batch / README.md
zeio's picture
feat(example): added link to the dataset usage examples in readme
f97338e verified
metadata
language:
  - ru
  - en
license: apache-2.0
tags:
  - social-networks
  - not-for-all-audiences
annotation_creators:
  - crowdsourced
language_creators:
  - crowdsourced
pretty_name: batch
size_categories:
  - 100K<n<1M
task_categories:
  - text-generation
  - text-classification
  - question-answering
dataset_info:
  - config_name: written
    features:
      - name: title
        dtype: string
      - name: topics
        sequence:
          - name: posts
            sequence:
              - name: text
                dtype: string
  - config_name: spoken
    features:
      - name: title
        dtype: string
      - name: speech
        dtype: audio
      - name: topics
        sequence:
          - name: posts
            sequence:
              - name: text
                dtype: string

Dataset card for batch

Table of contents

Dataset description

Dataset summary

This dataset contains threads parsed from the /b/ board of 2ch archive. See dataset viewer at the derivative repo. Examples of the dataset reading and usage are provided in this colab notebook.

Dataset structure

The dataset is represented in three formats - compressed, uncompressed and spoken:

  1. uncompressed representation is the default and simplest one - in this form the content of dataset is organised inside txt files which are grouped into clusters inside threads folder. The grouping is done due to git's constraints, namely, because it's not possible to have more than 10000 files in a single directory. That's why each cluster contains 10000 items (except the last one, which could contain fewer elements). Each cluster name has the format ${START_PAGE}-${END_PAGE}, where ${START_PAGE} is the index of the first page in the archive from which posts have been put into the cluster, and ${END_PAGE} is the last such paget respectively;
  2. compressed representation is slightly more sophisticated than the uncompressed one - in consists of a set of tar.xz files which are nothing more than the compressed clusters of txt files described above. This representation corresponds to the threads-compressed folder;
  3. spoken representation consists of mp3 files with speech generated for some threads using an alternating speaker voice pattern meaning that the 1st post is said by the first speaker, the 2nd post is said by the second speaker, the 3rd post is said by the first speaker, the 4th post is said by the second speaker and so on. The speech is generated automatically using a TTS engine. The mp3 files are located in the threads-spoken-compressed and are grouped using tar.xz archives in the same way as txt files in the compressed dataset representation.

Concerning particular txt files under threads/\*/ folder, each item here corresponds to one thread and is organised as follows:

  1. Each non-empty line corresponds to a single post from a user;
  2. If a non-empty line follows another non-empty line, then it should be treated as a comment to one of the posts above it, a response to a request above or as an answer to a question;
  3. If a non-empty line follows an empty line, it should be treated as a beginning of a discussion or a topic.

Therefore, the dataset consists of threads, which can be separated into topics, which, in turn, consist of posts. Posts are the lowermost units in the dataset and are not divided further - they should be interpreted as a plain text.

Dataset instance

The following code snippet contains text for the thread 0000-0019/119540414:

Всем привет. Нужна помощь богов фотошопа, на картинке надо изменить дату на 09/03/2016 и значения тесто на 86.500++
черес код елемента ебаш
Опять ты, сука ебаная? Хули тебе опять надо?

СПАСИБО
Размер шрифта не совпадает, але.

This thread consists of two topics, the first one of which includes 3 posts, and the second - 2 posts.

Therefore, this dataset entry can be represented in json in the following format:

{
  "title": "Всем привет. Нужна помощь богов фотошопа, на картинке надо изменить дату на 09/03/2016 и значения тесто на 86.500++",
  "topics": [
    {
      "posts": [
        {
          "text": "Всем привет. Нужна помощь богов фотошопа, на картинке надо изменить дату на 09/03/2016 и значения тесто на 86.500++"
        },
        {
          "text": "черес код елемента ебаш"
        },
        {
          "text": "Опять ты, сука ебаная? Хули тебе опять надо?"
        }
      ]
    },
    {
      "posts": [
        {
          "text": "СПАСИБО"
        },
        {
          "text": "Размер шрифта не совпадает, але."
        }
      ]
    }
  ]
}

Dataset fields

In written configuration the dataset is represented as a list of Thread objects, each Thread has a single property topics, which contains a list of Topic objects. Each Topic object has a single property posts, which points to the list of Post objects, making up the Topic. Each Post object contains a single property text which contains text representation of the post (essentially text is html code without tags and explicit links to other posts; there may still be implicit links to other posts in a form of quotes, prefixed with > symbol). As an additional field, each instance has a property title which is equivalent to the thread's main post content.
In spoken configuration the structure is basically the same, but some Thread objects have and additional property speech with a spoken representation of the thread.