DataComp-LM: In search of the next generation of training sets for language models

2024-06-17

Neural Information Processing Systems

AI Summary

This study addresses the inefficiency of dataset design in language model pretraining by proposing the DataComp for Language Models (DCLM) benchmark. Methodologically, it constructs a standardized 240-TB Common Crawl corpus, integrates model-driven data filtering, deduplication, and mixing strategies, and establishes a scalable pretraining pipeline alongside a comprehensive 53-task downstream evaluation suite. Its key contribution is the first systematic empirical validation that model-driven filtering critically enhances data quality; it also releases the DCLM-Baseline dataset. Experiments show that a 7B model trained on DCLM achieves 64.0% accuracy on MMLU (5-shot), outperforming MAP-Neo by 6.6 percentage points while reducing computational cost by 40%. Its performance matches that of Mistral-7B-v0.3 and Llama-3-8B, yet its training cost is only ~1/6.6 of theirs—demonstrating substantial gains in data efficiency.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

We introduce DataComp for Language Models (DCLM), a testbed for controlled dataset experiments with the goal of improving language models. As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations. Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing at model scales ranging from 412M to 7B parameters. As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set. The resulting dataset, DCLM-Baseline enables training a 7B parameter language model from scratch to 64% 5-shot accuracy on MMLU with 2.6T training tokens. Compared to MAP-Neo, the previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6.6 percentage point improvement on MMLU while being trained with 40% less compute. Our baseline model is also comparable to Mistral-7B-v0.3 and Llama 3 8B on MMLU (63%&66%), and performs similarly on an average of 53 natural language understanding tasks while being trained with 6.6x less compute than Llama 3 8B. Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation.

Problem

• Improving language models through controlled dataset experiments
• Standardizing data curation strategies for better training sets
• Reducing compute costs while enhancing model performance

Innovation

• Standardized corpus of 240T tokens
• Model-based filtering for quality data
• Effective pretraining with OpenLM framework