Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research

Language models have become a critical technology to tackling a wide range of natural language processing tasks, yet many details about how the best-performing language models were developed are not reported. In particular, information about their pretraining corpora is seldom discussed: commercial language models rarely provide any information about their data; even open models rarely release datasets they are trained on, or an exact recipe to reproduce them. As a result, it is challenging to conduct certain threads of language modeling research, such as understanding how training data impacts model capabilities and shapes their limitations. To facilitate open research on language model pretraining, we release Dolma, a three trillion tokens English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials. In addition, we open source our data curation toolkit to enable further experimentation and reproduction of our work. In this report, we document Dolma, including its design principles, details about its construction, and a summary of its contents. We interleave this report with analyses and experimental results from training language models on intermediate states of Dolma to share what we have learned about important data curation practices, including the role of content or quality filters, deduplication, and multi-source mixing. Dolma has been used to train OLMo, a state-of-the-art, open language model and framework designed to build and study the science of language modeling.

AboutMe: Using Self-Descriptions in Webpages to Document the Effects of English Pretraining Data Filters

Large language models' (LLMs) abilities are drawn from their pretraining data, and model development begins with data curation. However, decisions around what data is retained or removed during this initial stage is under-scrutinized. In our work, we ground web text, which is a popular pretraining data source, to its social and geographic contexts. We create a new dataset of 10.3 million self-descriptions of website creators, and extract information about who they are and where they are from: their topical interests, social roles, and geographic affiliations. Then, we conduct the first study investigating how ten "quality" and English language identification (langID) filters affect webpages that vary along these social dimensions. Our experiments illuminate a range of implicit preferences in data curation: we show that some quality classifiers act like topical domain filters, and langID can overlook English content from some regions of the world. Overall, we hope that our work will encourage a new line of research on pretraining data curation practices and its social implications.

"One-size-fits-all"? Observations and Expectations of NLG Systems Across Identity-Related Language Features

Fairness-related assumptions about what constitutes appropriate NLG system behaviors range from invariance, where systems are expected to respond identically to social groups, to adaptation, where responses should instead vary across them. We design and conduct five case studies, in which we perturb different types of identity-related language features (names, roles, locations, dialect, and style) in NLG system inputs to illuminate tensions around invariance and adaptation. We outline people's expectations of system behaviors, and surface potential caveats of these two contrasting yet commonly-held assumptions. We find that motivations for adaptation include social norms, cultural differences, feature-specific information, and accommodation; motivations for invariance include perspectives that favor prescriptivism, view adaptation as unnecessary or too difficult for NLG systems to do appropriately, and are wary of false assumptions. Our findings highlight open challenges around defining what constitutes fair NLG system behavior.

Words as Gatekeepers: Measuring Discipline-specific Terms and Meanings in Scholarly Publications

Scholarly text is often laden with jargon, or specialized language that divides disciplines. We extend past work that characterizes science at the level of word types, by using BERT-based word sense induction to find additional words that are widespread but overloaded with different uses across fields. We define scholarly jargon as discipline-specific word types and senses, and estimate its prevalence across hundreds of fields using interpretable, information-theoretic metrics. We demonstrate the utility of our approach for science of science and computational sociolinguistics by highlighting two key social implications. First, we measure audience design, and find that most fields reduce jargon when publishing in general-purpose journals, but some do so more than others. Second, though jargon has varying correlation with articles' citation rates within fields, it nearly always impedes interdisciplinary impact. Broadly, our measurements can inform ways in which language could be revised to serve as a bridge rather than a barrier in science.

Characterizing English Variation across Social Media Communities with BERT

Much previous work characterizing language variation across Internet social groups has focused on the types of words used by these groups. We extend this type of study by employing BERT to characterize variation in the senses of words as well, analyzing two months of English comments in 474 Reddit communities. The specificity of different sense clusters to a community, combined with the specificity of a community's unique word types, is used to identify cases where a social group's language deviates from the norm. We validate our metrics using user-created glossaries and draw on sociolinguistic theories to connect language variation with trends in community behavior. We find that communities with highly distinctive language are medium-sized, and their loyal and highly engaged users interact in dense networks.

Using Sentiment Induction to Understand Variation in Gendered Online Communities

We analyze gendered communities defined in three different ways: text, users, and sentiment. Differences across these representations reveal facets of communities' distinctive identities, such as social group, topic, and attitudes. Two communities may have high text similarity but not user similarity or vice versa, and word usage also does not vary according to a clearcut, binary perspective of gender. Community-specific sentiment lexicons demonstrate that sentiment can be a useful indicator of words' social meaning and community values, especially in the context of discussion content and user demographics. Our results show that social platforms such as Reddit are active settings for different constructions of gender.

Are distributional representations ready for the real world? Evaluating word vectors for grounded perceptual meaning

Distributional word representation methods exploit word co-occurrences to build compact vector encodings of words. While these representations enjoy widespread use in modern natural language processing, it is unclear whether they accurately encode all necessary facets of conceptual meaning. In this paper, we evaluate how well these representations can predict perceptual and conceptual features of concrete concepts, drawing on two semantic norm datasets sourced from human participants. We find that several standard word representations fail to encode many salient perceptual features of concepts, and show that these deficits correlate with word-word similarity prediction errors. Our analyses provide motivation for grounded and embodied language learning approaches, which may help to remedy these deficits.