Enhancing Multilingual LLM Pretraining with Model-Based Data Selection

TL;DR

This paper introduces a transparent, efficient model-based data filtering framework for multilingual LLM pretraining, improving performance and resource efficiency across diverse languages and scripts.

Contribution

It develops a novel, accessible filtering method for multilingual datasets that enhances LLM training efficiency and effectiveness, especially for low-resource languages.

Findings

Matching baseline MMLU scores with only 15% of training tokens

Improved performance across multiple benchmarks

Mitigated the curse of multilinguality

Abstract

Dataset curation has become a basis for strong large language model (LLM) performance. While various rule-based filtering heuristics exist for English and multilingual datasets, model-based filtering techniques have primarily focused on English. To address the disparity stemming from limited research on non-English languages, we develop a model-based filtering framework for multilingual datasets that aims to identify a diverse set of structured and knowledge-rich samples. Our approach emphasizes transparency, simplicity, and efficiency, leveraging Transformer- and FastText-based classifiers to ensure the broad accessibility of our technique and data. We conduct comprehensive ablation studies on the FineWeb-2 web crawl dataset across diverse language families, scripts, and resource availability to demonstrate the effectiveness of our method. Training a 1B-parameter Llama model for 70B and 119B tokens, our approach can match the baseline MMLU score with as little as 15% of the training tokens, while also improving across other benchmarks and mitigating the curse of multilinguality. These findings provide strong evidence for the generalizability of our approach to other languages. As a result, we extend our framework to 20 languages for which we release the refined pretraining datasets.