A dataset created by lightonai on April 17, 2026, composed of high-quality data sources with mined hard negatives. It is intended for training a strong retrieval model, particularly after large-scale contrastive pre-training. The dataset was originally created to follow the nv-retrieve setup.
Use Cases
- Fine-tuning retrieval models based on the inclusion of mined hard negatives.
- Training a strong retrieval model independently based on the high-quality data sources.
- Performing large-scale contrastive pre-training based on the dataset's curated version.
- Filtering false negatives in retrieval tasks based on the described bi-encoder similarity method.
Strengths
- Dataset includes mined hard negatives, a feature for improving retrieval model training.
- Created specifically for the nv-retrieve setup, indicating a targeted design.
- Last updated on April 17, 2026, suggesting recent maintenance.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- File formats and size are unknown, complicating download and integration planning.
Provenance
- Source
- huggingface
- Collection Method
- Composed from high-quality data sources with mined hard negatives.
- Freshness
- Last updated 2026-04-17 14:14:08