Estimating $\lambda$:

Fixing $\lambda=1$ is a good idea for most data, but in some situations-e.g., for sequence data (not that we necessarily recommend structure for sequence data!)-other values may work better. For this reason, you can get the program to estimate $\lambda$ for your data. You may want to do this once, perhaps for $K=1$, and then fix $\lambda$ at the estimated value thereafter, because there seem to be some problems with non-identifiability when trying to estimate too many of the hyperparameters ($\lambda$, $\alpha$, $F$) at the same time.