Any chance the model code or weights will be published? This would help researchers replicate your benchmark results without having to pre-train a 5B-param model. Thanks for a cool paper!