In neural machine translation experiments, the joint testing condition uses a training dataset containing both regular parallel data and a specific subset of idiom data (idiom-train data). During testing, the model's performance is measured on translating these same idioms, but within sentence contexts different from those seen in the training data.