In my experience, it really depends on the situation. While it may be attractive to adapt an existing model, you may eventually find that it is weak in specific contexts, and work is required to fll in the gaps to get it behaving as you might want. Existing models are frequently a black box over which you have no insight. Building your own model from specific sources and samples gives you full control over the context and balance of the project. One solution is to spend time doing both, and spending most of your time with the one that seems to work best.