README.md

metadata

license: apache-2.0
language:
  - en
task_categories:
  - text-ranking
  - text-retrieval
pretty_name: BlockRank MS MARCO Training Data (10% Sample)
size_categories:
  - 10K<n<100K

BlockRank MS MARCO Training Data (10% Sample)

Dataset Description

A 10% sample of MS MARCO passage ranking data formatted for training in-context ranking LLMs. This dataset is used in the training of the BlockRank project: Scalable In-context Ranking with Generative Models.

• Format: JSONL (in-context ranking format)
• Size: 50k training examples (10% sample)
• Documents per query: 30-50 candidates (mix of positives and hard negatives)

Source

• Original: MS MARCO Passage Ranking
• Retrieval annotations: sentence-transformers/msmarco-distilbert-base-v3
• Sampling: 10% random sample of training queries
• Negatives: Top dense retrieval results from msmarco-distilbert-base-v3 sbert model (excluding positives)

License

Apache 2.0 (following MS MARCO license)