GUARD-FT round 19 (AgentDocs' EN+FR legal-document PII token classifier), vocabulary-trimmed to the pieces its EN+FR training corpus actually uses (250,002 -> 53,751 pieces), fused (Attention/SkipLayerNormalization/BiasGelu via onnxruntime.transformers.optimizer) then weight-cast to fp16 (encoder + classifier head; graph I/O stays float32/int64 at the boundary).

Same checkpoint and same predictions as gauthierrobert2/guard-ft-v19 on GPU, and now also on CPU — measured near-identical across a 259-document evaluation corpus at every downstream masking stage (stage 3 leaked identical, F2 within 0.01pt). Loads correctly under ONNX Runtime 1.24.1's CPU execution provider at ORT_ENABLE_ALL (an earlier revision of this repo did not: fusing BEFORE the fp16 cast, instead of after, is what fixes it — see the model card history).

CUDA remains the fast path (~1.7x the CPU token-classifier throughput measured here); on CPU this export is correct but not faster than fp32 — ONNX Runtime has very limited native fp16 CPU kernels, so gauthierrobert2/guard-ft-v19-trimmed-int8 is the CPU-side sibling instead.