license: gpl-3.0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
dataset_info:
features:
- name: src
dtype: string
- name: dst
dtype: string
splits:
- name: train
num_bytes: 41095474
num_examples: 300000
- name: validation
num_bytes: 255868
num_examples: 500
- name: test
num_bytes: 365602
num_examples: 1000
download_size: 22810511
dataset_size: 41716944
task_categories:
- translation
language:
- en
tags:
- Yandex
- Young&&Yandex
size_categories:
- 100K<n<1M
Dataset Description
Dataset Summary
This dataset was used as homework #3 in the Yandex Machine Learning 2.0 training in November 2024.
This dataset is a simulation of automatically collected internet data for a low-resource language.
In 2084, humanity made first contact with an extraterrestrial civilization living on the planet Zeta in the constellation Andromeda. The aliens, called Zetans, have highly developed technology and are eager to share knowledge with Earthlings. To successfully establish contact and develop mutually beneficial relationships, it is extremely important to establish effective communication.
The Zetans provided humanity with an extensive text library in their language, including both original works and translations of known Earth texts. In turn, they were given access to Earth's libraries. However, machine translation algorithms have not yet coped well with the unusual structure of the Zetan language, which makes the translation inaccurate and incomplete. Need to train a model for translating from Zetan to English.
Supported Tasks and Leaderboards
Translation.
Languages
The dataset contains two languages, a real one, English, and a fictional one, Zetan.
Dataset Structure
In the original dataset consists of three files in JSON Lines format. Current dataset structure:
DatasetDict({
train: Dataset({
features: ['src', 'dst'],
num_rows: 300000
})
validation: Dataset({
features: ['src', 'dst'],
num_rows: 500
})
test: Dataset({
features: ['src', 'dst'],
num_rows: 1000
})
})
Data Splits
There are three splits in the dataset:
train- 300,000 translation examples. The data is specially noisy. There are approximately 235,000 qualitative examples.validation- 500 translation examples, are used to validate the model during the fine-tuning process.test- 1000 translation examples, used to submit the result to a special testing system - Yandex Contest. There is no translation from Zetan to English in this part of the dataset!
Data Fields
src - contains a phrase in the Zetan language.dst - contains a translation from Zetan to English.
Data Instances:
{
"src":"◲▦◠▦◬▦■ ◉◗▢◕◗ ◍◗▱◎ ▽◠▽▪▦◠ ◕▴◉◗▦▼▴ ◀◗◓◉◧▨ ◎▴◞◠▸ ◠▱◈▪▨ ◚◪ ◀◨ ◎◪◞◠▸▱◠◓◬▦ ◀◠▢▪▱◠◓▪ ▻◪▨ ◈◂◞▫◉◠ ◈▴◐◫▱◈◗▵",
"dst":"Believe me, we had many messages and some of these messages that they were all over the world."
}