Network security research requires network traffic data sets of sufficient size, variety, and completeness in order to perform tasks such as training intrusion detection systems. While the standard is to use testbeds to create data sets or capture data sets from real systems, Generative Adversarial Networks have proven successful in generating new packet samples for protocols such as ICMP, DNS, HTTP, and SIP. However, existing approaches have problems with quality evaluation due to insufficient sampling, or they require non-generalised criteria to be created specifically for the data set being trained on. This paper proposes a new and generalised two-pass approach to evaluating the quality of samples produced by the generator to produce a filtered, higher-quality output data set. Compared against SIP-GAN, which is a Generative Adversarial Network model targeting Session Initiation Protocol samples, we reduced the ratio of malformed SIP samples from between 9.6% and 19.8% down to 1.2%.