPT-4, despite achieving high syntactic accuracy, raise concerns regarding data privacy and economic accessibility. The thesis reviews the existing literature on commercial GPT models and the specialized LogicLLaMA solution. The practical contribution consists of fine-tuning two base models from different generations, LLaMA-2-7B and LLaMA-3-8B, using the LoRA method on the MALLS dataset. The results were evaluated using four metrics (SynV, BSC, BLEU, LE) on the gold-standard FOLIO dataset. The experimental results demonstrated that LoRA fine-tuning significantly improved the performance of both models while fine-tuning less than (1 %) of all model parameters. The fine-tuned LLaMA-2-7B surpassed the direct translation stage of LogicLLaMA, while the fine-tuned LLaMA-3-8B outperformed all LogicLLaMA variants and achieved results practically identical to GPT-4 in the 5-shot prompting, despite being substantially smaller in size.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Bukauskas, Andenis,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308337155&prefLang=en_US