The crystallization of modeling methods around the Transformer architecture has been a boon for practitioners. Simple, well-motivated architectural variations can transfer across tasks and scale, increasing the impact of modeling research. However, with the emergence of state-of-the-art 100B+ parameters models, large language models are increasingly expensive to accurately design and train. Notably, it can be difficult to evaluate how modeling decisions may impact emergent capabilities, given that these capabilities arise mainly from sheer scale alone. In the process of building BLOOM–the Big Science Large Open-science Open-access Multilingual language model–our goal is to identify an architecture and training setup that makes the best use of our 1,000,000 A100-GPU-hours budget. Specifically, we perform an ablation study at the billion-parameter scale comparing different modeling practices and their impact on zero-shot generalization. In addition, we study the impact of various popular pre-training corpora on zero-shot generalization. We also study the performance of a multilingual model and how it compares to the English-only one. Finally, we consider the scaling behaviour of Transformers to choose the target model size, shape, and training setup. All our models and code are open-sourced at https://huggingface.co/bigscience.

High quality datasets for question answering exist in a few languages, but far from all. Producing such datasets for new languages requires extensive manual labour. In this work we look at different methods for using existing datasets to train question-answering models in languages lacking such datasets. We show that machine translation followed by cross-lingual projection is a viable way to create a full question-answering dataset in a new language. We introduce new methods both for bitext alignment, using optimal transport, and for direct cross-lingual projection, utilizing multilingual BERT. We show that our methods produce good Swedish question-answering models without any manual work. Finally, we apply our proposed methods on Spanish and evaluate it on the XQuAD and MLQA benchmarks where we achieve new state-of-the-art values of 80.4 F1 and 62.9 Exact Match (EM) points on the Spanish XQuAD corpus and 70.8 F1 and 53.0 EM on the Spanish MLQA corpus, showing that the technique is readily applicable to other languages.