Re-chunk of ai4data/datause-dataset into <=384-token windows (max_tokens=384, overlap=50) so GLiNER's window is never truncated during evaluation.
{"train": {"orig": 1779, "rechunked": 1822}, "validation": {"orig": 415, "rechunked": 458}, "holdout": {"orig": 1149, "rechunked": 1166}}
