UKPLab/ProReviewer-Dataset

ProReviewer Dataset

A version-matched dataset of ICLR paper submissions paired with their initial peer reviews and scores. Each sample contains the initial submission of a paper matched with the reviewers' initial ratings (before the discussion phase), enabling research on automated scientific peer review.

Dataset Description

This dataset supports training and evaluating automated peer review systems. All papers and reviews are version-matched: the paper content corresponds to the initial submission, and the review scores reflect the initial ratings assigned before any author-reviewer discussion.

Splits

Split	Source	Samples	Description
train	ICLR 2025	4,011	Training set for RL/SFT
test	ICLR 2026	1,000	Held-out evaluation set

Features

Field	Type	Description
paper_id	string	OpenReview paper ID
arxiv_id	string	arXiv identifier
title	string	Paper title
markdown	dict	Paper content (initial submission) in markdown with metadata (authors, char/line counts)
reviews	list	List of peer reviews (see below)
scores	dict	Aggregated initial scores across reviewers (rating_avg, confidence_avg, etc.)
metadata	dict	Submission metadata (primary_area, track, venue)
meta_review	dict	Area chair meta-review and additional comments
decision	dict	Final acceptance decision (e.g., Accept (Poster), Reject)

Each review contains:

Field	Type	Description
id	string	Reviewer ID
initial_rating	int	Rating before discussion (1-10)
confidence	int	Reviewer confidence score
soundness	int	Technical soundness score
contribution	int	Contribution significance score
presentation	int	Presentation quality score
summary	string	Paper summary by reviewer
strengths	string	Identified strengths
weaknesses	string	Identified weaknesses
questions	string	Questions for the authors

Usage

from datasets import load_dataset

ds = load_dataset("UKPLab/ProReviewer-Dataset")

# Access a training sample
paper = ds["train"][0]
print(paper["title"])
print(f"Avg initial rating: {paper['scores']['rating_avg']}")
print(f"Decision: {paper['decision']['decision']}")
print(f"Number of reviews: {len(paper['reviews'])}")

# Read paper content (initial submission)
content = paper["markdown"]["content"]

# Access individual reviews
for review in paper["reviews"]:
    print(f"Initial rating: {review['initial_rating']}, Confidence: {review['confidence']}")

Version Matching

All data is version-matched to ensure consistency:

• Paper content: Corresponds to the initial submission (before any revisions in response to reviews)
• Review scores: Reflect the initial ratings assigned by reviewers (before the discussion/rebuttal phase)

This version matching is important for training review agents, as it ensures the model learns to evaluate papers as they were first submitted, with scores that reflect first impressions rather than post-discussion adjustments.

Source

Papers and reviews were collected from OpenReview for the ICLR 2025 and ICLR 2026 venues.

Associated Resources

• Paper: From Passive Generation to Investigation: A Proactive Scientific Peer Review Agent
• Code: UKPLab/arxiv2026-ProReviewer

Citation

@article{fang2026passive,
  title={From Passive Generation to Investigation: A Proactive Scientific Peer Review Agent},
  author={Fang, Haishuo and Feng, Yue and Gurevych, Iryna},
  journal={arXiv preprint arXiv:2606.13349},
  year={2026}
}

License

This dataset is released under the MIT License.