ParaPaper

English | 中文

Overview

ParaPaper is a private research dataset containing JSON files converted from accepted paper PDFs across six AI/NLP/ML conferences:

• AAAI
• ACL
• EMNLP
• ICLR
• ICML
• NeurIPS

The dataset is intended for internal research workflows such as paper-level context construction, scientific writing models, research-idea reconstruction, reviewer-risk analysis, and SFT data generation for the ParadoxGPT project.

It does not include original PDF files. It contains parsed and cleaned JSON representations derived from PDFs.

Repository Structure

The data is uploaded as one compressed archive per conference:

data/
  aaai.tar.zst
  acl.tar.zst
  emnlp.tar.zst
  iclr.tar.zst
  icml.tar.zst
  neurips.tar.zst
manifest.json
SHA256SUMS
CITATION.cff
README.md
LICENSE

After extraction, each conference archive follows this layout:

<conference>/
  <year>/
    <paper_id_or_slug>/
      cleaned_paper.json
      paper_record.json
      context_pack.json

Dataset Statistics

Total successful parsed papers: 16,471

Conference	Years	Papers
AAAI	2024, 2025, 2026	3,963
ACL	2023, 2024, 2025	2,804
EMNLP	2023, 2024	1,684
ICLR	2024, 2025, 2026	2,800
ICML	2023, 2024, 2025	1,959
NeurIPS	2023, 2024, 2025	3,261

Each successfully parsed paper contains:

• cleaned_paper.json
• paper_record.json
• context_pack.json

Excluded:

• original PDFs
• _failed parsing outputs
• raw MinerU intermediate files

Archive Sizes

Archive	Size
data/aaai.tar.zst	74 MB
data/acl.tar.zst	49 MB
data/emnlp.tar.zst	30 MB
data/iclr.tar.zst	296 MB
data/icml.tar.zst	141 MB
data/neurips.tar.zst	250 MB

Checksums are provided in SHA256SUMS.

How to Download

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="bhxdianzhang/ParaPaper",
    repo_type="dataset",
    local_dir="./ParaPaper",
    token=True,
)
print(local_dir)

Because this is a private dataset, you need a Hugging Face token with access to the repository.

How to Extract

mkdir -p parapaper_json
tar --zstd -xf data/iclr.tar.zst -C parapaper_json
tar --zstd -xf data/icml.tar.zst -C parapaper_json
tar --zstd -xf data/neurips.tar.zst -C parapaper_json
tar --zstd -xf data/aaai.tar.zst -C parapaper_json
tar --zstd -xf data/acl.tar.zst -C parapaper_json
tar --zstd -xf data/emnlp.tar.zst -C parapaper_json

Verify checksums:

sha256sum -c SHA256SUMS

File Schemas

cleaned_paper.json

This is the most complete cleaned representation of one paper.

Important top-level fields:

Field	Type	Description
paper_id	string	Stable paper identifier used by the local pipeline.
source	object	Source paths and parsing provenance, including PDF path and MinerU outputs.
metadata	object	Title, authors, venue, year, DOI/arXiv field if available, and paper hash.
section_outline	array	Detected section headings and outline structure.
blocks	array	Cleaned text/table/formula/caption blocks with block-level structure.
abstract	object	Abstract text and source block ids.
introduction	object	Introduction text and paragraph segmentation.
section_texts	object	Extracted coarse sections such as preliminary, method, experiments, conclusion.
figure_captions	array	Parsed figure captions.
table_captions	array	Parsed table captions.
source_map	array	Mapping between cleaned content and source blocks.
full_text	string	Full cleaned body text after processing.
body_text_no_references	string	Body text with references/appendix-like trailing material removed when detected.
cleaning_report	object	Quality and parsing signals.

paper_record.json

This is a normalized paper-level record for downstream annotation.

Important top-level fields:

Field	Type	Description
paper_id	string	Stable paper identifier.
source	object	Paths to cleaned JSON and original source information.
metadata	object	Title, authors, venue, year, DOI/arXiv field if available, paper hash.
abstract	object	Abstract text and block ids.
introduction	object	Introduction text, paragraphs, paragraph count, and block ids.
preliminary	object	Preliminary/background section text and presence flag.
method	object	Method section text and presence flag.
experiments	object	Experiment section text and presence flag.
conclusion	object	Conclusion section text and presence flag.
figure_captions	array	Figure captions.
table_captions	array	Table captions.
section_outline	array	Section outline.
cleaning_report	object	Cleaning diagnostics.
quality_signals	object	Boolean quality indicators used by downstream filters.

context_pack.json

This is the compact downstream context used by annotation and SFT construction scripts.

Important top-level fields:

Field	Type	Description
paper_id	string	Stable paper identifier.
metadata	object	title, venue, and year.
targets	object	Target writing fields: abstract, introduction_full, and introduction_paragraphs.
context	object	Supporting context: body_without_intro_abstract, figure_table_captions, section_outline.
usable	boolean	Whether this paper passed minimal context checks.
stats	object	Character counts and paragraph counts for abstract, introduction, body context, and captions.

Intended Use

This dataset is intended for:

• building paper-level context packs
• constructing SFT data for scientific writing assistants
• retrospective research-idea reconstruction
• experiment-design and reviewer-risk annotation
• internal evaluation of paper-level reasoning pipelines

Limitations

• The data is automatically parsed from PDFs and may contain parsing errors.
• Some section boundaries may be imperfect.
• Mathematical notation, tables, captions, and references may be partially noisy.
• The dataset should not be treated as a canonical source of the original papers.
• Users should cite and consult the original papers for authoritative content.

License and Use Restrictions

This repository is marked with license: other.

The JSON files are derived from copyrighted academic papers. Copyright of the original papers remains with their authors and/or publishers. This private dataset is provided for internal research and engineering use only. Redistribution or public release should be reviewed separately against the source venues' policies and applicable copyright rules.

Citation

If you use this dataset internally or reference it in related artifacts, cite:

@dataset{zhang2026parapaper,
  title        = {ParaPaper: Parsed JSON Research Corpus from Six AI/NLP/ML Conferences},
  author       = {Heng Zhang},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/bhxdianzhang/ParaPaper}},
  note         = {Private dataset}
}

---

中文

概述

ParaPaper 是一个私有研究数据集，包含从 6 个 AI/NLP/ML 顶会论文 PDF 转换得到的 JSON 文件：

• AAAI
• ACL
• EMNLP
• ICLR
• ICML
• NeurIPS

这个数据集服务于 ParadoxGPT 项目的内部研究流程，例如 paper-level context 构建、科研写作模型、idea 反推重构、reviewer risk 分析，以及 SFT 数据构造。

本数据集不包含原始 PDF，只包含由 PDF 解析和清洗得到的 JSON 表示。

仓库结构

数据按会议压缩上传：

data/
  aaai.tar.zst
  acl.tar.zst
  emnlp.tar.zst
  iclr.tar.zst
  icml.tar.zst
  neurips.tar.zst
manifest.json
SHA256SUMS
CITATION.cff
README.md
LICENSE

解压后，每个会议目录结构如下：

<conference>/
  <year>/
    <paper_id_or_slug>/
      cleaned_paper.json
      paper_record.json
      context_pack.json

数据统计

成功解析论文总数：16,471

会议	年份	论文数
AAAI	2024, 2025, 2026	3,963
ACL	2023, 2024, 2025	2,804
EMNLP	2023, 2024	1,684
ICLR	2024, 2025, 2026	2,800
ICML	2023, 2024, 2025	1,959
NeurIPS	2023, 2024, 2025	3,261

每篇成功解析论文包含：

• cleaned_paper.json
• paper_record.json
• context_pack.json

不包含：

• 原始 PDF
• _failed 解析失败输出
• MinerU 原始中间文件

下载方式

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="bhxdianzhang/ParaPaper",
    repo_type="dataset",
    local_dir="./ParaPaper",
    token=True,
)
print(local_dir)

由于这是 private dataset，需要使用有访问权限的 Hugging Face token。

解压方式

mkdir -p parapaper_json
tar --zstd -xf data/iclr.tar.zst -C parapaper_json
tar --zstd -xf data/icml.tar.zst -C parapaper_json
tar --zstd -xf data/neurips.tar.zst -C parapaper_json
tar --zstd -xf data/aaai.tar.zst -C parapaper_json
tar --zstd -xf data/acl.tar.zst -C parapaper_json
tar --zstd -xf data/emnlp.tar.zst -C parapaper_json

校验：

sha256sum -c SHA256SUMS

字段说明

cleaned_paper.json

这是每篇论文最完整的清洗后表示。

重要字段：

字段	类型	说明
paper_id	string	本地流水线使用的稳定论文 ID。
source	object	来源路径和解析来源信息，包括 PDF 路径和 MinerU 输出。
metadata	object	标题、作者、会议、年份、DOI/arXiv 字段和 paper hash。
section_outline	array	检测到的章节标题和大纲结构。
blocks	array	清洗后的文本、表格、公式、caption 等 block。
abstract	object	摘要文本和来源 block id。
introduction	object	Introduction 文本和段落切分。
section_texts	object	preliminary、method、experiments、conclusion 等粗粒度章节。
figure_captions	array	图 caption。
table_captions	array	表 caption。
source_map	array	清洗内容与原始 block 的映射。
full_text	string	清洗后的完整正文文本。
body_text_no_references	string	尽量移除 references/appendix 后的正文文本。
cleaning_report	object	清洗和解析质量信号。

paper_record.json

这是面向下游标注的标准论文记录。

重要字段：

字段	类型	说明
paper_id	string	稳定论文 ID。
source	object	cleaned JSON 和原始来源路径。
metadata	object	标题、作者、会议、年份、DOI/arXiv 字段和 paper hash。
abstract	object	摘要文本和 block ids。
introduction	object	Introduction 文本、段落、段落数和 block ids。
preliminary	object	Preliminary/background 文本和是否存在。
method	object	Method 文本和是否存在。
experiments	object	Experiment 文本和是否存在。
conclusion	object	Conclusion 文本和是否存在。
figure_captions	array	图 caption。
table_captions	array	表 caption。
section_outline	array	章节大纲。
cleaning_report	object	清洗诊断信息。
quality_signals	object	下游过滤使用的质量信号。

context_pack.json

这是下游 annotation 和 SFT 构造最常用的紧凑上下文。

重要字段：

字段	类型	说明
paper_id	string	稳定论文 ID。
metadata	object	title、venue 和 year。
targets	object	写作目标字段：abstract、introduction_full、introduction_paragraphs。
context	object	支撑上下文：body_without_intro_abstract、figure_table_captions、section_outline。
usable	boolean	是否通过最小上下文质量检查。
stats	object	abstract、intro、body context、captions 的字符数和段落数。

适用场景

本数据集适用于：

• 构建 paper-level context pack
• 构造科研写作助手 SFT 数据
• retrospective research idea reconstruction
• 实验设计和 reviewer risk 标注
• paper-level reasoning pipeline 的内部评估

局限性

• 数据由 PDF 自动解析得到，可能存在解析错误。
• 章节边界可能不完美。
• 数学符号、表格、caption 和 references 可能有噪声。
• 不应把本数据集视为原论文的权威版本。
• 权威内容应以原论文为准，并应引用原论文。

许可和使用限制

本仓库标记为 license: other。

JSON 文件由受版权保护的学术论文解析得到。原论文版权归原作者和/或出版方所有。这个 private dataset 仅用于内部研究和工程用途。任何再分发或公开发布都应单独检查来源会议政策和相关版权规则。

引用方式

如果在内部项目、实验记录或相关产物中使用本数据集，可以引用：

@dataset{zhang2026parapaper,
  title        = {ParaPaper: Parsed JSON Research Corpus from Six AI/NLP/ML Conferences},
  author       = {Heng Zhang},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/bhxdianzhang/ParaPaper}},
  note         = {Private dataset}
}