Hacker News Reader

Comment by llm_trw | Hacker News Reader

We don't know since no one is releasing their data.

Calling these models open source is like calling a binary open source because you can download it.

Which in this day and age isn't far from where were at.

A big distinction is that you can built on top (fine-tune) thus released models as well as if they released the pre-training data.

You can also build on top of binaries if you use gotos and machine code.

This seems intentionally obtuse. What you say is true, but it is very obvious that this is much more of a pain than if you had the source code. On the other hand, fine tuning is just as easy, regardless of whether you have the original training data.

If you don't know the original training data statistical distribution then catastrophic forgetting is guaranteed with any extra training.

I was going to ask for a reference to support this (although please provide one if handy), but the search term “catastrophic forgetting” is a great entry into the literature. Thanks.

One could also disassemble an executable and build on top of it. Not for the faint of heart and probably illegal, but possible unless it was deliberately obfuscated. Compared to that, it is impossible with state-of-the-art methods to systematically extract the training data from an LLM model. Fragments yes, but not all of it.

You can do better - generate synthetic data covering all topics. And to make it less prone to hallucination, use RAG or web search for reference material. The Phi-1.5 model was trained on 300B of synthetic tokens generated with chatGPT and it showed a 5x bump in efficiency, punching well above its line.

Synthetic data can be more diverse if you sample carefully with seeded concepts, and it can be more complex than average web text. You can even diff against a garden variety Mistral or LLaMA and only collect knowledge and skills they don't already have. I call this approach "Machine Study", where AI makes its own training data by studying its corpus and learning from other models.

Or shell scripts

You can fine tune without the pre training data too.

Mistral models are one example, they never released pre training data and there are many fine tunes.

We are in agreement -- that's exactly what I am saying :)

We should just call it open weight models at this point.

FWIW the Grok repo uses the term "open weights".

Exactly. Training data is key for any trained AI system. And I strongly suspect that every single company active in these modern AI systems is still struggling with how to tune their training data. It's easy to get something out of it, but it's hard to control what you get out of it.

> We don't know since no one is releasing their data.

Is anyone else just assuming at this point that virtually everyone is using the pirated materials in The Pile like Books3?

I think it's really, really clear that the majority of the data used to train all of these things was used without permission.

The model is open weight, which is less useful than open source, but more useful than fully propriety (akin to the executable binaries you compare to)

How about "weights available" as similar to the "source available" moniker?

weights available or model available, but yes.

Their data is the twitter corpus which is public. Or do you want a dump of their database for free too?

Twitter tweet data in itself is both highly idiosyncratic and short by design, which alone is not conductive towards training a LLM.

Saying "It's just the twitter public corpus." is like saying "Here's the Linux Kernel, makefiles not included."

Or even "here's the Linux Kernel makefiles, no sources included, enjoy".

Reply on news.ycombinator.com