Year	Reviewable	Complete	%	Reject	Poster	Spotlight	Oral
2020	2,213	2,190	99.0%	1,526	531	108	48
2021	2,594	2,562	98.8%	1,735	692	114	53
2022	2,617	2,543	97.2%	1,523	865	174	55
2023	3,792	3,612	95.3%	2,219	1,202	280	91
2024	5,780	5,498	95.1%	3,519	1,808	367	86
2025	8,727	7,956	91.2%	5,019	3,115	380	213
2026	15,948	9,189	57.6%	0	0	0	0
Total	41,671	33,550	80.5%	15,541	8,213	1,423	546

Notes:

Reviewable = Total submissions minus Withdrawn and Desk Rejected
Complete = Has MD file with detected Abstract and References sections
Decision counts are for reviewable papers only

Schema

Each row contains:

Field	Type	Description
`submission`	dict	Paper metadata (id, title, abstract, decision, authors, etc.)
`review_1` to `review_12`	dict/None	Reviewer assessments (None if fewer reviews)
`meta_review`	dict/None	Area Chair assessment
`raw_md`	str	Raw markdown from PDF conversion
`clean_md`	str	Cleaned markdown (Introduction → References)
`clean_md_sections`	dict	Mapping of section titles to content
`md_path`	str	Local path to markdown file
`pdf_path`	str/None	Local path to PDF file (None if not found)

Submission Fields

{
    "id": "abc123xyz",           # OpenReview ID
    "title": "Paper Title",
    "abstract": "Abstract text...",
    "decision": "Accept (Poster)",
    "authors": ["Author One", "Author Two"],
    "keywords": ["machine learning", "nlp"],
    "venue": "ICLR 2024 poster",
    "pdf_url": "/pdf/abc123xyz.pdf",
    "created_date": 1699574400,  # Unix timestamp
    "modified_date": 1699574400,
    "tldr": "Short summary...",
    "primary_area": "machine learning",
    "google_scholar_citations": 42,  # Google Scholar citation count (None if not found)
}

Review Fields

Review schemas vary by year due to OpenReview changes:

Year	Rating Field	Key Content Fields
2020	`rating` (str)	`review`, `experience_assessment`, `review_assessment_*`
2021	`rating` (str)	`review`, `confidence`
2022	`recommendation` (str)	`main_review`, `summary_of_the_paper`, `correctness`, `novelty`
2023	`recommendation` (str)	`strength_and_weaknesses`, `summary_of_the_paper`, `correctness`, `novelty`
2024-2026	`rating` (int)	`summary`, `strengths`, `weaknesses`, `soundness`, `presentation`, `contribution`

All reviews include:

number_of_author_responses: Author replies to this review
number_of_reviewer_responses_to_author: Reviewer follow-ups

Rating Scales

Year	Rating Scale	Confidence	Sub-scores
2020	1, 3, 6, 8 (str)	N/A	N/A
2021	1-10 (str)	1-5 (str)	N/A
2022-2023	1-10 (str)	1-5 (str)	correctness, novelty (str 1-4)
2024-2026	1-10 (int)	1-5 (int)	soundness, presentation, contribution (int 1-4)

Markdown Normalization

The clean_md field contains normalized markdown produced from the raw PDF conversion. The normalization pipeline:

1. Validation

Paper must have both Abstract and References sections detected
Papers failing validation are excluded from the dataset

2. Content Clipping

Start: First section header AFTER Abstract (typically "Introduction")
End: End of References section (before Appendix/Supplementary)
This removes: title, authors, abstract, appendices, supplementary material

3. Section Removal

Acknowledgements: Removed entirely (to preserve anonymity for blind review analysis)
Reproducibility: Removed entirely (often contains author-identifying information)

4. Artifact Cleaning

Line numbers: Removed (e.g., **054 055 056** remnants from submitted PDFs)
Standalone number lines: Removed (bare PDF line numbers like 327, 337 338)
Page anchors: Lines containing <span id="page-X">...<sup> removed entirely
Code/GitHub refs: Entire sentences containing code...https://github... removed (author code)
Footnotes: Removed except those referencing figures
Dagger markers: Removed (†, ‡) except figure references

5. Header Normalization

All headers normalized to single # level
Titles converted to UPPERCASE
Span tags and bold markers removed
Example: ## 3.1 **<span>Methods</span>** → # 3.1 METHODS

6. Whitespace Normalization

Multiple blank lines collapsed to single blank line
Trailing whitespace stripped

Section Breakdown

The clean_md_sections field provides a dict mapping normalized section titles to content:

{
    "INTRODUCTION": "Section content...",
    "RELATED WORK": "Section content...",
    "METHODS": "Section content...",
    "EXPERIMENTS": "Section content...",
    "CONCLUSION": "Section content...",
    "REFERENCES": "Reference list..."
}

Note: Section titles vary by paper. Common sections include INTRODUCTION, RELATED WORK, METHOD/METHODS, EXPERIMENTS, RESULTS, DISCUSSION, CONCLUSION, REFERENCES.

Usage

from datasets import load_dataset

# Load a specific year (as a config/subset)
ds = load_dataset("skonan/iclr-data-2020-2026", "2024")

# Load default (most recent year)
ds = load_dataset("skonan/iclr-data-2020-2026")

# Access data
for row in ds["train"]:
    print(row["submission"]["title"])
    print(row["submission"]["decision"])

    # Access reviews
    if row["review_1"]:
        print(row["review_1"]["rating"])

    # Access sections
    intro = row["clean_md_sections"].get("INTRODUCTION", "")
    print(intro[:500])

Data Source

Data extracted from OpenReview using the OpenReview API. Paper PDFs converted to markdown using Marker.

License

Apache 2.0

Citation

If you use this dataset, please cite:

@misc{iclr-data-2020-2026,
  title={ICLR Reviews Dataset 2020-2026},
  author={OpenReview Community},
  year={2024},
  howpublished={HuggingFace Datasets},
  url={https://huggingface.co/datasets/skonan/iclr-data-2020-2026}
}

Downloads last month: 82

Size of downloaded dataset files:

4.97 GB

Size of the auto-converted Parquet files:

4.97 GB

Number of rows:

33,550