Predictive Data Selection: The Data That Predicts Is the Data That Teaches

TL;DR

This work introduces PreSelect, a predictive data selection method for language model pretraining that identifies high-impact documents by measuring how well losses on a seed set predict downstream performance. A lightweight fastText scorer is trained on a small, labeled set of positives and negatives derived from a ranking-based predictive strength metric, enabling scalable document-level data selection without heavy model training. Across 1B and 3B parameter scales and multiple data pools, PreSelect yields consistent improvements over strong baselines (average 3.1% across 15 benchmarks) and can achieve substantial compute savings (up to 10x) while maintaining or surpassing performance achieved with far larger token budgets. The authors also provide extensive analyses of data characteristics, feature contributions, and length distributions, and release their scorer and curated datasets for public use.

Abstract

Language model pretraining involves training on extensive corpora, where data quality plays a pivotal role. In this work, we aim to directly estimate the contribution of data during pretraining and select pretraining data in an efficient manner. Specifically, we draw inspiration from recent findings showing that compression efficiency (i.e., the normalized loss) of diverse models on certain text correlates strongly with their downstream performance, when the text domain aligns with the downstream benchmarks(Huang et al., 2024). Building on this observation, we hypothesize that data on which model losses are predictive of downstream abilities also contribute effectively to learning, which shares similar intuition with Thrush et al.(2024). To leverage this insight, we introduce predictive data selection (PreSelect), a lightweight and efficient data selection method that requires training and deploying only a fastText-based scorer. Through comprehensive experiments with 1B and 3B parameter models, we demonstrate that models trained on 30B tokens selected with PreSelect surpass the performance of the vanilla baseline trained on 300B tokens, achieving a 10x reduction in compute requirements. Furthermore, PreSelect significantly outperforms other competitive data selection baselines, such as DCLM and FineWeb-Edu on a scale of 3B models trained on 100B tokens. We open-source our trained data selection scorer along with the curated datasets at https://github.com/hkust-nlp/PreSelect.

Figures (8)

• Figure 1: PreSelect outperforms random selection by an average of 3.1% accuracy across 15 downstream benchmarks and achieves 10x reduction in compute requirements on the scale of 1B models using RefinedWeb corpus.
• Figure 2: The overall framework of PreSelect. It first samples a subset of pre-training data for computing the predictive strength score of each document (§\ref{['section-approach-subsection-task-formulation']}). Then a fastText-based scorer is trained based on the identified positive data and negative data. Finally the fastText-based scorer is trained to enable scalable data selection.
• Figure 3: Length distribution of different data selection baselines on RefinedWeb measured by the number of characters. The length annotations are averaged characters.
• Figure 4: The matching score distribution on RefinedWeb subset where reflect the alignment between compression efficiency and averaged downstream performance. The higher the better.
• Figure 5: The Top-10 domains where the identified positive documents from, measured by percentage of positive documents inside that domain.