README.md

metadata

license: cc-by-4.0
configs:
  - config_name: COCOtext
    data_files:
      - split: train
        path: COCOtext/train-*
      - split: val
        path: COCOtext/val-*
  - config_name: TextOCR
    data_files:
      - split: train
        path: TextOCR/train-*
      - split: val
        path: TextOCR/val-*
  - config_name: pubtables-1m
    data_files:
      - split: train
        path: pubtables-1m/train-*
      - split: val
        path: pubtables-1m/val-*
      - split: test
        path: pubtables-1m/test-*
dataset_info:
  - config_name: COCOtext
    features:
      - name: sample_id
        dtype: string
      - name: dataset_name
        dtype: string
      - name: task_name
        dtype: string
      - name: query
        sequence: string
      - name: annotations
        sequence: string
      - name: img_id
        dtype: string
      - name: query_info
        struct:
          - name: notes
            dtype: string
          - name: source_attribution
            sequence: string
          - name: source_license
            dtype: string
          - name: source_url
            dtype: string
      - name: annotations_info
        struct:
          - name: notes
            dtype: string
          - name: source_attribution
            sequence: string
          - name: source_license
            dtype: string
          - name: source_url
            dtype: string
      - name: image_info
        struct:
          - name: notes
            dtype: string
      - name: image_sha256
        dtype: string
    splits:
      - name: train
        num_bytes: 66261559
        num_examples: 46970
      - name: val
        num_bytes: 12498555
        num_examples: 8892
    download_size: 15280439
    dataset_size: 78760114
  - config_name: TextOCR
    features:
      - name: sample_id
        dtype: string
      - name: dataset_name
        dtype: string
      - name: task_name
        dtype: string
      - name: query
        sequence: string
      - name: annotations
        sequence: string
      - name: img_id
        dtype: string
      - name: query_info
        struct:
          - name: notes
            dtype: string
          - name: source_license
            dtype: string
          - name: source_url
            dtype: string
      - name: annotations_info
        struct:
          - name: notes
            dtype: string
          - name: source_license
            dtype: string
          - name: source_url
            dtype: string
      - name: image_info
        struct:
          - name: notes
            dtype: string
      - name: image_sha256
        dtype: string
    splits:
      - name: train
        num_bytes: 123664898
        num_examples: 43484
      - name: val
        num_bytes: 17736515
        num_examples: 6232
    download_size: 35928078
    dataset_size: 141401413
  - config_name: pubtables-1m
    features:
      - name: sample_id
        dtype: string
      - name: dataset_name
        dtype: string
      - name: task_name
        dtype: string
      - name: query
        sequence: string
      - name: annotations
        sequence: string
      - name: img_id
        dtype: string
      - name: query_info
        struct:
          - name: notes
            dtype: string
          - name: source_license
            dtype: string
          - name: source_url
            dtype: string
      - name: annotations_info
        struct:
          - name: notes
            dtype: string
          - name: source_license
            dtype: string
          - name: source_url
            dtype: string
      - name: image_info
        struct:
          - name: notes
            dtype: string
      - name: image_sha256
        dtype: string
    splits:
      - name: train
        num_bytes: 4515375098
        num_examples: 1371294
      - name: val
        num_bytes: 568892238
        num_examples: 172887
      - name: test
        num_bytes: 561231801
        num_examples: 170466
    download_size: 1398777256
    dataset_size: 5645499137

BigDocs-7.5M

Training data for the paper: BigDocs: An Open and Permissively-Licensed Dataset for Training Multimodal Models on Document and Code Tasks

🌐 Homepage | 📖 arXiv

🔔 News

• [2025-04-23]: Initial release of the BigDocs-7.5M data.

Guide on Data Loading

Some parts of BigDocs-7.5M are distributed without their "image" column, and instead have an "img_id" column. The file get_bigdocs_75m.py, part of this repository, provides tooling to substitutes such images back in.

from get_bigdocs_75m import get_bigdocs_75m

cocotext = get_bigdocs_75m("COCOtext", user_local_path=".../train2014")
pubtables1m = get_bigdocs_75m("pubtables-1m", user_local_path=".../PubTables-1M-Detection/images")
textocr = get_bigdocs_75m("TextOCR", user_local_path=".../train")

When specified, user_local_path must point to one of the third-party datasets listed below.

- COCOtext: http://images.cocodataset.org/zips/train2014.zip
- pubtables-1m: https://www.microsoft.com/en-us/research/publication/pubtables-1m
- TextOCR: https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip

See the docstring in get_bigdocs_75m.py for more details.

Licensing

The part of this repository generated by us is Copyright ServiceNow 2024 and licensed under the CC-BY-4.0 license.

Multiple datasets, documents, and tools were involved in the generation of BigDocs-Bench. We document these dependencies on a per-sample basis through the query_info, annotation_info and image_info fields, respectively documenting the query, annotations and image fields of our datasets.