Qdrant-FineWeb-10B is a public dataset containing 10 billion records and nearly 120,000 ground-truth queries, designed to benchmark vector retrieval systems at production-scale workloads.
Qdrant has released Qdrant-FineWeb-10B, a large-scale dataset designed for benchmarking vector retrieval systems. Built from a 10-billion-document slice of Hugging Face’s FineWeb corpus, the dataset is intended to help developers test vector databases and retrieval systems under workloads closer to those found in large production environments.
Each document in the dataset is represented using both dense and sparse embeddings generated with the gte-multilingual-base model. Alongside the embeddings, Qdrant-FineWeb-10B retains the original document text, metadata and payload information, providing a large corpus for testing different retrieval approaches.
The benchmark includes approximately 120,000 queries covering dense, sparse and filtered retrieval. Qdrant generated exact top-1,000 ground-truth results for these queries by performing brute-force nearest-neighbour calculations across the complete 10-billion-vector corpus. According to Qdrant, this required more than one quadrillion distance calculations using GPU-accelerated infrastructure.
Alongside the dataset, Qdrant has released Supernova, an open-source distributed benchmarking framework used to generate the embeddings and calculate the ground truth. The framework covers embedding generation, brute-force ground-truth calculation, database loading and benchmark evaluation, allowing others to create and test large-scale vector datasets on their own infrastructure.
Qdrant-FineWeb-10B is available publicly through Hugging Face and is released under the ODC-BY licence. By making both the dataset and its benchmarking tools available to the community, Qdrant aims to provide developers with a reproducible platform for evaluating dense, sparse, hybrid and filtered vector retrieval systems at a much larger scale than conventional benchmark datasets.
