True, but 4M posts is heavy (even when bypassing the API itself and using unauthenticated requests), so was curious.

Right! I used the official Reddit API. I created an APP, got the credentials for the API. Then used the Python library PRAW to consume the API. https://praw.readthedocs.io/en/latest/

It took me 36 hours to collect the 4M posts. Reddit API returns results in batches of 100 results, and then sleeps for 2 seconds.

You can find some more details on how it was built here https://blog.valohai.com/machine-learning-pipeline-classifyi...

I can publish on Github the repository that runs two commands to collect the data if your are interested.