In the Hyperparameter Choice page, an experiment is conducted to determine the best combination of features and hyperparameters for an SVM used within GlitchProber. The parameters considered include the regularization parameter C and the polynomial kernel's degree. The results indicate that a higher value of C and degree leads to better performance, specifically achieving the highest F1-score with C=1 and degree=3, thus selecting these for the tool's operation. The decision is based on the tool's ability to achieve optimal performance without necessitating post-processing adjustments.
In the Finetuning page, the approach involves fine-tuning language models (LLMs) to reduce the incidence of "glitch tokens". A specific LLM, Llama-2-7b-chat, is fine-tuned using a dataset containing 3,000 Q&A pairs, and its performance is evaluated against standard datasets. The findings highlight that while fine-tuning can target specific issues (like glitch tokens), it adversely affects the model’s core capabilities in other tasks, such as code writing and math problem solving. In contrast, \tool offers a method that adjusts the model’s processing without altering its underlying parameters, preserving its foundational skills.
Together, these parts highlight the careful balance between enhancing model performance through hyperparameter optimization and the potential downsides of altering core model parameters through fine-tuning.