README.md

---
license: cc-by-4.0
task_categories:
- image-text-to-text
- image-feature-extraction
language:
- en
tags:
- pdf
- ocr
- legal
- government
size_categories:
- 100K<n<1M
dataset_info:
- config_name: index
  features:
  - name: filename
    dtype: string
  - name: filepath
    dtype: string
  - name: broken_pdf
    dtype: bool
  - name: num_pages
    dtype: float64
  - name: created_date
    dtype: string
  - name: modified_date
    dtype: string
  - name: title
    dtype: string
  - name: author
    dtype: string
  - name: subject
    dtype: string
  - name: file_size_mb
    dtype: float64
  - name: error_message
    dtype: string
  splits:
  - name: train
    num_bytes: 39695484
    num_examples: 229917
  download_size: 19387703
  dataset_size: 39695484
- config_name: sample
  features:
  - name: pdf
    dtype: pdf
  - name: num_pages
    dtype: float64
  - name: created_date
    dtype: string
  - name: modified_date
    dtype: string
  - name: title
    dtype: string
  - name: author
    dtype: string
  - name: subject
    dtype: string
  - name: file_size_mb
    dtype: float64
  - name: broken_pdf
    dtype: bool
  - name: error_message
    dtype: string
  splits:
  - name: train
    num_bytes: 879832.0
    num_examples: 5000
  download_size: 400528
  dataset_size: 879832.0
configs:
- config_name: index
  data_files:
  - split: train
    path: index/train-*
- config_name: sample
  data_files:
  - split: train
    path: sample/train-*
---

# govdocs1: source PDF files

> [!NOTE]
> Converted versions of other document types (word, txt, etc) are available [in this repo](https://huggingface.co/datasets/BEE-spoke-data/govdocs1-by-extension)

This is ~220,000 open-access PDF documents (about 6.6M pages) from the dataset [govdocs1](https://digitalcorpora.org/corpora/file-corpora/files/). It wants to be OCR'd.


- Uploaded as `tar` file pieces of ~10 GiB each due to size/file count limits with an [index.csv](data/index.csv) covering details
- 5,000 randomly sampled PDFs are available unarchived in `sample/`. Hugging Face supports previewing these in-browser, for example [this one](sample/001070.pdf)

## Recovering the data

Download the `data/` directory (with `huggingface-cli download` or similar) extract the tar pieces:

```sh
cat data_pdfs_part.tar.* | tar -xf - && rm data_pdfs_part.tar.*
```
 
## processing details

### duplicates

exact duplicate PDFs were removed with `jdupes`. See the [log file](exact_duplicate_removal.log) for details.  

---


## By the numbers

Based on the [index.csv](data/index.csv)

### Dataset Overview

| Metric | Value | Percentage |
|--------|-------|------------|
| Total Documents | 229,917 | 100% |
| Successfully Processed | 229,824 | 99.96% |
| Broken/Corrupted | 93 | 0.04% |
| Unique Filenames | 229,917 | 100% |

### Document Structure

#### Page Count Distribution

| Pages | Count | Percentage |
|-------|-------|------------|
| 2 pages | 21,887 | 9.5% |
| 1 page | 19,282 | 8.4% |
| 4 pages | 14,640 | 6.4% |
| 3 pages | 12,861 | 5.6% |
| 6 pages | 9,770 | 4.3% |

| Statistic | Value |
|-----------|-------|
| **Range** | 1 - 3,200 pages |
| **Mean** | 27.8 pages |
| **Median** | 10 pages |
| **Standard Deviation** | 67.9 pages |

#### File Size Distribution

| Size (MB) | Count | Percentage |
|-----------|-------|------------|
| 0.02 | 13,427 | 5.8% |
| 0.03 | 12,142 | 5.3% |
| 0.04 | 12,085 | 5.3% |
| 0.05 | 11,850 | 5.2% |
| 0.01 | 9,929 | 4.3% |

| Statistic | Value |
|-----------|-------|
| **Range** | 0 - 68.83 MB |
| **Mean** | 0.565 MB |
| **Median** | 0.15 MB |
| **Standard Deviation** | 1.134 MB |

### Metadata Completeness Crisis

| Field | Missing | Present | Completeness |
|-------|---------|---------|--------------|
| **Subject** | 182,430 | 47,487 | **20.6%** |
| **Author** | 78,269 | 151,648 | **66.0%** |
| **Title** | 51,514 | 178,403 | **77.6%** |
| **Created Date** | 3,260 | 226,657 | **98.6%** |

#### Title Quality Breakdown

| Title Type | Count | Percentage |
|------------|-------|------------|
| Missing (None) | 51,514 | 22.4% |
| Generic "Document" | 11,699 | 5.1% |
| "untitled" | 2,081 | 0.9% |
| Meaningful titles | ~165,000 | 71.6% |

#### Top Authors

| Author | Count |
|--------|-------|
| U.S. Government Printing Office | 11,838 |
| Unknown | 3,477 |
| Administrator | 1,630 |
| U.S. Government Accountability Office | 1,390 |

#### Top Subjects

| Subject | Count |
|---------|-------|
| Extracted Pages | 11,692 |
| NIOSH HHE REPORT | 466 |
| CMS Opinion Template | 353 |
| SEC Financial Proposals Summary | 230 |

### Processing Errors

| Error Type | Count | Percentage |
|------------|-------|------------|
| Could not read Boolean object | 46 | 49.5% |
| cryptography>=3.1 required for AES | 15 | 16.1% |
| Stream ended unexpectedly | 9 | 9.7% |
| 'NullObject' has no attribute 'get' | 5 | 5.4% |
| Other errors | 18 | 19.4% |

### Temporal Coverage

| Date Field | Range | Issues |
|------------|-------|--------|
| **Modified Date** | 1979-12-31 to 2025-03-31 | (dates in 2023-2025 are incorrect/defaulted to) |
| **Created Date** | Various formats | 1,573 invalid "D:00000101000000Z" |

### Critical Assessment

> [!NOTE]
> Generated by Claude Sonnet-4, unsolicited (_as always_)

#### Data Quality Issues

| Issue | Severity | Impact |
|-------|----------|---------|
| **Metadata Poverty** | **CRITICAL** | 79% missing subjects kills discoverability |
| **Title Degradation** | **HIGH** | 28% generic/missing titles |
| **Date Inconsistencies** | **MEDIUM** | Invalid formats, future dates |
| **Processing Errors** | **LOW** | 0.04% failure rate acceptable |

#### Key Insights

**Document Profile**: Typical government PDF = 10 pages, 0.15 MB, metadata-poor

**Fatal Flaw**: This dataset has excellent technical extraction (99.96% success) but catastrophic intellectual organization. You're essentially working with 230K unlabeled documents.

**Bottom Line**: The structural data is solid, but without subject classification for 79% of documents, this is an unindexed digital landfill masquerading as an archive.

---