Msmarco Hard Negatives Llm Scores

100 candidate documents per query from the MS MARCO train split annotated with re-ranking scores from bi-encoder and cross-encoder teacher models. The dataset utilizes the opensearch-neural-sparse-encoding-doc-v1 model for initial retrieval to identify relevant hard negatives for training retrieval systems. These annotations facilitate the development of neural search models through knowledge distillation and contrastive learning.

Use Cases

• Train a neural retriever using the bi-encoder and cross-encoder scores as supervision signals for distillation
• Improve ranking models by utilizing the top-100 candidates as hard negatives during contrastive training
• Analyze the score distribution between bi-encoder and cross-encoder teachers for specific MS MARCO queries

Strengths

• Top-100 candidate documents per query retrieved using opensearch-neural-sparse-encoding-doc-v1
• Includes re-ranking scores from both bi-encoder and cross-encoder teacher models
• Derived from the standard MS MARCO train split for compatibility with existing IR research