ro_sft_finepdfs / README.md
mihaimasala's picture
Update README.md
40d4977 verified
|
Raw
History Blame Contribute Delete
2.02 kB
metadata
license: odc-by
language:
  - ro
dataset_info:
  features:
    - name: id
      dtype: string
    - name: images
      sequence: image
    - name: text
      dtype: string
    - name: messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
  splits:
    - name: train
      num_bytes: 282401531337.178
      num_examples: 379198
  download_size: 303410496330
  dataset_size: 282401531337.178
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Description

FinePDFs is the largest publicly available corpus sourced exclusively from PDFs, containing about 3 trillion tokens across 475 million documents in 1733 languages.

Here we provide the Romanian split of FinePDFs training set, prepared for OCR: pairs of images (pages) and extracted text. This dataset is part of the instruction finetune protocol for Romanian VLMs proposed in "Înțelegi românește?" A Recipe for Romanian Vision-Language Models (Masala et al., 2026).

Citation

@misc{kydlicek2025finepdfs,
      title={FinePDFs}, 
      author={Hynek Kydl{\'\i}{\v{c}}ek and Guilherme Penedo and Leandro von Werra},
      year={2025},
      publisher = {Hugging Face},
      journal = {Hugging Face repository},
      howpublished = {\url{https://huggingface.co/datasets/HuggingFaceFW/finepdfs}}
}
@misc{masala2026intelegi,
      title={``\^{I}n\c{t}elegi Rom\^{a}ne\c{s}te?'' A Recipe for Romanian Vision-Language Models},
      author={Mihai Masala and Marius Leordeanu and Mihai Dascalu and Traian Rebedea},
      year={2026},
      eprint={2605.31401},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.31401},
}