metadata
license: mit
dataset_info:
features:
- name: id
dtype: int64
- name: English
dtype: string
- name: Uzbek
dtype: string
splits:
- name: train
num_bytes: 450765
num_examples: 1428
download_size: 298131
dataset_size: 450765
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
task_categories:
- translation
- question-answering
- text2text-generation
language:
- uz
- en
pretty_name: Fleurs Uzbek
size_categories:
- 1K<n<10K
This is a clean copy of google/fleurs dataset that has translation of text data in 100s of languages with speech.
The primary goal of this dataset is to use it in fine-tuning phase of an LLM so that it could better translate English to Uzbek.
You should add one more column called 'system instruction' and add some command like below:
"Ushbu matnni ingliz tilidan o'zbek tiliga tarjima qiling."
Prompt: English text
Response: Uzbek text