When working on adaptive scheduling for heterogeneous architectures ^JVGP2009, we performed online modeling of program behavior vs dataset parameters vs CPU/GPU execution. We would like to unify this approach in cM and provide a mechanism to quickly scan large dataset space (using our probabilistic approach ^FT2010, ^FOTP2005 or ANOVA to focus exploration on points with unusual behavior and quickly reduce dimensions) combined with some off-the-shelf decision trees or hybrid Multivariate adaptive regression splines (MARS) ^LCWP2009. We have prototypes ready and plan to add off-the-shelf models from R or Scientific Python to cM. Collaborations and help are welcome!