<p>Landslide hazards cause human fatalities and damage infrastructure in hilly areas worldwide. Establishing landslide susceptibility maps (LSMs) is a critical step in assessing and reducing landslide-related risk. Landslide prediction models have been extensively studied to produce LSMs for many years. However, the models often focus on comparing accuracy metrics, while the reliability of the LSMs they yield is often ignored. Moreover, non-landslide samples still need a thorough study. To address this research gap, we first present a study on non-landslide samples, which are combined from areas with low slopes and high slopes, and an evaluation of machine learning models in terms of accuracy metrics. We then also employ two statistical metrics to assess the reliability of LSMs. Our study was conducted in the Tam Chung area, Vietnam. We collected 13 influence factors and a landslide inventory of 173 landslide points and 55 delineated landslide polygons. We constructed five datasets, including landslide samples and non-landslide samples. Support vector machine (SVM), K-nearest neighbours (KNN), Logistic regression (LR), and random forest (RF) were used to build landslide prediction models. Empirical results have indicated that RF is the best method for accuracy metrics. However, LSMs yielded by SVM are the best choice in terms of accuracy and reliability, and the ratio of area with low slope and area with high slope for non-landslide samples is recommended to be 3:1.</p>