---
pretty_name: Deduped Paper Text Dataset
viewer: true
tags:
- datasets
- arxiv
- scientific-papers
- text
size_categories:
- 1M<n<10M
---

# Deduped Paper Text Dataset

This dataset merges an existing paper-text parquet export with one or more
backfill parquet shards, then keeps exactly one row per `canonical_paper_id`.

One row corresponds to one canonical arXiv paper. Where multiple versions are
available, the merge keeps the strongest text row according to the repo's
selection logic: prefer non-partial text, then prefer raw-PDF-derived text,
then newer versions, then longer text.

## Important provenance note

This is a merged release, not a fresh one-pass export. It combines an earlier
deduped paper-text dataset with later backfill shards, then dedupes again on
`canonical_paper_id`.

Use per-paper `license` filtering before broad downstream publication. Many
arXiv records still carry `nonexclusive-distrib` rather than a general reuse
license.

This release is capped deterministically at exactly `2000000` deduped papers. The selected inputs contained `2364986` deduped papers before applying the cap.

## Rows

- `train`: `2000000` papers
- unique canonical papers: `2000000`
- base input rows considered: `1000000`
- backfill input rows considered: `1364986`
- deduped rows before target cap: `2364986`


## Main columns

- `paper_id`
- `canonical_paper_id`
- `paper_version`
- `pdf_path`
- `title`
- `abstract`
- `authors`
- `categories`
- `license`
- `text`
- `text_char_count`
- `token_count`
- `page_count`
- `token_types`

## Files

- `train_00000.parquet`
- `train_00001.parquet`
- `train_00002.parquet`
- `train_00003.parquet`
- `train_00004.parquet`
- `train_00005.parquet`
- `train_00006.parquet`
- `train_00007.parquet`
- `train_00008.parquet`
- `train_00009.parquet`
- `train_00010.parquet`
- `train_00011.parquet`
- `train_00012.parquet`
- `train_00013.parquet`
- `train_00014.parquet`
- `train_00015.parquet`
- `train_00016.parquet`
- `train_00017.parquet`
- `train_00018.parquet`
- `train_00019.parquet`