guard-ft-v19

guard-ft-v19 (vocabulary-trimmed)

GUARD-FT round 19 (AgentDocs' EN+FR legal-document PII token classifier). This repo's main now carries the VOCABULARY-TRIMMED export: XLM-R-large's 250,002-piece vocabulary reduced to the 53,751 pieces the EN+FR training corpus actually used (plus every byte-level/single-character fallback piece), tools/guard-train/trim_vocab.py. Same checkpoint, same weights on every kept row — this is row-selection on the embedding matrix, not a retrain, and measured byte-identical to the untrimmed export's masking decisions across a 259-document evaluation corpus.

• Weights: ~2.1 GB (untrimmed) -> ~1.4 GB (this revision), now a single self-contained model.onnx (under the 2 GB protobuf limit, so no companion model.onnx_data any more).
• The untrimmed original remains reachable at the pinned historical revision 09451885cc53531ecd28aa3bbc1be7b7986ffb9e if you need the full 250,002-piece vocabulary.
• Sibling exports: guard-ft-v19-trimmed-fp16 (CUDA-only, ~716 MB) and guard-ft-v19-trimmed-int8 (CPU-only, ~360 MB, measurable precision trade-off).
• Labels: PERSON, ORG, PUBLIC_BODY, ROLE, ADDRESS, PLACE, DOB, CONTACT (BIO scheme)