SunbirdAI / parallel-text-EDA

Exploratory data analysis and data preparation the ug-language-parallel-text dataset

Language parallel text exploratory data analysis and data preparation

Exploratory data analysis and data preparation scripts for model training, done on the ug-language-parallel-text dataset

Step by step outline of our process of data preparation and model training:

• Concatenate any new chunks of data into the publicly available json dataset by running the Data EDA and cleaning.ipynb notebook
• Create the model training dataset files and put them in the right folder structure by running the Model Training Data Prep.ipynb and Multilingual Data Prep.ipynb notebooks, in that order
• Upload the dataset to the sunbird-translate bucket on AWS S3 (as a versioned dataset, for example v4-dataset.zip) and make sure the resource is public
• Update the dataset URL in the SunbirdAI/datasets repository with the new dataset link (as shown in the next section of this README). The code in this file picks up the models from S3 bucket whose link we added as the dataset URL
• Run the SunbirdAI language model training notebook on AWS SageMaker. The training notebook refers to SunbirdAI/datasets for the datasets to be used in training
• Save the checkpoints and upload them to the sunbird-translate AWS S3 bucket, in the models folder
• Upload the models to Hugging Face
  

How to add the dataset link in SunbirdAI/datasets

Find the _URL constant in the datasets/sunbird/sunbird.py file on the init-sunbird-dataset branch of the SunbirdAI/datasets repository.
The image below shows an example of this: