README.md

---
license: other
license_name: pdfa-eng-train
license_link: LICENSE
task_categories:
- image-to-text
size_categories:
- 10M<n<100M
language:
- en
splits:
- name: train
  num_examples: 2159432

---
# Dataset Card for PDF Association dataset (PDFA)

## Dataset Description

- **Point of Contact from curators:** [Peter Wyatt, PDF Association CTO](mailto:peter.wyatt@pdfa.org)
- **Point of Contact Hugging Face:** [Pablo Montalvo](mailto:pablo@huggingface.co)

### Dataset Summary

PDFA dataset is a document dataset filtered from the SafeDocs corpus, aka CC-MAIN-2021-31-PDF-UNTRUNCATED. The original purpose of that corpus is for comprehensive pdf documents analysis. The purpose of that subset differs in that regard, as focus has been done on making the dataset machine learning-ready for vision-language models. 

<center>
    <img src="https://huggingface.co/datasets/pixparse/pdfa-english-train/resolve/main/doc_images/Nexsen_pruet.png" alt="A brochure with visible bounding boxes for lines and words" width="600" height="300">
    <p><em>An example page of one pdf document, with added bounding boxes around words (red), lines (blue) and embedded images (green). </em></p>
</center>

### Usage

This instance of PDFA is in [webdataset](https://github.com/webdataset/webdataset/commits/main) .tar format and can be used with derived forms of the `webdataset` library.
For dataloading, we recommend to use it with the [chug](https://github.com/huggingface/chug) library, an optimized library for distributed data loading.


```python
dataset = load_dataset('pixparse/pdfa-english-train', streaming=True)
print(next(iter(dataset['train'])).keys())
>> dict_keys(['__key__', '__url__', 'json', 'ocr', 'pdf', 'tif'])
```

For faster download, you can use directly the `huggingface_hub` library. Make sure `hf_transfer` is installed prior to downloading and mind that you have enough space locally. 

```python
 import os
 
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 
 from huggingface_hub import HfApi, logging
 
 #logging.set_verbosity_debug()
 hf = HfApi()
 hf.snapshot_download("pixparse/pdfa-english-train", repo_type="dataset", local_dir_use_symlinks=False)
 
```
On a normal setting, the 1.5TB can be downloaded in approximately 4 hours.

### Efficient dataloading with `chug`

# TODO The following example uses fitz which is AGPL. We should also recommend the same with pypdf.

```python
from chug import create_wds_loader, create_doc_anno_pipe

# TODO import image transforms and text transforms from pixparse.data




decoder = create_decoder_pipe(
    image_preprocess=image_preprocess, # callable of transforms to image tensors
    anno_preprocess=anno_preprocess, # callable of transforms to text into tensors of labels and targets
    image_key="pdf",
    image_fmt="RGB",
)

loader = create_wds_loader(
    "/my_data/pdfa-eng-train-*.tar",
    decoder,
    is_train=True,
    resampled=False,
    start_interval=0,
    num_samples=2159432,
    workers=8,
    batch_size=32, # adjust to your architecture capacity
    seed=seed, # set a seed
    world_size=world_size, # get world_size from your training environment
)

```

Further, a metadata file `_pdfa-english-train-info-minimal.json` contains the list of samples per shard, with same basename and `.json` or `.pdf` extension, 
as well as the count of files per shard. 


#### Words and lines document metadata

Initially, we started from the readily available ~11TB zip files from PDFA in their initial [data release](https://digitalcorpora.org/corpora/file-corpora/cc-main-2021-31-pdf-untruncated/).
From the pdf digital files, we extracted words, bounding boxes and image bounding boxes that are available in the pdf file. This information is then reshaped into lines organized in reading order, under the key `lines`. We keep non-reshaped word and bounding box information under the `word` key, should users want to use their own heuristic.

The way we obtain an approximate reading order is simply by looking at the frequency peaks of the leftmost word x-coordinate. A frequency peak means that a high number of lines are starting from the same point. Then, we keep track of the x-coordinate of each such identified column. If no peaks are found, the document is assumed to be readable in plain format.
The code to detect columns can be found here.
```python
def get_columnar_separators(page, min_prominence=0.3, num_bins=10, kernel_width=1):
    """
    Identifies the x-coordinates that best separate columns by analyzing the derivative of a histogram
    of the 'left' values (xmin) of bounding boxes.

    Args:
        page (dict): Page data with 'bbox' containing bounding boxes of words.
        min_prominence (float): The required prominence of peaks in the histogram.
        num_bins (int): Number of bins to use for the histogram.
        kernel_width (int): The width of the Gaussian kernel used for smoothing the histogram.

    Returns:
        separators (list): The x-coordinates that separate the columns, if any.
    """
    try:
        left_values = [b[0] for b in page['bbox']]
        hist, bin_edges = np.histogram(left_values, bins=num_bins)
        hist = scipy.ndimage.gaussian_filter1d(hist, kernel_width)
        min_val = min(hist)
        hist = np.insert(hist, [0, len(hist)], min_val)
        bin_width = bin_edges[1] - bin_edges[0]
        bin_edges = np.insert(bin_edges, [0, len(bin_edges)], [bin_edges[0] - bin_width, bin_edges[-1] + bin_width])

        peaks, _ = scipy.signal.find_peaks(hist, prominence=min_prominence * np.max(hist))
        derivatives = np.diff(hist)

        separators = []
        if len(peaks) > 1:
            # This finds the index of the maximum derivative value between peaks
            # which indicates peaks after trough --> column
            for i in range(len(peaks)-1):
                peak_left = peaks[i]
                peak_right = peaks[i+1]
                max_deriv_index = np.argmax(derivatives[peak_left:peak_right]) + peak_left
                separator_x = bin_edges[max_deriv_index + 1]
                separators.append(separator_x)
    except Exception as e:
        separators = []
    return separators
```

<center>
    <img src="https://huggingface.co/datasets/pixparse/pdfa-english-train/resolve/main/doc_images/columnar_detection.png" alt="A graph of leftmost x positions in a 2-columns document" width="600" height="300">
    <p><em>A graph of leftmost x-positions of bounding boxes on a 2-column (arxiv) document. Peaks are visibly detected. </em></p>
</center>

For each pdf document, we store statistics on the file size, number of words (as characters separated by spaces), number of pages, as well as the rendering times of each page for a given dpi. 
#### Filtering process

File size and page rendering time are used to set thresholds in the final dataset: the goal is to remove files that are larger than 100 MB, or that take more than 500ms to render on a modern machine, to optimize dataloading at scale. Having "too large" or "too slow" files would add a burden to large-scale training pipelines and we choose to alleviate this in the current release. Finally, a full pass over the dataset is done, trying to open and decode a bytestream from each raw object and discarding any object (pair pdf/json) that fails to be opened, to remove corrupted data.

As a last step, we use XLM-Roberta to restrict the dataset to an english subset, specifically `papluca/xlm-roberta-base-language-detection` , on the first 512 words of the first page of each document. 
Be aware that some documents may have several languages embedded in them, or that some predictions might be inaccurate. A majority of documents from the original corpus are in English language.
  
<center>
    <img src="https://huggingface.co/datasets/pixparse/pdfa-english-train/resolve/main/doc_images/languages_pdfa_xlmroberta.png" alt="A histogram of languages count in the PDFA dataset." width="600" height="300">
    <p><em>A histogram of language distribution taken on a fraction of the original -non-filtered on language- PDFA dataset. </em></p>
</center>

At the end, each document exists as a pairing of a pdf and a json file containing extensive OCR annotation as well as metadata information about rendering times. The filterings and packaging in 
webdataset format are tailored towards multimodal machine learning at scale, specifically image-to-text tasks. 

### Data, metadata and statistics.

Pdf files are coming from various sources. They are in RGB format, and contain multiple pages, and they can be rendered using the engine of your choice, here [pdf2image](https://github.com/Belval/pdf2image) .

```python
from pdf2image import convert_from_bytes

pdf_first_page = convert_from_bytes(sample['pdf'], dpi=300, first_page=1, last_page=1)[0]
```

<center>
  <img src="https://huggingface.co/datasets/pixparse/pdfa-english-train/resolve/main/doc_images/pdf_first_page.png" alt="Rendering of an image for a Grade 8 lesson plan" width="400" height="600">
</center>

The metadata for each document has been formatted in this way. Each `pdf` is paired with a `json` file with the following structure. Entries have been shortened for readability.
```json
{
  "pages": [
    {
      "words": [
        {
          "text": [
            "Health", "Smart", "Virginia", "Sample", "Lesson", "Plan", "Grade", "8", "-", "HP-7"
          ],
          "bbox": [
            [0.117647, 0.045563, 0.051981, 0.015573],
            [0.174694, 0.045563, 0.047954, 0.015573],
            [0.227643, 0.045563, 0.05983, 0.015573],
            [0.292539, 0.045563, 0.061002, 0.015573],
            [0.357839, 0.045563, 0.058053, 0.015573],
            [0.420399, 0.045563, 0.035908, 0.015573],
            [0.716544, 0.04577, 0.054624, 0.016927],
            [0.776681, 0.04577, 0.010905, 0.016927],
            [0.793087, 0.04577, 0.00653, 0.016927],
            [0.805078, 0.04577, 0.044768, 0.016927]
          ],
          "score": [
            1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
          ],
          "line_pos": [
            [0, 0], [0, 8], [0, 16], [0, 24], [0, 32], [0, 40], [0, 48], [1, 0], [2, 0], [3, 0]
          ]
        }
      ],
      "lines": [
        {
          "text": [
            "Health Smart Virginia Sample Lesson Plan Grade", "Physical", "Disease", "Health", "2020", "Grade 8 Sample Lesson Plan:"
          ],
          "bbox": [
            [0.117647, 0.045563, 0.653521, 0.016927],
            [0.716546, 0.063952, 0.07323199999999996, 0.016927],
            [0.716546, 0.082134, 0.07102200000000003, 0.016927],
            [0.716546, 0.100315, 0.05683300000000002, 0.016927],
            [0.716546, 0.118497, 0.043709, 0.016927],
            [0.27, 0.201185, 0.459554, 0.028268]
          ],
          "score": [
            1.0, 1.0, 1.0, 1.0, 1.0, 1.0
          ],
          "word_slice": [
            [0, 7], [7, 8], [8, 9], [9, 10], [10, 11], [11, 16]
          ]
        }
      ],
      "images_bbox": [
        [0.37353, 0.090907, 0.253736, 0.100189]
      ],
      "images_bbox_no_text_overlap": [
        [0.37353, 0.090907, 0.253736, 0.100189]
      ]
    }
  ]
}

```

The top-level key, `pages`, is a list of every page in the document. The above example shows only one page. 
`words` is a list of words without spaces, with their individual associated bounding box in the next entry.
`bbox` contains the bounding box coordinates in `left, top, width, height` format, with coordinates relative to the page size. 
`line_pos`, for words, is a list of tuples indicating the index of the line the word belongs to, then the starting position in that line, character-wise. 

`lines` are lines (parts of sequences, strings separated by spaces) grouped together using the heuristic detailed above. 
`bbox` contains the bounding box coordinates in `left, top, width, height` format, with coordinates relative to the page size. 

For each page, 
`images_bbox` gives the bounding boxes of the images embedded in the page. 
`images_bbox_no_text_overlap` gives a reduced list of bounding boxes that have no overlap with text found in the pdf. Text might be present as a drawing or another representation, however.
``

`score` is a placeholder of value 1.0 for the entire dataset. 

Such a formatting follows the multimodal dataset from the Industry Document Library, `https://huggingface.co/datasets/pixparse/IDL-wds`. 
Estimating the number of tokens is done using a `LlamaTokenizer` from `tokenizers`. There is a clear power law distribution with respect to data length.

<center>
    <img src="https://huggingface.co/datasets/pixparse/pdfa-english-train/resolve/main/doc_images/token_count_distribution.png" alt="A histogram of token count distribution per page" width="600" height="300">
    <p><em>A histogram of token count distribution per page, taken from a subset of the dataset. There is a visible power law. </em></p>
</center>

### Data Splits

#### Train
* `pdfa-eng-train-*.tar`
* Downloaded on 2024/01/22
* 1800 shards, 2,159,432 samples, 18M pages, 9.7 billion tokens (around 5 billion words)

## Additional Information

### Dataset Curators

Pablo Montalvo, Ross Wightman

### Disclaimer and note to researchers

This dataset is intended as an OCR-heavy pretraining basis for vision-language models. As a corpus, it does not represent the intent and purpose from  CC-MAIN-2021-31-PDF-UNTRUNCATED. The original is made to represent extant pdf data in its diversity and complexity. In particular, common issues related to misuse of pdfs such as mojibake (garbled text due to decoding erros) are yet to be addressed systematically, and this dataset present simplifications that can hide such issues found in the wild. In order to address this biases, we recommend to examine carefully both the simplified annotation and the original `pdf` data, beyond a simple rendering. 

Further, the annotation is limited to what can be extracted and is readily available - text drawn in images and only present as a bitmap rendition might be missed entirely by said annotation. 

Finally, the restriction to English language is made to alleviate difficulties related to multilingual processing so that the community can be familiarized with this optimized multimodal format. A later release will be done on the full PDFA, with splits per languages, layout types, and so on.

### Licensing Information

Data has been filtered from the original corpus. As a consequence, users should note [Common Crawl's license and terms of use](https://commoncrawl.org/terms-of-use) and the [Digital Corpora project's Terms of Use](https://digitalcorpora.org/about-digitalcorpora/terms-of-use/).