We demonstrate that a model pretrained with our objective obtains state-of-the-art results on SentEval