Stepami/rubert-tiny-bilingual-nli-onnx

RuBERT tiny bilingual NLI: ONNX export

ONNX export of cointegrated/rubert-tiny-bilingual-nli, source revision 3907306f5e5d65e281a524c536883b766c89ded2. Original model by cointegrated. No additional training or quantization.

FP32 weights, ONNX opset 17. Inputs: int64 input_ids, attention_mask, token_type_ids, with dynamic batch and sequence dimensions. Output: logits for not_entailment (0) and entailment (1); apply softmax. Maximum pair length: 512 tokens.

Use actual paired BERT tokenization with separate premise and hypothesis. Concatenating strings with [SEP] does not construct the correct segment IDs. The original model card does not specify a license; this export does not assign a new license.

Contains only the ONNX graph and original tokenizer/configuration files. Exported with torch.onnx.export and checked against the source model using ONNX Runtime CPU.