1.7B SPP models: bases publish their training trajectory as step-* revisions, plus persona-binding instruct counterparts.