Comment Re:Flash is costly? (Score 5, Informative) 37
Creating the training dataset is the *last* step. I have dozens of TB of raw data which I use to create training datasets that are only a few GB in size. Of which I'll have a large number sitting around at any point in time.
Take a translation task. I start with several hundred gigs of raw data. This inflates to a couple terabytes after I preprocess it into indexed matching pair datasets (for example, if you have an article that's published in N different languages, it becomes (N * N-1) language pairs - so, say, UN, World Bank, EU, etc multilingual document sets greatly inflate). I may have a couple different versions of this preprocessed data sitting around at any point in time. But once I have my indexed matching pair datasets, I'll weighted-sample only a relatively small subset of it - stressing higher-quality data over lower quality and trying to ensure a desired mix of languages.
But what I do is nothing compared to what these companies do. They're working with common crawl. It grows at a rate of 200-300 TB per month. But the vast majority of that isn't going to go into their dataset. It's going to be markup. Inapplicable file types. Duplicates. Junk. On and on. You have to whittle it down to the things that are actually relevant. And in your various processing stages you'll have significant duplication. Indeed, even the raw training files... I don't know about them, but I'm used to working with jsons, and that adds overhead on its own. Then during training there's various duplications created for the various processing stages - tokenization, patching with flash attention, and whatnot.
You also use a lot of disk space for your models. It's not just every version of the foundation you train (and your backups thereof) - and remember that enterprise models are hundreds of billions to trillions of FP16 parameters in their raw states - but especially the finetune. You can make a finetune in like a day or so; these can really add up.
Certainly disk space isn't as big of a cost as your GPUs and power. But it is a meaningful cost. As a hobbyist I use a RAID of 6 20TB drives and one of 2 4TB SSDs. But that's peanuts compared to what people working with common crawl and having hundreds of employees each working on their own training projects will be eating up in an enterprise environment.