diff --git a/orttraining/orttraining/core/graph/loss_func/bert_loss.cc b/orttraining/orttraining/core/graph/loss_func/bert_loss.cc index 3b8b60ab56..2f3b3f4e90 100644 --- a/orttraining/orttraining/core/graph/loss_func/bert_loss.cc +++ b/orttraining/orttraining/core/graph/loss_func/bert_loss.cc @@ -122,7 +122,7 @@ GraphAugmenter::GraphDefs BertLoss::operator()(const Graph& graph, const LossFun "Reshape_label")); std::vector attrs; - attrs.push_back(ONNX_NAMESPACE::MakeAttribute("ignore_index", static_cast(-1))); + attrs.push_back(ONNX_NAMESPACE::MakeAttribute("ignore_index", static_cast(0))); attrs.push_back(ONNX_NAMESPACE::MakeAttribute("reduction", "mean")); new_nodes.emplace_back(NodeDef("SoftmaxCrossEntropyLoss",