Most retrieval gains come from realistic, hard negatives — not just random passages. This dataset supplies per-query ranked negatives (up to 2048) alongside the positive and relevance scores across seven high-quality sources, letting you experiment with nv-retrieve-style filtering, thresholding, and sampling strategies without destructive pre-filtering.
What Sets It Apart
- Full-ranked negatives preserved: every query includes the positive followed by up to 2048 mined negatives plus their scores, so you can apply any nv-threshold or custom sampling instead of using a single pre-filtered set. This enables reproducible comparisons between filtering strategies.
- Multi-dataset coverage at scale: combines FiQa, NaturalQuestions, HotpotQA, MSMARCO, FEVER, SquadV2 and TriviaQA — ~1.88M score rows, ~1.22M queries, and ~50.6M documents across splits, suitable for both fine-tuning and downstream retrieval evaluation.
- Format & tooling: provided in parquet/optimized-parquet and organized as three configs (documents, queries, scores) that can be mapped to (query, positive, negative_1..n) contrastive rows; example mapping code is included by the publisher for common frameworks.
- Transparent mining model: negatives were mined using the gte-modernbert bi-encoder, and scores are provided so you can reproduce nv-retrieve (filtering by a fraction of positive score) or perform knowledge distillation experiments.
Who It's For and Trade-offs
Great fit if you want a ready-made, large-scale contrastive dataset for retrieval/embedding fine-tuning, especially when you need to compare sampling/filtering policies or apply nv-retrieve-style negative filtering. It is also useful for knowledge-distillation and hard-negative curriculum experiments. Look elsewhere if you need curated, license-cleared proprietary content (the dataset page lists no explicit license), dense single-vector evaluation suites out-of-the-box, or extremely small/lightweight datasets for quick prototyping. Also plan for large storage and memory needs: combined score and document files are tens of terabytes uncompressed across splits.
Where It Fits
Use this after (or alongside) a broad contrastive pre-training phase: the publisher recommends pairing it with large-scale pre-training datasets for best results. Because it preserves raw mined negatives and scores, it is more flexible than pre-filtered training sets but requires you to decide thresholds and sampling policies.
