3 · Data processing (coming later)
This section is reserved for the next step: building the unified training targets
(compact block markers, x-first 0-1000 boxes, figure blocks without text, HTML tables, LaTeX), the manifest fixes
(hw-vs image extensions), the leftover cleaning (unclosed figure descriptions, literal \n, synthetic line breaks),
and per-batch source quotas with sequence packing for the 32K context.