Remove the same 271 stable IDs from full-answer and short-answer SFT datasets: 20 same mathematical tasks and 251 derivatives exposing pinned benchmark content and forward answers. Preserve all retained values, steps, row order and schema. Keep 8 modified variants pending policy review and 3 different mathematical tasks. Update card, manifest, provenance, and validation evidence.
Add 352470 examples with shortened answers replacing answer
