Compare larger language models under supervised fine-tuning

Determine whether larger language models, specifically models larger than LLaMA-3.1-8B-Instruct, produce different or improved results for supervised fine-tuning on within-turn Transition Relevance Place prediction.

Background

The supervised fine-tuning experiments use LLaMA-3.1-8B-Instruct with LoRA adapters rather than the larger LLaMA-3.1-70B-Instruct model used in the prompting experiments. The choice was motivated by the absence of a consistent scale benefit in prompt-based prediction and by computational considerations.

The paper explicitly leaves unresolved whether larger models behave differently when their parameters are updated through supervised fine-tuning. This comparison would clarify whether the limited and non-monotonic fine-tuning gains reflect model size, adaptation method, or the formulation of the within-turn TRP task.