Suuret kielimallit relaatiotietokantojen opettajina
Ruusumaa, Severi (2024)
2024
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2024-12-12
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2024120810851
https://urn.fi/URN:NBN:fi:tuni-2024120810851
Tiivistelmä
Relational database systems are the most common database systems in current software industry. Relational databses saves data in tables that are build with rows and columns. SQLlanguage is used to both change the tables and observe the data in them. Therefore it is important for software engineers to learn SQL. SQL-queries contains SELECT-part. These queries are used to get data from tables. Queries result in another table. If user makes a mistake in their query that computers can’t understand, user gets an error message instead of a table as a result.
Large-Language Models have become really popular in recent years. Models like GPT-4o from OpenAi and Llama 3 from Meta are capable of understanding and producing natural language. These models have been trained with a vast amount of data. When the amount of data increases, also the answers of language model improves.
PG4N is a software that utilizes a Large-Language model if user makes and SQL-error. PG4N uses a cloud service to run the model. Previously used model has been taken off the cloud service since there has been added newer, better ones. Therefore we are looking for a new model. Models are compared by attributes we want to take in to account when choosing the model. Attributes like SQL-skills and prices are researched.
In this study we chose ten language models to be tested. There is 16 prompts that was given to the models. After first try some models were dropped since they were already found too inaccurate for PG4N. One was found too expensive when compare to others. Then all of the prompts were given to models twice more. This was done to ensure models actually knows SQL since their answers can change. In this work an example prompt was designed that PG4N could use. We also determined the prices of each model and the servers they are running. After going through all these metrics the chosen large language models were Llama 3 models.
Large-Language Models have become really popular in recent years. Models like GPT-4o from OpenAi and Llama 3 from Meta are capable of understanding and producing natural language. These models have been trained with a vast amount of data. When the amount of data increases, also the answers of language model improves.
PG4N is a software that utilizes a Large-Language model if user makes and SQL-error. PG4N uses a cloud service to run the model. Previously used model has been taken off the cloud service since there has been added newer, better ones. Therefore we are looking for a new model. Models are compared by attributes we want to take in to account when choosing the model. Attributes like SQL-skills and prices are researched.
In this study we chose ten language models to be tested. There is 16 prompts that was given to the models. After first try some models were dropped since they were already found too inaccurate for PG4N. One was found too expensive when compare to others. Then all of the prompts were given to models twice more. This was done to ensure models actually knows SQL since their answers can change. In this work an example prompt was designed that PG4N could use. We also determined the prices of each model and the servers they are running. After going through all these metrics the chosen large language models were Llama 3 models.