Hacker News Reader

Comment by NitpickLawyer | Hacker News Reader

Weights are not binary. A model is created at init time, with random values. After that, it is being modified using data. The key point is that the labs modify the models "as weights". That means that weights are the intended / preferred way of modifying a model. Which, coincidentally, matches the definition of source in Apache 2.0. There is no "higher level" place where editing takes place. It all happens in weight space. Through the license you get the same rights as the lab that created it: view, inspect, run, modify, re-release. That's it. That's the only thing a license can grant you.

The rest is semantics, misunderstandings, and FUD. A model released under an open source license is open source. Training data is lab knowhow / IP. Which, historically, has never been required for any open source release.

View on HN

Well, you can't add or alter data in pre-training from just the weights. Which, as I understand it, means you can't fundamentally increase core knowledge or cognitive ability, only what the model likes to do with those. You can only post-train, and you're subject as a result to catastrophic forgetting.

To explain simply as far as I can tell (would love to be corrected) the large number of pre-training tokens only works because the documents are randomly ordered.

So if you e.g. took a foundation model with open weights, then tried post-training it all the new data since its cut-off period, it would then end up over-trained on that new data, and forget older things.

As I live next to EPFL, I'll give you example from them: their Meditron-70B model is adapted to the medical domain from Llama-2-70B through continued pretraining. They took weights of Llama-2-70B and continued training on PubMed, medical guidelines and general data.

Weights aren't just executable artifact that's consumed by users. Third parties actually use released parameter state as the editable starting point for further training and produce new foundation models from it.

Nice example - although it seems it ended up specialised in medical texts, so did indeed ("catastrophically") forget other knowledge?

No, it didn't.

It lost 69.2% -> 67.8% on MMLU while improving medical performance. If you're trying to argue that loss of ~1.4 points is "catastrophic forgetting" (it's not) then look at later work, ie. Me-LLaMA that clearly demonstrates continued pretraining that improved both general MMLU and medical performance.

Not sure why you're fixating on catastrophic forgetting. How do you think model training works? Model training is just a sequence of checkpoints: pretraining produces it, training resumes from last, continued pretraining starts from last, supervised fine tuning starts from last, RL/post-training starts from last - it's just a sequence of checkpoints.

There isn't some fundamental distinction where original author continuing training from checkpoint X is training but a third party downloading checkpoint X and continuing training from it suddenly isn't.

ie. checkpoint doesn't somehow become a different kind of artifact when it's published.

Reply on news.ycombinator.com