README.md

---
license: mit
language:
- ru
- en
- fr
size_categories:
- 10K<n<100K
---

# Paragraph dataset

Dataset for a lines classifier that predicts the beginning of a new paragraph (see [paragraph classifier](https://huggingface.co/dedoc/paragraph_classifier)).

The repository structure:
```
└─paragraph_dataset
  ├─labeling
    ├─paragraph_classifier_834997.zip
    └─results
      ├─law_classifier_000000_AKs.json
      ├─ ...
      └─law_classifier_000049_d9X.json
  ├─.gitattributes
  ├─README.md
  └─paragraph_dataset.zip
```

`labeling` directory contains tasks for labeling (`paragraph_classifier_834997.zip`) and labeled lines (`results` directory).
`paragraph_dataset.zip` contains files with the resilting dataset.

Tasks for labeling are formed using [dedoc labeling API](https://github.com/ispras/dedoc/tree/master/labeling).
See [dedoc documentation](https://dedoc.readthedocs.io/en/latest/tutorials/add_new_structure_type/dataset_creation.html) for more details.

The structure of the `paragraph_dataset.zip` archive after unzipping goes as follows:
```
└─paragraph_dataset
  └─labeled.json
  ├─parameters.json
  ├─task_manager.log
  └─original_documents
    ├─1709714642_887.png
    ├─ ...
    └─1709714823_806.pdf
```

* `labeled.json` contains lines with their labels. For each line, line's text, metadata and formatting are extracted by [dedoc](https://github.com/ispras/dedoc).
* `parameters.json` contains `dedoc` parameters using which lines were extracted.
* `task_manager.log` contains logs from labeling process.
* `original_documents` folder contains documets from which lines were extracted.