Decision bias refers to a systematic tendency to approach or avoid a particular option in two-alternative forced-choice (2AFC) tasks. With the advent of easy-to-use tools, diffusion-decision model (DDM) is widely used to distinguish two types of latent cognitive processes, bias in starting point or in efficiency, providing a powerful tool for understanding a variety of psychological phenomena across different fields. However, correctly estimating these two types of bias requires specifying the model to match the type of bias being investigated. So far, no study has systematically evaluated the different model configurations for modelling decision bias in HDDM, the most widely used python package for DDM. Here we evaluated nine model specifications using the three main modeling modules in HDDM across both simulated and empirical data. Our results show that the default function, which uses accuracy coding of data without flipping parameters, fails to recover either the starting point bias or the drift rate bias. Only model specifications that directly estimate the drift bias parameter, drift criterion (dc), or apply appropriate parameter flipping recover true bias parameters. By applying optimal specification to empirical data, we uncovered a novel psychological effect that went undetected in the original study. We provide concrete, evidence based recommendations for model specification and fully reproducible code. This work ensures accurate parameter estimation and the correct interpretation of parameter when modeling decision biases.