Scripts for training data generation and processing of generated tweets by
@fortunes_teller, a human-curated
tweetbot made with GPT-2 trained on various fortunes packages, sourced
from github (full links in data_generation.R), the R fortunes package,
and by running fortune fortunes > fortunes.txt until I was pretty confident
I had all the fortunes included in fortune-mod 1:1.99.1-7build1.
data_generation.R walks through generating the training data; model training
and tweet generation happened in Max Woolf's
fantastic Google colab notebook.
Post-processing of generated tweets can be found in cleaning_generated_tweets.R,
and a simple rtweet script to post tweets in hourly_post.R. I have the
post script called by a bash loop managed by systemd that posts a tweet then
sleeps an hour.
This bot can be scarily on-the-nose, so this repo also includes a small
shiny app in did_i_do_that to check bot outputs against the training
dataset to make sure anything I'm choosing as a tweet was actually created
and not just stolen. You can see a demo of the app live here.