Bo10k PDF performance dataset
A larger corpus for throughput and robustness experiments across many documents. Its shorter median document makes document construction and loading costs especially relevant.
The document selection follows NVIDIA NeMo Retriever's 26.03 evaluation resources. PDFs come from Digital Corpora's CC-MAIN-2021-31-PDF-UNTRUNCATED collection; numeric filenames preserve the source document IDs.
Corpus statistics
| Metric | Value |
|---|---|
| PDF documents | 10,000 |
| Total pages | 213,538 |
| PDF size | 29.326 GB (27.312 GiB) |
| Mean pages/document | 21.35 |
| Median pages/document | 6.00 |
| Minimum / maximum pages | 1 / 2,227 |
| 25th / 75th percentile | 2.00 / 19.00 |
| 90th / 95th / 99th percentile | 50.00 / 92.00 / 226.01 |
| Single-page documents | 1,895 |
Statistics count every PDF in pdf/, using pypdfium2 5.14.0, on 2026-11-06.
All PDFs could be opened and their pages counted; this is not a rendering or PDF-conformance check.
Percentiles use linear interpolation. Sizes sum PDF file bytes and exclude Git/LFS storage.
Document length and page distribution
| Pages/document | Documents | Document share | Pages | Page share |
|---|---|---|---|---|
| 1 | 1,895 | 18.9% | 1,895 | 0.9% |
| 2-10 | 4,473 | 44.7% | 20,016 | 9.4% |
| 11-50 | 2,655 | 26.6% | 61,848 | 29.0% |
| 51-100 | 534 | 5.3% | 37,658 | 17.6% |
| 101-500 | 423 | 4.2% | 74,522 | 34.9% |
| 501-1000 | 15 | 0.1% | 10,443 | 4.9% |
| 1001+ | 5 | 0.1% | 7,156 | 3.4% |
The left panel compares the share of documents with the share of pages in each length range. The right panel orders documents from longest to shortest and shows their cumulative share of all pages. The diagonal represents equal page counts per document.
Documents with more than 100 pages: 443, containing 43.1% of all pages. The ten longest documents account for 5.3% of all pages. Page counts describe corpus size; images, fonts and page content also affect parsing and rendering time.
Longest documents
| Pages | Size (MB) | |
|---|---|---|
| 1593260.pdf | 2,227 | 71.72 |
| 1781090.pdf | 1,537 | 66.41 |
| 1302892.pdf | 1,246 | 123.59 |
| 2216123.pdf | 1,120 | 169.66 |
| 2048697.pdf | 1,026 | 138.79 |
Source and evaluation context
- NeMo Retriever 26.3.0 extraction notebooks introduce Bo20, Bo767 and Bo10k.
- The 26.03 download notebook defines Bo20's IDs and downloads the selected PDFs.
- Bo767 IDs and Bo10k IDs define the larger selections.
- Digital Corpora source archives contain the original PDF files.
- The Bo767 retrieval notebook evaluates retrieval after extraction and embedding, using queries and ground truth beyond these PDFs. Parser throughput measurements do not reproduce that retrieval evaluation.
For reproducible performance comparisons, record the dataset commit, exact PDF filenames, page counts and file hashes alongside the parser version and benchmark configuration.
The generated statistics.json contains these descriptive statistics and per-document page counts and sizes.
Adding PDFs with Git LFS
PDFs must match *.pdf filter=lfs diff=lfs merge=lfs -text in .gitattributes before staging.
git lfs install --local
git lfs track "*.pdf"
cp /path/to/new.pdf pdf/
git add .gitattributes pdf/new.pdf
git lfs status
git commit -s -m "Add PDF"
git lfs push origin main
git push origin main
Recompute the statistics, figures and this README whenever the corpus changes.
- Downloads last month
- 291
