Decontaminate test sets: drop sentences that appear verbatim in training data (pdmp -6.3%, 12m -16.5%)

Update training recipe (60 iters, prune 0.04) and F1 to 98.00/98.06

Merge RELEASE_TRAINING_DATA.md into README; simplify data provenance

Normalize tokenization to PreSegment convention (retag2 char-type); F1 97.70/97.48

Reorg data into dataset/ with wapic-cws-data-* names; drop nolabel

Add wapic-cws train/test data (retag2, jsonl+BMES, gzip)

Add files using upload-large-folder tool

initial commit