Finetune Huggingface t5 for text2text generation

About T5:

• T5 , or Text-to-Text Transfer Transformer developed by Google, is a Transformer based architecture that uses a text-to-text approach.

Goal:

We have multiple small/simple notes in the following fashion.

Input:

• James Elliot leaving the company was a loss.
• The firm drew comfort from Shrenick Shah's experience and involvement in the strategy since its inception in 2012.

Target output:

• Elliot's departure was a loss, but we draw comfort from Shah's experience and involvement in the strategy since its 2012 inception.

Predicted output from t5:

• Elliot's departure from the firm was a loss, and the firm drew comfort from Shah's experience and involvement in the strategy since its inception in 2012.

Implementation details:

• We have used a T5-based Huggingface model to finetune on our dataset
• Dataset split into train and test
• Ran for 12 epochs
• Paraphrased using another Paraphraser