reporting failures: I finetuned gemini and it went badly. over the weekend I had the urge to try to get a virtual lesswronger who could yell at me for bad takes endlessly. I collected a curated (but not really well curated) dataset of my favorite contrarians replying to threads. I trained gemini-2.5-flash-lite on this, intending it to be a cheap test pass to see if it was worth pursuing further. it was not cheap, I thought it'd cost $10 - counting tokens in the dataset, which used fairly long contexts because it included the posts and had the favorite user as the model's response - but it instead cost $100, presumably counting total tokens inferred! and of course, being flash-lite, the output is pretty disappointing. idk, might be interesting as a source of terrible takes to be annoyed at and get inspired by, at best. instead of being an online contrarian it's an endless source of someone being wrong on the internet. sigh. quite annoyed at the lost time and money.
yeah, preprocessing with an AI's help will be my next go at it. likely a bit more than just curation, also various data cleaning and reformatting. I might try very low epoch count, very high learning rate, minimum lora size, to see if that can get me anywhere. Also going to try local tuning.
my suspicion is that in-context learning is stronger, but I currently expect to prefer fine-tuning if it works at all, because it can force lower dimensional update (only in the low millions rather than gbs of kv cache) and I'd expect that to generalize better. probably there are like 30 papers on this, actually, brb
edit: hmmm...
[...]
hmmmmmmmmn
[...]
hmmm
wat that's not alignment
ok looks tasty, read me further?
pac bounds for icl?? what's the gotcha. but in the intro they say
[...]
okay so yeah my intuition is updated a bit from this paper browsing session but the main thing I'm getting so far is "lol idk try it and see what sticks, both of these things are unreliable and sometimes one is better than the other"
[...]
it can do what now?
[...]
okay, so icl is good and then bad, whereas sft is bad and then good? thanks for clearing that up for me. looking at their charts... weird, they're clustering multiple-choice questions by answer letter? their charts of clustering and intrinsic dimension sure would be cool if I was reading closely enough to be sure they meant anything
https://arxiv.org/abs/2502.04580 "Our findings reveal a striking dichotomy: while ICL initially matches the efficiency of a Bayes optimal estimator, its efficiency significantly deteriorates in long context" sure yes makes sense, this feels overwrought but probably true
I think the thing I want out of SFT is probably just that it works to change implicit priors in the model without the model having verbal reference to the things changed, so it can bypass a bunch of verbal tics and go directly for imitation
Yeah "try it and see" is the gold standard. I do know that for stuff which boils down to "monitor for patterns in text data that is too large to plausibly be examined by a team of humans we could afford to hire" I've been favoring the approach of
Grab 100 random data points, run the most obvious possible prompt on them to get reasoning + label(s) + confidence
Spot check the high confidence ones to make sure you're not getting confident BS out of the model (you can alternatively start by writing two very different prompts for the same labeling task and see... (read more)