|
--- |
|
license: gpl |
|
--- |
|
|
|
# Introduction |
|
|
|
We have scrapped all the collected works of Mohandas Karamchand Gandhi (aka Mahatma Gandhi) from [here](http://www.gandhiashramsevagram.org/gandhi-literature/collected-works-of-mahatma-gandhi-volume-1-to-98.php). Cleaned the text so that it contains only the writings of Gandhi without footnotes, titles, and other texts. |
|
|
|
|
|
We observed that (after cleaning), Gandhi wrote 755468 sentences. |
|
|
|
|
|
We first fine-tuned gpt-2 for 1 epoch on the English corpus (after cleaning*) of [Ai4Bharat](https://ai4bharat.iitm.ac.in/corpora). |
|
|
|
Since the above dataset contains news regarding Indian subcontinent. We thought that with this fine-tuning, the model will get familiary with India specific terms. |
|
|
|
Then we further fine-tuned this model on sentences written by Gandhi (for 3 epochs). |
|
|
|
[Here](https://colab.research.google.com/drive/1NAUCRoIf3fXu2wGDjiRhIrRe8sB80h81?usp=sharing) is the colab link with a working example. |
|
|
|
*Before cleaning #sents = 54M, after cleaning 42M. We simply took those English sentences which ends with a full-stop. |
|
|
