Hacker Remix

Guide to Fine-Tuning LLMs

154 points by ignoramous 2 days ago | 16 comments

kcorbitt 2 days ago

I saw this when it was making the rounds on X a few days ago. Fair warning: it seems like at least some sections are AI-generated, and there isn't much insight to be gained from reading the actual sections compared to eg. reading the relevant category pages on Huggingface.

danielhanchen 2 days ago

I took a skim through it in the morning - I like the LoRA Learns Less and Forgets Less paper more https://openreview.net/forum?id=aloEru2qCG - it has much more signal in a few pages - also the original QLoRA paper from Dettmers https://arxiv.org/abs/2305.14314 has so many more important morsels.

But all in all, the review is a reasonable "manual" I guess. I would have liked maybe more instructive comprehensive practical examples, and maybe more mention of other OSS packages for finetuning :))

YetAnotherNick 2 days ago

Not only the it seems to be AI generated, it seems these guys don't even know about best practices or even what works. e.g. It contains archaic comparison of optimizers and its pros and cons, but for LLMs no optimizer other than Adam and new ones like Lion works.

abc-1 2 days ago

Is there a paper on this? Why do no other optimizers give good results? Adam requires insane amounts of memory so alternatives would be welcome.

daghamm 2 days ago

I would not say that, as long as it is a good summary there is a value in having everything in the same document.

Obviously they should have stated that this is partially generated, but at least they are dog fooding it :)

worstspotgain 2 days ago

Glancing at the authors' names, it's possible that none of them are native English speakers. Any chance that the sections you're referring to were just AI-polished rather than AI-generated?

qeternity 2 days ago

No, this paper was edited yesterday. The original (you can verify on arxiv) contained this incredible section: "6.10 Optimised Routing and Pruning Operations (ORPO)"

The actual ORPO paper is "Odds Ratio Preference Optimisation" and it has nothing to do with pruning. This goes way beyond native language preference.

espadrine 2 days ago

It takes no time at all to find other major mistakes. For instance, the Mixtral diagram § 6.6.1 shows a single router that selects separate 32-layer transformers. Instead, Mixtral has one router per layer (inside of each block), and it doesn’t select a transformer block: it selects a feedforward.

cubefox 2 days ago

Wow, so significant parts of the paper could still be LLM confabulation.

anothername12 2 days ago

Well, it sucks that we’re at the “best practices” phase already

p1esk 2 days ago

It sucks that we’re still at “best practices” phase. We’ve been in this phase for the last three decades [1], and I really hope we enter “good theory” phase soon.

[1] https://cseweb.ucsd.edu/classes/wi08/cse253/Handouts/lecun-9...

kleiba 2 days ago

Why is that?

make3 2 days ago

there's likely still an infinite amount of things to figure out, transformers haven't been out for 10 years yet

raymond_goo 2 days ago

Ctrl-F: Unsloth --> no results == bad paper

youoy 2 days ago

But you can find "delve"

aubanel 2 days ago

FYI, some sections are LLM-generated bullshit: https://x.com/sam_paech/status/1848332471953448972

The example shown in the tweet has been edited out of the paper since, but there must be others. High noise, low signal content.