Hacker News Reader

Well that was kind of my point: you need to manually figure out what’s clean and what isn’t.

But it's easy to do that for a small subset for finetuning compared to cleaning up the entire dataset.