Pretraining data design is critically under-documented and often guided by empirically unsupported intuitions. We pretrain models on data curated (1) at different collection times, (2) with varying toxicity and quality filters, and (3) with different domain compositions. First, we find that temporal shift between evaluation data and pretraining data leads to performance degradation, which is not overcome by finetuning. Second, we measure the effect of quality and toxicity filters, showing a trade-off between performance on standard benchmarks and risk of toxic generations. We also find that the effects of different types of filtering are not predictable from text domain characteristics. Third, we empirically validate that heterogeneous data sources, like books and web, are beneficial and warrant greater prioritization. To date, these experiments constitute the single largest publicly documented empirical study of the effects of pretraining data. Spanning 28 unique 1.5 billion parameter models pretrained from scratch, these findings validate, quantify, and expose many undocumented intuitions about text pretraining, which ultimately support more informed data-centric decisions in model development.

Large language models achieve high performance on many but not all downstream tasks. The interaction between pretraining data and task data is commonly assumed to determine this variance: a task with data that is more similar to a model’s pretraining data is assumed to be easier for that model. We test whether distributional and example-specific similarity measures (embedding-, token- and model-based) correlate with language model performance through a large-scale comparison of the Pile and C4 pretraining datasets with downstream benchmarks. Similarity correlates with performance for multilingual datasets, but in other benchmarks, we surprisingly find that similarity metrics are not correlated with accuracy or even each other. This suggests that the relationship between pretraining data and downstream tasks is more complex than often assumed.

Much of the progress in contemporary NLP has come from learning representations, such as masked language model (MLM) contextual embeddings, that turn challenging problems into simple classification tasks. But how do we quantify and explain this effect? We adapt general tools from computational learning theory to fit the specific characteristics of text datasets and present a method to evaluate the compatibility between representations and tasks. Even though many tasks can be easily solved with simple bag-of-words (BOW) representations, BOW does poorly on hard natural language inference tasks. For one such task we find that BOW cannot distinguish between real and randomized labelings, while pre-trained MLM representations show 72x greater distinction between real and random labelings than BOW. This method provides a calibrated, quantitative measure of the difficulty of a classification-based NLP task, enabling comparisons between representations without requiring empirical evaluations that may be sensitive to initializations and hyperparameters. The method provides a fresh perspective on the patterns in a dataset and the alignment of those patterns with specific labels.

Images can give us insights into the contextual meanings of words, but current image-text grounding approaches require detailed annotations. Such granular annotation is rare, expensive, and unavailable in most domain-specific contexts. In contrast, unlabeled multi-image, multi-sentence documents are abundant. Can lexical grounding be learned from such documents, even though they have significant lexical and visual overlap? Working with a case study dataset of real estate listings, we demonstrate the challenge of distinguishing highly correlated grounded terms, such as “kitchen” and “bedroom”, and introduce metrics to assess this document similarity. We present a simple unsupervised clustering-based method that increases precision and recall beyond object detection and image tagging baselines when evaluated on labeled subsets of the dataset. The proposed method is particularly effective for local contextual meanings of a word, for example associating “granite” with countertops in the real estate dataset and with rocky landscapes in a Wikipedia dataset.