This large-scale dataset is designed for pre-training state-of-the-art text embedding models. Its goal is to reproduce and build upon the data recipe described in the mGTE technical report (Zhang et al., 2024), which details the data sources used to train the GTE family of embedding models but does not release the data itself.
We assembled this dataset as part of a research effort to understand how data composition affects retrieval model quality. Our experiments confirmed… See the full description on the dataset page:
https://huggingface.co/datasets/lightonai/embeddings-pre-training.
