Conversation

Boris Dayma

@borisdayma

May 27, 2020

I've been working on #huggingtweets, a fun project to generate tweets based on your favorite twitter account using

@huggingface

. You can fine-tune a neural network and log the predictions automatically into

@weights_biases

. The demo runs in 2-3mn: colab.research.google.com/github/borisda

150

501

Boris Dayma

@borisdayma

May 27, 2020

The predictions are pretty good despite using a “small” gpt-2 network and having very small datasets (typically 100-200 kB). People with longer tweets such as

@jackclarkSF

@fchollet

typically generate better predictions.

Boris Dayma

@borisdayma

May 27, 2020

We can definitely recognize the personality of the

team.

It’s interesting to see that the model learns hashtags, @ mentions, pic links, or even

emojis.

What amazes me is that everybody has similar loss functions except for

@karpathy

who appears to be much more unpredictable to the model. I actually ran it multiple times to confirm these results!

Boris Dayma

@borisdayma

I had fun running it on some people I like to follow. Here are a few predictions on

and

8:04 PM · May 27, 2020Twitter Web App

Replying to

And here are other predictions on

. I really feel like we can recognize people’s personalities.

Boris Dayma

@borisdayma

May 27, 2020

Later I’ll pre-train the model on a large tweet dataset for the model to learn the “tweet” language. Then set up a pipeline for continuous learning of the pre-trained model with data/model versioning. For more details, check my report and try the demo! bit.ly/wb-huggingtwee

Boris Dayma

@borisdayma

May 27, 2020

Try it with your friends tweets!

colab.research.google.com

huggingtweets-demo.ipynb

Run, share, and edit Python notebooks