The Common Pile V0.1: an 8tb Dataset of Public Domain and Openly Licensed Text

Personally I don't believe in model collapse. Has anyone demonstrated it occurring in the wild, outside of the tiny set of papers that deliberately caused it to happen?

I think model collapse gets talked about so much because it is irresistible schadenfreude. The idea of models eating their own tails in a way that leads to their inevitable demise is captivating to a lot of people, especially AI skeptics.

"Model collapse" isn't real. It's a laboratory failure mode that doesn't happen in real world environments.

It's popular because some people latched onto the idea - desperately wanting something to stop the AI tech from advancing. It, quite obviously, doesn't stop the AI tech from advancing.

Now, you can write an entire research paper on why model collapse happens or fails to happen. But a simple way to think of it is: looping AI onto itself multiple times amplifies that AI's own deficiencies, distortions and idiosyncrasies - until, after enough iterations, they come to completely dominate its outputs.

This doesn't apply at all to training an LLM on Whisper outputs that are, in turn, based on human-generated videos. The LLM will inherit some Whisper quirks, but most of the data in Whisper outputs comes from the videos themselves.

No, I don’t think it applies here. The semantics and speech patterns were generated by a human, Whisper just transcribed them.

There is some risk that Whisper transcribed inaccurately, but that’s less model collapse and more “the dataset is bad”.

Phi models are notorious for using mostly synthetic data

The reality of working with humongous datasets is they're always bootstrapped like this, in multiple steps. In LLMs in particular, the entire post-training step is always done on synthetic data. There are ways to avoid failure modes typical for that (like model collapse), you need much less real data to keep the model in check than you probably think.

Its accuracy can be and is quantified.