nashtree is Curbcut's long-term fork of stochtree, maintained for the NASH+ system of models. It trains BART and BCF models for nash and carries the native forest core that serves them.
It diverges from upstream in:
- sampler correctness fixes (MCMC proposal probabilities and split extrema, GFR NaN ordering, the Gaussian random-effects variance conditional, tree reconstruction with recycled node IDs);
- random-intercept support in retention modes and for unseen groups;
- selected and importance-only posterior retention and bulk sample retention;
- batched numeric preprocessing and specialized residual updates;
- the
nashtree_forest_v1binary format, its R-independent inference core, the producer-neutral writer and build receipts (native/), and the inference-onlynashtreeInferencebindings built from the same core (native/r-package/).
NASH-FORK.md records each change and its evidence.
nash pins an exact commit in its DESCRIPTION Remotes field, and builds its
inference bindings from the same commit, so install nashtree through nash
(pak::pak("Curbcut/nash")) rather than on its own. To work on the fork,
install a checkout into an isolated library with R CMD INSTALL ..
nashtree and stochtree register the same model classes and S3 methods. Never load both in one R session; compare them in separate processes.
The package, DLL and configure identity are nashtree. The C++ StochTree
namespace, the public R classes and APIs, and historical JSON fields (including
stochtree_version) are unchanged, so models saved by the stochtree fork load
in nashtree and serialize identically.
Upstream changes are not merged wholesale. Each one comes in only after a
focused regression review against NASH fits. upstream-main mirrors
upstream's main for reference. Original author and licence information is
kept in DESCRIPTION, LICENSE and inst/COPYRIGHTS.