Signed-off Chiboard 1.1 Plan 04.5 supplemental training candidates.
Rows retain per-record source and license provenance. Mix English source types by supervised target tokens at 95% zero-LanguageTool-finding Wikipedia records and 5% independently authored project technical controls.
The rows use completion-only Chiboard serialization, preserve protected spans,
exclude the frozen 50,000-row core plus 1,000-row evaluation addition ledger,
and… See the full description on the dataset page:
https://huggingface.co/datasets/johnbean393/chiboard-1.1-english-sft.
