MS MARCO with BGE-M3: Query-Document Pairs for Retrieval Training

MS MARCO dataset documents have been matched to queries using the BGE-M3 embedding model and scored by the bge-reranker-v2-m3 model. The dataset was created by lightonai and was last updated on the Hugging Face platform in September 2025. It is intended for training retrieval models via knowledge distillation techniques.

Use Cases

• Train a retrieval model using knowledge distillation based on the described query-document pairs.
• Fine-tune embedding models using the BGE-M3 generated similarities mentioned in the description.
• Benchmark reranking models using scores from the bge-reranker-v2-m3 model.
• Implement retrieval training pipelines, for example using the PyLate framework referenced in the description.

Strengths

• Documents are matched to queries using the BGE-M3 model, a specific and modern embedding technique.
• Document-query relevance is scored by the bge-reranker-v2-m3 model, providing a second signal for training.
• Dataset is hosted on Hugging Face and was updated on 2025-09-11, indicating recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Description metadata is limited; actual data quality and structure require manual inspection after download.

Provenance

Source

Based on the MS MARCO dataset, processed by lightonai.

Collection Method

Documents similar to queries were mined using BGE-M3 and then scored by bge-reranker-v2-m3.

Freshness

Last updated 2025-09-11 12:01:12; freshness should be verified.