Tostino

51 ms·

901.
by Tostino 3y ago
It takes work, and breaking your problem down into very simple and clear tasks. It is possible to get decent data transformations from gpt 3.5 though. I generally start with gpt4 when prototyping an idea, and then after it's working c
902.
by Tostino 3y ago
I used: https://github.com/artidoro/qlora but there are quite a few others that likely work better. It was literally my first attempt at doing anything like this, and took the better part of an evening to work through
903.
by Tostino 3y ago
Falcon is also slow for inference compared to LLaMA models of similar size. Speed can turn into a quality all its own when you break up problems into small pieces and use the LLM to iterate.
904.
by Tostino 3y ago
Look at QLoRA. The QLoRA can be attached to all layers, allowing you to alter behavior with much less data than the original LoRA implementation. It seems to "stick" better. I just fine tuned a ~30b parameter model on my 2x 3090s
905.
by Tostino 3y ago
It really depends on what you are using it for. There are plenty of high-value use cases where you can free up resources that are spending time inefficiently due to handling tasks that a domain specific LLM could help with. Maybe that allow
906.
by Tostino 3y ago
I used windows 2000 for many, many years. Pretty sure until 7 came out.
907.
by Tostino 3y ago
Fair enough, totally missed that it was the same account.
908.
by Tostino 3y ago
The GP comment didn't read like a ChatGPT response to me. What made you say that?
909.
by Tostino 3y ago
Military provides a bunch of benefits until it's actually used for its purpose. Then you send a bunch of underage kids to die.
910.
by Tostino 3y ago
That is really disappointing to hear... a whole lot of wasted compute.
911.
by Tostino 3y ago
The bots will break...for a few days until selenium based alternative APIs pop up to interact with the site with. Personally, I'll be purging my 10+ years of content from the site and moving on. I just about only interact with Reddit t
912.
by Tostino 3y ago
All it would take to remedy that, is actually providing good support going forward and a bit of advertising. Not a huge barrier IMO.
913.
by Tostino 3y ago
That depends. You can implement an error checking algorithm with the LLM itself to check for drift in meaning and loss of original context from version to version as you improve your training dataset, as well as measure performance as you t
914.
by Tostino 3y ago
3090 can run 30b parameter models, 2x can run 65b parameter models. 4090 can run the same, very slightly faster for much more money.
915.
by Tostino 3y ago
Thank you for that. Google utterly failed to bring...any relevant results when I was searching for what others were doing for ML rigs. I'll stick with 2x 3090s for now, but maybe I'll go for an 8x monster in a few years when the h
916.
by Tostino 3y ago
Good to know, but I've racked up a decent bill for just my GPT 3.5 use. I can get by with experiments using my ChatGPT Plus subscription, but I really need my own API access to start using it for anything serious.
917.
by Tostino 3y ago
Not GP, but it's still the latter...i've been (im)patiently waiting. From their blog post the other day: With these updates, we’ll be inviting many more people from the waitlist to try GPT-4 over the coming weeks, with the intent
918.
by Tostino 3y ago
So I've been working on a side project to make a Youtube channel I watch have its content be more discoverable through text. I've had great results by scraping the Youtube transcription, and running that for a few passes through
919.
by Tostino 3y ago
What is on offer optimized for running these LLMs?
920.
by Tostino 3y ago
I had the same issue... Turned out it was because I used the flat pack version of intellij idea and it had problems with paths. Running from a plain terminal worked fine.
921.
by Tostino 3y ago
I think curriculum training is going to come back around now that we have found an approach that generally works. That seems like the direction a lot of these papers are going.
922.
by Tostino 3y ago
I don't personally like model censorship, but you need to come up with a better argument than...whatever that is. It is entirely unconvincing, and seems like you misunderstand what free speech is.
923.
by Tostino 3y ago
There is a happy medium... somewhere. After following Postgres development for the better part of a decade, I think it's definitely closer to the python side of things... But man they (python) do make it hard to like using that ecosyst
924.
by Tostino 3y ago
Most of them are merged models, so you don't need the base model. It's stupidly simple to get going.
925.
by Tostino 3y ago
I'm interested in seeing this scaled up to larger parameter size models (30b+ parameters), and the dataset expanded with more high-quality data (scientific papers, more books, more code, etc).
926.
by Tostino 3y ago
I could see a Jira plugin that does this by looking through all people working issues and figuring out how would be best to handle this by looking at prior tasks completed, and notes associated with them, along with workload among the team.
927.
by Tostino 3y ago
Completely agree here. And the fact of the matter is, long term memory is actively being worked on. You can already approximate reasonably with a Vector DB and embeddings, feeding the relevant results into the context. As is increasing the
928.
by Tostino 3y ago
I don't think that's true at all. Think of it like setting up conversation constraints to reduce the potential pitfalls for a model. You can vastly improve the capability of just about any LLM I've used by being clear about w
929.
by Tostino 3y ago
There are a lot of people who are entirely okay with the censorship but think it should be done in a different layer than the main LLM itself, as not to hurt the cognitive performance. Alignment is just fine-tuning... any type of fine tuni
930.
by Tostino 3y ago
Try getting GPT-4 to spit out that much code and have it be coherent and run together.

More ›