GNSS ambiguity acceptance test is the key procedure to improve the reliability of GNSS ambiguity resolution. The hypothesis test model more the ambiguity acceptance test has been well modelled while how to reasonably determine the threshold is still an open problem. The empirical threshold and the model-driven fixed failure rate has been extensively studied. While these methods is difficult to adapt observation condition variation since they are not sensitive to abrupt data quality change. This study proposed a new data-driven threshold determination method called radical basis function support vector machine (RBF-SVM), which takes the ambiguity acceptance test as a classification problem and makes a decision based on historical data. The implicit non-linear relationship between the correctness of the fixed integer and the test statistics are captured with a training process and then applying this relationship to make decision. Numerical results indicates that the proposed RBF-SVM approach achieves fairly low failure rate although no failure rate tolerance is applied, while it improves success rate comparing to the empirical threshold and the fixed failure rate approach.