kiyam/ddro-nq-dataset

DDRO — NQ320K Processed Dataset

This dataset contains the preprocessed Natural Questions (NQ320K) corpus used to train and evaluate the DDRO generative retrieval models from:

📄 Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval (SIGIR 2025)

The raw NQ data (from Google) is processed into a unified format with document text, queries, and relevance annotations, ready for use in generative retrieval training pipelines.

---

Files

File	Description	Size
nq_merged.json	Full NQ320K document corpus (109,739 unique docs, JSONL format)	5.5 GB
nq_train.gz	Training split — queries + document fields (TSV, gzipped)	6.5 GB
nq_val.gz	Validation split — queries + document fields (TSV, gzipped)	159 MB

---

Format

nq_merged.json (JSONL)

One document per line:

{
  "id": "5655493461695504401",
  "query": "what is the most common use of opt-in email marketing",
  "long_answer": "...",
  "short_answer": "a newsletter sent to an advertising firm's customers",
  "title": "email marketing",
  "abstract": "...",
  "content": "...",
  "document_url": "https://en.wikipedia.org/...",
  "doc_tac": "<title> + <abstract> + <content>",
  "language": "en"
}

nq_train.gz / nq_val.gz (TSV, gzipped)

Tab-separated, no header. Columns: query, id, long_answer, short_answer, title, abstract, content, document_url, doc_tac, language

---

Dataset Statistics

Split	Examples
Train	~307,373 (raw) → 87,791 (after dedup by title)
Dev	~7,830 (raw) → 21,948 (after dedup)
Unique documents	109,739

---

Usage

Load the corpus

import json

docs = {}
with open("nq_merged.json") as f:
    for line in f:
        doc = json.loads(line)
        docs[doc["id"]] = doc

Load train/val splits

import gzip
import pandas as pd

columns = ["query", "id", "long_answer", "short_answer", "title",
           "abstract", "content", "document_url", "doc_tac", "language"]

with gzip.open("nq_train.gz", "rt") as f:
    train_df = pd.read_csv(f, sep="\t", names=columns)

---

Corresponding Models

Trained on this dataset:

Model	Docid Type	MRR@10	R@10
kiyam/ddro-nq-pq	PQ (Product Quantization)	55.51	67.31
kiyam/ddro-nq-tu	TU (Title + URL)	45.99	55.98

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

---

Source Data

This dataset is derived from the Google Natural Questions (NQ) dataset:

Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural Questions: a Benchmark for Question Answering Research. Transactions of the Association of Computational Linguistics.

• Homepage: https://ai.google.com/research/NaturalQuestions
• License: Creative Commons Attribution-ShareAlike 3.0
• Raw files: gs://natural_questions/v1.0-simplified/

The original data is made available by Google under CC BY-SA 3.0. This processed version inherits the same license terms.

See process_nq_dataset.py for the full preprocessing pipeline.

---

Citation

@article{mekonnen2025lightweight,
author = {Mekonnen, Kidist Amde and Tang, Yubao and de Rijke, Maarten},
title = {Lightweight and Direct Document Relevance Optimization for Generative Information Retrieval},
year = {2025},
isbn = {9798400715921},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3726302.3730023},
doi = {10.1145/3726302.3730023},
abstract = {Generative information retrieval (GenIR) is a promising neural retrieval paradigm that formulates document retrieval as a document identifier (docid) generation task, allowing for end-to-end optimization toward a unified global retrieval objective. However, existing GenIR models suffer from token-level misalignment, where models trained to predict the next token often fail to capture document-level relevance effectively. While reinforcement learning-based methods, such as reinforcement learning from relevance feedback (RLRF), aim to address this misalignment through reward modeling, they introduce significant complexity, requiring the optimization of an auxiliary reward function followed by reinforcement fine-tuning, which is computationally expensive and often unstable. To address these challenges, we propose direct document relevance optimization (DDRO), which aligns token-level docid generation with document-level relevance estimation through direct optimization via pairwise ranking, eliminating the need for explicit reward modeling and reinforcement learning. Experimental results on benchmark datasets, including MS MARCO document and Natural Questions, show that DDRO outperforms reinforcement learning-based methods, achieving a 7.4\% improvement in MRR@10 for MS MARCO and a 19.9\% improvement for Natural Questions. These findings highlight DDRO's potential to enhance retrieval effectiveness with a simplified optimization approach. By framing alignment as a direct optimization problem, DDRO simplifies the ranking optimization pipeline of GenIR models while offering a viable alternative to reinforcement learning-based methods},
booktitle = {Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval},
pages = {1327–1338},
numpages = {12},
keywords = {document relevance optimization, generative information retrieval, learning to rank, ranking optimization, supervised fine-tuning},
location = {Padua, Italy},
series = {SIGIR '25}
}