rubert-nli-threeway (ONNX: fp32 / fp16 / int8)
Three-way NLI (entailment / contradiction / neutral) cross-encoder, ONNX export. Three precision variants; all keep fp32 I/O (int64 ids, float32 logits) so consumers are unchanged:
model.onnx- fp32 (reference).model.fp16.onnx- fp16 weights (GPU; ~half size).model.int8.onnx- dynamic-int8 weights on the linear (MatMul) layers (CPU; smaller, embeddings kept in fp32 to preserve class parity).
Inputs input_ids+attention_mask+token_type_ids; output logits [b,3] with id2label
{0: entailment, 1: contradiction, 2: neutral}. Validated on 8 generic (premise, entailing,
contradicting) triples: predicted class preserved on every pair for all variants (int8 vs fp32
max probability drift <=0.03). kbx nli download [--fp16|--int8] --to <dir> fetches a variant
(saved as model.onnx). Base license apache-2.0.