Benchmarking down-scaled (not so large) pre-trained language models

Large Transformer-based language models are pre-trained on corpora of varying sizes, for a different number of steps and with different batch sizes. At the same time, more fundamental components, such as the pre-training objective or architectural hyperparameters, are modified. In total, it is therefore difficult to ascribe changes in performance to specific factors. Since searching the hyperparameter space over the full systems is too costly, we pre-train down-scaled versions of several popular Transformer-based architectures on a common pre-training corpus and benchmark them on a subset of the GLUE tasks (Wang et al., 2018). Specifically, we systematically compare three pre-training objectives for different shape parameters and model sizes, while also varying the number of pre-training steps and the batch size. In our experiments MLM + NSP (BERT-style) consistently outperforms MLM (RoBERTa-style) as well as the standard LM objective. Furthermore, we find that additional compute should be mainly allocated to an increased model size, while training for more steps is inefficient. Based on these observations, as a final step we attempt to scale up several systems using compound scaling (Tan and Le, 2019) adapted to Transformer-based language models.