README.md

metadata

language:
  - mr
license: cc
size_categories:
  - n<1K
source_datasets:
  - extended|ai4bharat/IndicXParaphrase
task_categories:
  - text-generation
pretty_name: Aya Paraphrase Marathi
dataset_info:
  features:
    - name: inputs
      dtype: string
    - name: targets
      dtype: string
    - name: template_lang
      dtype: string
    - name: template_id
      dtype: int64
  splits:
    - name: train
      num_bytes: 683937
      num_examples: 1001
  download_size: 245473
  dataset_size: 683937
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Description

This dataset is derived from the already existing dataset made by AI4Bharat. We have used the IndicXParaphrase dataset of AI4Bharat to create this instruction style dataset. We have used the malayalam split of the above mentioned dataset to create this one. This was created as part of Aya Open Science Initiative from Cohere For AI.

IndicXParaphrase is multilingual, and n-way parallel dataset for paraphrase detection in 10 Indic languages. The original dataset(IndicXParaphrase) was made available under the cc-0 license.

Template

The following templates(Marathi) where used for converting the original dataset:

#Template 1
prompt:
खालील वाक्य दुसरे-भिन्न शब्द वापरून लिहा:  "{original_sentence}"

completion:
{paraphrased_sentence}

#Template 2
prompt:
खालील वाक्य वेगळ्या प्रकारे पुन्हा लिहा:  "{original_sentence}"

completion:
{paraphrased_sentence}

#Template 3
prompt:
खालील वाक्य दुसरे शब्द वापरून रूपांतरित-अनुवादित करा: "{original_sentence}"

completion:
{paraphrased_sentence}

Acknowledgement

Thank you, Yogesh Haribhau Kulkarni for helping with the preparation of this dataset by providing the Marathi translation of the above mentioned English prompts.