Year: 2025
Type: dissertation
Abstract: Testing is a vital part in software- and hardware development to ensure that requirements are met. Use of language models within Safety-critical testing is an under-studied field, with previous work focusing on automated tools requiring a large number of documents, whereas models may be able to produce the same output with far less data. This thesis covers generating code, pseudo-code and summaries fortest cases in safety-critical applications using small language models, fine-tuned on task specific data. Models are fine-tuned on test code and summaries provided by ABB and are tasked with producing newtest code from requirements. Additionally, synthetic data produced by Claude 3.7 Sonnet Thinking isused for knowledge distillation to a smaller model to generate pseudo-code, which in turn is used to traina model to create new pseudo-code which is capable of outputting pseudo-code from directly from safetyrequirements. Models fine-tuned on summaries gave results equivalent to the much larger Copilot language model, and at times better, showing that small, fine-tuned language models can perform similar,or better in comparison with large language models. For the test and pseudo-code cases, results point tosmall language models being a promising path forward for both complete test code generation and as ahelpful tool in generating pseudo code to speed-up work for test developers.
Authors: Adrian, Mosqueda, Lindholm, Johan
Institution: Mälardalen University
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7045587679