v15 training set: building+unit addresses, facility/heading/relationship hard negatives

Add v14_train.jsonl (73415 records)

Upload v13_train.jsonl with huggingface_hub

Upload train_pii_layout_synth.py with huggingface_hub

Upload v13_train.jsonl with huggingface_hub

Train script: v12

Add v12_train.jsonl: dedicated en-dash KV templates, gen_kv en-dash rate 25%->52%

Fix: remove UTF-8 BOM so uv reads PEP 723 deps correctly

Fix: HF_DEBUG=1, try/except on download, remove Unicode chars

Use hf_hub_download to bypass Xet Storage issue in datasets

Add hf_xet dependency for Xet Storage support

Update train script for v11: USERNAME->O in LABEL_MAP_TRAIN

Add v11_train.jsonl: en-dash phone weights boosted, ext@40%, lab results last-col fix, USERNAME->O

Update train script for v10

Add v10_train.jsonl: boosted en-dash phone weights, short ref-code hard negatives

Update train script for v9: new dataset, warmup_steps=200

Add v9_train.jsonl: CUSTOMER_NAME removed, en-dash SSN, cc_bare_hyphen phone, tab/space SSN, vdp_table_body template

Upload v5_plus_gen5_rev8.jsonl with huggingface_hub

Upload v5_plus_gen5_rev7_norm.jsonl with huggingface_hub

Replace rev6 with normalized span_labels (array throughout)

Add rev6 dataset (72415 examples, updated generators)

Upload train_pii_layout_synth.py with huggingface_hub

Upload v5_plus_gen5_rev5.jsonl with huggingface_hub

Upload v5_plus_gen5_rev4.jsonl with huggingface_hub

Upload v5_plus_gen5_rev3.jsonl with huggingface_hub

Upload v5_plus_gen5_rev3.jsonl with huggingface_hub

Upload v5_plus_gen5_rev2.jsonl with huggingface_hub

Add training data

initial commit