Determine the effect of extended and higher-capacity QLoRA adaptation

Determine whether training the Qwen2.5-VL-7B-Instruct hallucination-detection adapter for more than one epoch improves or causes overfitting, and whether higher LoRA rank, 8-bit or full-precision adaptation, or a larger image-token budget changes the reported Contrastive Instability (CI) of 0.035.

Background

HalDetect adapts Qwen2.5-VL-7B-Instruct with 4-bit NF4 QLoRA, rank-8 adapters, a fixed image-token budget, and one training epoch. These choices were constrained by the authors' dual-T4 compute budget rather than selected through a systematic hyperparameter sweep.

The paper explicitly states that the authors could not test whether additional epochs improve performance or induce overfitting, and that they cannot rule out performance changes from longer training, higher adapter rank, alternative numerical precision, or a larger image budget. Consequently, the reported CI of 0.035 is a result for a compute-bounded configuration rather than an established optimum.