NVIDIA launches Nemotron-CC, a large-scale dataset for LLM pre-training

NVIDIA announced the launch of Nemotron-CC, a groundbreaking 6.3 trillion token English language dataset designed to improve dictionary training of large-scale language models (LLMs). This dataset, derived from Common Crawl, aims to increase the accuracy and efficiency of LLM through innovative data curation techniques, including the use of 1.9 trillion synthetically generated data tokens, according to NVIDIA.

Enhancing LLM pre-education

NVIDIA’s initiative addresses a critical need in LLM training, where the quality of pre-training datasets plays a pivotal role. Recent models, such as Meta’s Llama series, have been based on datasets consisting of up to 15 trillion tokens, but the exact composition of these datasets is largely unknown. Nemotron-CC seeks to fill this gap by providing the wider community with high-quality datasets that can support both short-term and long-term Token Horizon training.

Existing datasets often sacrifice up to 90% of data to improve benchmark accuracy, limiting their usefulness for widespread training. However, Nemotron-CC demonstrates how advanced methods such as classifier ensembles and synthetic data reconstruction can transform the Common Crawl data into a superior dataset that outperforms the Llama 3.1 8B model.

important results

The efficacy of Nemotron-CC is demonstrated by its performance on a variety of benchmarks. When training an 8B parameter model on 1 trillion tokens, the high-quality subset Nemotron-CC-HQ outperforms key datasets such as DCLM, increasing the MMLU score by 5.6 points. Additionally, the full 6.3 trillion token dataset matches MMLU’s DCLM while providing 4x more unique real-world tokens. This allowed the Nemotron-CC trained model to outperform Llama 3.1 8B on several metrics, including a 5-point increase in MMLU and a 3.1-point increase in ARC-Challenge score, enabling effective training over long token periods.

Innovative data curation technology

The development of Nemotron-CC involved several key insights. By combining different model-based classifiers, NVIDIA was able to select a wider range of high-quality tokens. Paraphrasing techniques also reduces noise and errors, creating diverse and valuable data transformations. The decision to disable the existing heuristic filter further improved the quality of the data set without compromising accuracy.

NVIDIA leveraged the NeMo Curator tool to extract and refine data from Common Crawl, applying filters for language, deduplication, and quality classification. This process was complemented by synthetic data generation, contributing approximately 2 trillion tokens to the dataset.

future prospects

Nemotron-CC has established itself as an essential resource for pre-training cutting-edge LLMs across a diverse range of tokens. To further enhance LLM capabilities, NVIDIA plans to expand its product by releasing more specialized datasets, including datasets focused on specific areas such as mathematics.

Image source: Shutterstock