kiyam/ddro-msmarco-doc-dataset-300k

DDRO — MS MARCO Top-300K Processed Dataset

This dataset contains the preprocessed MS MARCO Top-300K document corpus used to train and evaluate the DDRO generative retrieval models from:

[Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval (SIGIR 2025)](https://arxiv.org/abs/2504.05181)

Files

Format

msmarco-docs-sents.top.300k.json (JSONL)

One document per line:

{
  "docid": "D1650436",
  "url": "https://...",
  "title": "...",
  "body": "...",
  "sents": ["sentence 1", "sentence 2", "..."]
}

Dataset Statistics

Note: Only the Top-300K split is used. Random sampling is not used in any experiments.

Corresponding Models

Trained on this dataset:

SFT reference policies: `kiyam/ddro-msmarco-pq-sft`, `kiyam/ddro-msmarco-tu-sft`

Source Data

This dataset is derived from the Microsoft MS MARCO Document Ranking dataset:

Tri Nguyen, Mir Rosenberg, Xia Song, Jauhar Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016.

MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. NIPS 2016 Workshop on Cognitive Computation.

• —Homepage: https://microsoft.github.io/msmarco/Datasets.html#document-ranking-dataset
• —License: Microsoft Research License Terms (ms-pl)

The original data is provided by Microsoft under the MS-PL license. This processed version inherits the same license terms.

See `sample_top_docs.sh` for the full preprocessing pipeline.

Citation

If you use this dataset, please cite:

@article{mekonnen2025lightweight,
  title={Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval},
  author={Mekonnen, Kidist Amde and Tang, Yubao and de Rijke, Maarten},
  journal={arXiv preprint arXiv:2504.05181},
  year={2025}
}