518 comments

[ 2.2 ms ] story [ 173 ms ] thread
No way, I thought this was months away still. So excited
It's released but still cooking in the oven as we speak.
Completed training is. These are early "Alpha" checkpoints.
Are they releasing the weights for download? The links to HuggingFace in the readme are giving me 404. This dataset they built on-top of "The Pile" sounds interesting - looking forward to evaluating their claim that 3-7 billion param models can perform on par with 175 billion param GPT-3
Did they claim this? I didn’t see that claim made in the above post.
"The richness of this dataset gives StableLM surprisingly high performance in conversational and coding tasks, despite its small size of 3 to 7 billion parameters (by comparison, GPT-3 has 175 billion parameters)."

So they did not explicitly say it is comparable, but implicitly compared the two. I'm curious to evaluate what "surprisingly high performance" means exactly.

I'm curious if this will give better results than llama 7B? Llama 7B felt like a toy that, while cool to be able to run locally, did not feel useful in any way when contrasted to the state of GPT. Here's hoping for better and/or release of larger parameter models with low performance requirements soon :)

EDIT: my first question times out when ran online, seems like huggingface is getting hugged to death.

They have on their team people from Anthropic which have stuff like Claude Instant which is likely running a very light model, ie the tricks from Anthropic have likely been incorporated into the model they’re presenting here
Even if it doesn't initially, the fact that it's being released so permissively is massive - stable diffusion was made far more powerful by being hackable at all levels and I can't imagine we won't see the same here.

I imagine things like control nets that restrict output to parsable types, LoRa style adaptations that allow mixable "attitudes", that sort of thing.

Very different underlying architecture from diffusers, ofc. But the action of open source is the same - a million monkeys with a million xterms and so forth.

I'm really hoping for the ability to load in different sets of trained material as embeddings/textual inversions like in Stable Diffusion. Imagine scanning in some of your favorite philosophy and design books and throwing them with small weighting as a little flavor to your answer. The crossovers between LLM and Stable Diffusion type models (like Loras) is such a fascinating space to explore.
This already exists for LLaMA, I'm sure it'll be implemented for StableLM as well.

https://github.com/lxe/simple-llm-finetuner

good looks on the link! I was experimenting with script writing the other day and thought "gee, I really wish I could finetune on Beckett plays specifically".

I don't know if anyone else has experienced this same tipping point, but when I used to have ideas, I would look them up and discover that implementing them was probably out of scope. These days, I think "wouldn't it be cool..." and immediately stumble on a way to make it happen, by accident.

This has been around for GPT models for a while in the form of "soft prompts", which are rather approximate to textual inversion in the Stable Diffusion space.
If you're comparing to ChatGPT performance then Vicuna 13B would be a best comparison point for something Llama-based.
Vicuna 13B performance is an order of magnitude below ChatGPT for all but gimmicky conversational stuff. Try giving both somewhat large, task-based prompts with steps and see what happens.
> Vicuna 13B performance is an order of magnitude below ChatGPT for all but gimmicky conversational stuff.

Until you connect it to external resources, I tend to think of anything you do with “brain-in-a-jar” isolated ChatGPT as gimmicky conversational stuff.

ChatGPT is still going to be way more capable when you use it's API to connect to external resources.
Maybe I should have phrased that better! I didn't mean that Vicuna was comparable to ChatGPT, just that it's the best Llama-based comparison you can make (since it's at least been conversationally trained).
Isn't ChatGPT a 165B parameter model?
No. OpenAI haven't disclosed parameter count of GPT-3.5 or GPT-4, which are models used by ChatGPT. You may be thinking of GPT-3, which is indeed a 175B parameter model.
Ah, interesting. Thought GPT-3.5 had the same structure as GPT-3, for some reason. GPT-4 would obviously be different.
GPT-3.5 is likely a finetuned Curie 13B using output from the full size GPT-3 175B.
StabilityAI is the real OpenAI. Thanks for this.
Well said. Openai is a promise unkept. Thanks StabilityAI for existing.
Unfortunately, due to the law of names, StabilityAI will in the future hit the same issue as OpenAI and do a 180, unleashing very unstable AI to the world.
I'd rather have that than closed AI models controlled by afew corporations
then they both have ironic names now
More like Stability will turn out to be an unstable company. Last we heard they were struggling to raise more funding and might lose their CEO due to unclear business models:

https://www.semafor.com/article/04/07/2023/stability-ai-is-o...

The company can cease operations tomorrow, but the model they open sourced (and all of its derivatives built by the community) will continue to exist. If OpenAI disappears then all of the work they have done goes with it.
when has opensource ever spearheaded independent innovation? they usually follow along.

Fred Wilson once did a take on all trends in SV. First some firm comes out with a product that changes the landscape and makes a massive profit. Then some little firm comes along and does the same for a cheaper price. Then some ambitious group out of college comes out with an open-source version of the same.

Open source has never been a trailblazer of innovation. Open "research" was the original mantra for open ai. And an entrepreneur in residence put together a great product. If they were any more open, it would not make sense.

> Open source has never been a trailblazer of innovation.

Except for, you know, all the major programming languages and Linux, which make all that innovation possible in the first place. Also, everything OpenAI is doing is based on open source stuff from Google and others, so…

True. Should have added open research has given lots of puzzle pieces for innovation.

And open source products has led to many individual contributions.

But again it's never been a trailblazer for innovation.

This could not be more untrue.

The world is littered with businesses that operate as commercial wrappers around open source technology. Ever heard of GitHub? What about MacOS? AWS?

Ahh. Looks like I misspoke.

The first line should have been "Paradigm shifting innovations have never started as open source."

Yes, open source has helped many people innovate.

Not fully. The instruction tuned model is CC-BY-NC-SA
That's a limitation of the dataset used for that particular tuned model. Probably not a great choice on their part given that people aren't reading past the headline, but the actual base model is not restricted.
Its CC-BY-NC-SA because of the upstream sources used for instruction training. There’s open resources being developed for that that I’ve seen, but probably nothing ready.
This is amazing. They even let the developers use it for commercial purposes;

“Developers can freely inspect, use, and adapt our StableLM base models for commercial or research purposes, subject to the terms of the CC BY-SA-4.0 license.“

You can use this link to interact with the 7B model;

https://huggingface.co/spaces/stabilityai/stablelm-tuned-alp...

Is there a way to check the queue for this ?

I sent it one small text (actually a task) five minutes ago. Its still loading.

same. Probably hugged to death.
It just finished building.
Just tested it. I tried to get it to write a conclusion after giving it my report and while is was correct, it was kind of silly, a far cry from chatgpt. But again, this is the 7B variant and its open source.
This is the Alpha of the 7B variant only partially through its initial training set (800B tokens out of 1.5 Trillion)
(comment deleted)
Finally, the true race to zero is now on.

Stability and others are already at the finish line in the race to zero. O̶p̶e̶n̶AI.com cannot get any cheaper and eventually will have to increase their prices.

There is no escape and as I said before, open source AI models will eventually swamp and survive longer and intergrate faster than even O̶p̶e̶n̶AI.com's ChatGPT.

Expect open source models to get smaller and even better such that it will fit in an iPhone, on device.

Stay tuned.

(comment deleted)
(comment deleted)
Why are you referring to OpenAI as ai.com?
I think he's just emphasizing that OpenAI is in fact not open, thusly it's crossed out.
This is absolutely huge. LLaMA weights leaking was a big deal, but you couldn’t actually use them without attracting Meta’s ire. Would love to see some benchmarking vs. LLaMA and GPT.
What is IRE?
anger. probably a lawsuit in this context
There were no lawsuits around LLaMA. They used DCMA requests to take down some GitHub and HuggingFace releases but the majority of releases (Alpacas and other animals) was left alone. It was enough to prevent commercial use, though. Edit: a lawsuit would be quite interesting, to clear up many things around how copyright works for LLMs.
They're going to court with the guy who made llamaDL actually.
Ire is the result of irritation.
Are you saying we shouldn’t scratch Meta’s itch?
It's a department at meta: Intellectual Rights Enforcement. /s

"Ire" is a synonym for "anger" or "wrath"

That's actually what I thought it might be. Thanks lol.
Your initialization hinted it. I must say, if Meta had a IP-litigation department called "IRE" or "MIRE" (Meta Intellectual Rights Enforcement), that would be a little on the nose.The unofficial motto would be "We will bog you down in the courts for years"
It’s a noun meaning anger or wrath. Usually it’s specific not just to triggering anger the emotion, but an angry response. So attracting Meta’s ire means attracting an angry response from meta - like the legal response users of LLaMA have triggered.

It’s not an acronym.

The real thing we should be asking is if the 3B or 7B model is capable of correctly answering this question.
Do we even know if model weights are copyrightable or not?
No but unless you enjoy being sued by Meta you have to act as if they are.
I was asking myself the same question and am 99% sure it isn't protected by IP laws. It is another story for the training data and the source code used to run the model.
A JPG is copyrightable and that's just an array of numbers. Model weights seem pretty similar as a class.
Couldn't you just transform the weights in some predictable way and then untransform them before use in runtime? The weights at rest would be completely distinct from the meta weights.
It's not the JPG data you can copyright, but the recognisable image it produces. Every time you re-save or resize the image, the data changes, but the recognizable image remains.

It's not clear how this process applies to model weights. Once you run another training epoch on them, the data has changed. What is the essential copyrightable, trademarkable or patentable thing that remains? A legally untested question for sure.

But the data inside the jpg isn't generated by a machine, model weights are.
It depends on how the JPG is made. Some JPGs are not copy writable, like that picture a monkey took of itself. Model weights are probably (legally) more like a selfie of a monkey than a photographer's photo.
> Supportive. We build models to support our users, not replace them. We are focused on efficient, specialized, and practical AI performance – not a quest for god-like intelligence. We develop tools that help everyday people and everyday firms use AI to unlock creativity, boost their productivity, and open up new economic opportunities.

Refreshing take on the peak alarmism we see from tech "thought leaders"

I dunno... god-like intelligence would be pretty useful. I'll take a brochure.
do you trust god?
Well, ...

OK, I withdraw the comment.

Why should I need to? Isn't God on the blockchain? (j/k)
Depends on which one. All the ones described in religious books seem to have very poor alignment, though.
Doesn't sell as much, though
>Refreshing take on the peak alarmism we see from tech "thought leaders"

It's not alarmism when people have openly stated their intent to do those things.

Its alarmism to support government regulation to reinforce the moat when industry leaders say they intend to do it, but also that the danger of it being done is why competition with them must be restricted by the State (and why they can’t, despite being, or being a subsidiary of, a nonprofit founded on an openness mission, share any substantive information on their current models.)
This is just marketing. They're positioning themselves as somehow "more human" while building the exact same technology. When a model supports me by doing the work I'd otherwise hire someone to do, the model just replaced someone. And this goes without saying, but a large amount of outsourced tasks today don't exactly require "god-like intelligence".
That was probably said about the automobile, when it replaced horses, or about electrical lamps, when replaced oil-based lamps, no?

I mean, every city had an army of people to light up and down oil lamps in the streets, and these jobs went away. But people were freed up to do better stuff.

(comment deleted)
It is different this time. I bet that was also said when the transformations that you mentioned occurred, but this time it really is different.

LLM models are pretty general in their capabilities, so it is not like the relatively slow process of electrification, when lamplighters lost their jobs. Everyone can lose their jobs in a matter of months because AI can do close to everything.

I am excited to live in a world where AI has "freed" humans from wage slavery, but our economic system is not ready to deal with that yet.

> but this time it really is different

I'm skeptical. This will drastically change what it means to do a job in a way that has never happened before, but humans will find a way to deal with the fallout. We don't have a choice. Besides, if we were able to disrupt the very foundations of our economy for a minor virus, we can and will do the same to deal with this if required.

Either way this change has already arrived and we are starting to adapt our lives in response to it like we have many times in the past.

tldr: This change is significant but we'll manage.

I wouldn’t say the handling of COVID was smooth to say the least.

Yes we handled it, we are still paying the bill for that handling (inflation).

I think AI will have the disruption level of COVID, but there will not be an end in sight, 5%, 10, 20, 50% of people will lose jobs and even if they can refrain and handle it, it will take 5-10 years for those people to handle it. Can the countries have people on unemployment for that long ?

If COVID was to be worse on young, healthy people, instead of elders and debilitated, we'd be in serious trouble today. It was very badly handled...
Not like AI revolution will be better handled. It will be even worse because there are very obvious economic incentives to handle badly
Covid is intrinsically bad.

I don't think this is the case for AI.

I see a completely different picture.

Productivity will skyrocket and with it the standard of living. Humans will always enjoy having other humans doing stuff for them.

Sure, it will be faster this time and there will be some growth pains.

It's not a matter of being ready, it's a matter of needing this. If you look at society's problems today, we're in a deadlock. I believe the benefits of AI can help alleviate a lot.

But to whose pockets will that productivity go? I think the gap between the haves and have-nots will widen and just increase society's problems
It will most likely widen, but who cares? What matters to me is the quality of my life, not others. If they're managing to get better than me while doing something useful to society, good for them.

What really matters is: the poor of tomorrow will laugh at the life of today's rich.

I mean, the poor won't have the Bezos' yatch, but they'll have access to some life amenities, health resources, etc, that Bezos can't even dream of having today.

That’s bull, the poorest will have to fight for water
Yeah all the Terminator energy around these AI things is so off-putting. They aren't like that. They're big matrices and they are very cool tools!
What if big matrices are the last missing piece to research going on since the 50s…
But the concerns about AI taking over the world are valid and important; even if they sound silly at first, there is some very solid reasoning behind it. They’re big matrices, yes, but they’re Turing-complete which means they can theoretically do any computational task

See https://youtu.be/tcdVC4e6EV4 for a really interesting video on why a theoretical superintelligent AI would be dangerous, and when you factor in that these models could self-improve and approach that level of intelligence it gets worrying…

This comment basically implies I don't get it, but I will if I watch a Youtube video. I get it. ChatGPT isn't that. That's the point. You can have concerns about AGI. That's fine. But they have nothing to do with LLMs unless you are trying to play a shell game.
But you were talking about AI in general and dismissing the risk entirely as sci-fi.

I think a large enough LLM, or at least a slightly modified one, could lead to AGI and we’re not as far from it as you think

> They're big matrices and they are very cool tools!

Well, your mom is a etc

Edit: Since this is getting downvoted I'll be more explicit: The human brain may well be also just described as some simple sort of thing, but that doesn't mean humans are not dangerous, nor hypothetical humans with a brain ten times as large and a million times faster. The worry about AIs killing all humans soon is not naive just by sounding naive.

Sure, it's not naive just because it sounds naive. It's naive for other reasons (for one thing, we're really no closer to super-intelligent AIs than we were before the LLM craze began).
A lot of people would disagree with that. You can hardly deny that progress has sped up in the last few years, so I don't know why we shouldn't extrapolate this speed into the coming years.
Well, it's to their benefit to portray their models as working alongside and enhancing humans, as opposed to replacing us. So it sounds a bit like marketing speak to me.

And it's to the benefit of many of those tech "thought leaders" to be alarmist since they don't have much of the AI pie

Well exactly. AI _is_ a tool and a very good one at that.
"It is refreshing to hear opinions I already agree with. People with other opinions are unintelligent"

Is that what you were trying to convey? If not, I'm curious to know what you find refreshing about it and why those who disagree are wrapped in double quotes.

the models: https://huggingface.co/stabilityai/stablelm-base-alpha-3b, https://huggingface.co/stabilityai/stablelm-base-alpha-7b

There are also tuned version of these models: https://huggingface.co/stabilityai/stablelm-tuned-alpha-3b https://huggingface.co/stabilityai/stablelm-tuned-alpha-7b, these versions are fine-tuned on various chat and instruction-following datasets.

The Github repo mentions that the models will be trained on 1.5T tokens, this is pretty huge in my opinion, the alpha models are trained on 800B tokens. The context lenght is 4096.

These models are huge. I assume they are not quantized down to 4bits yet.
Quantized versions will pop up on huggingface very soon, if they arent already there. It takes basically no time, much less than something like a alpaca finetune.
I love stability ai so much
Thanks for breaking ground here and being truly OSS
On the off-note, can anybody tell me what's going on with embeddings, & vector databases? Certainly it would seem that forward-pass completion is pretty much solved, & a smaller, better model will appear eventually. Let's say you even managed to solve both complete() and embed() but what do you do with it, how are you going to organise, query, and multiply this dataset? Now the question I know that text-embedding-ada-002 has twice as many dimensions as mainstream Sentence transformers. Do we need all the extra dimensions? If not, how do I make it work better for my specific dataset with lots of jargon and abbreviations and stuff like that? What are the hardware requirements for that? I.e. could I do a fine-tuning job on some specific jargon-heavy text to get better embeddings for them? For one, the more I look into similarity-based use-cases the more I see that it's not normally speaking "top-percentile nearest-neightbour search" but the data is also terribly relational, i.e. it's probably like a slowly changing dimension, and there's a tree traversal type structure in how documents are generated as output from other documents as inputs? So you kind of have to think about these complete/embed ops both in aggregate; for batching but also in particular, from the cost/reward ROI type calculation. Not just in aggregate but also in terms of memory usage patterns to further optimise layout— tiering and stuff like that really comes to light.

Also: vector database shilling on HN is getting out of hand; multiple companies literally plugging every mention on the radar, some actively begging for upvotes. Looking at it all makes you really appreciate pgvector[1] to a point where you would be more willing to buy 3.2 TB of high-bandwidth NVMe and dedicate it to a large IFV index than ever have to deal with all of this "purpose-built vector database" bullshit.

[1]: https://github.com/pgvector/pgvector

Yes, you need all of the dimensions. All of the dimensionality reduction techniques, including SOTA ones (UMAP or better) are going to massively harm your embeddings.
Perhaps I didn't word by question correctly, I'm looking to compare capability of Sentence transformers vs. OpenAI Ada-based embeddings relative to their respective dimensionality?
No you don't need the extra dimensions and OpenAI is generally the worst at everything except being the first to market.

Also, ditto your comments on vector database shilling. Vector Databases are just like any other database in that I'll host them myself. I don't need a dedicated VC backed company for a database.

Just because you say it doesn't make it true.

Dimensionality reduction is an extremely destructive operation. Losing even the wrong single vector component of an embedding is massively damaging to down stream performance.

The AI war has officially begun.
No, that comes when China catches up in the quest for superintelligence.
Stop worrying about China…wow
I don’t think China has enough skilled AI workers to compete in any quick timespan. Top talent doesn’t want to live and work there generally.
I'm curious why you'd think that. China as a country has many people to start. Some percentage of these people will end up in AI. Assuming people from all countries are roughly equally intelligent, the numbers clearly favor China. Universities over there are quite good, there's a pretty strong "work hard" mentality I see from all our Chinese students. Plenty of Chinese graduating or starting university these days during the AI hype peak. China as a country isn't sleeping on AI either. I think China as an AI hub looks quite promising. Anecdotally, China also retains quite a lot of talent or people go abroad to study and return to China. Compared to some European countries or India that "leak" a lot of talent to the U.S. I think China is quite a bit more stable.

On the hardware side, things tend to be produced there as well.

China definitely "leaks" a lot of talent to American companies - most AI papers that I've seen from respected Western universities include at least one Chinese name.

One challenge for China has been the university enrollment rate. While in Western countries half of each cohort has been going to university for decades, China is not there yet. In 2019, just 17% of Chinese adults have degrees compared to 44% in the US.

So the large Chinese population is offset by its relative lack of access to education, while the US can draw from its own highly educated population in addition to attracting the best and the brightest from the rest of the world, including China.

(comment deleted)
There are plenty of authoritarians in the US to conduct warfare against our institutions that are up to speed. We don't need to wait for anyone else overseas to get this party started.
I’m more than ready. It’s exciting to see what can come out of this.
It's fantastic that more orgs are releasing open-source models trained on more than 300B or so tokens. Here's my take from the details I could find.

Pros

  - 4096 context width (vs 2048 for llama, gpt-j, etc)
  - 3B to 65B released or in progress
  - RL tuned models available
  - Trained on more tokens than existing non-llama models
  - 128 head dim, so can use flash attention (unlike GPT-J)
Cons

  - No benchmarks released, or details about the model
  - Somewhat restrictive license on the base models, and NC license on the RL models
  - Small models only trained on 800B tokens, compared to 1T for llama-7B, and potentially more for other upcoming alternatives (RedPajama, etc).  I'd like to see their loss curves to see why they chose 800B.
High-level, this is likely to be more accurate than existing non-llama open source models. It's hard to say without benchmarks (but benchmarks have been gamed by training on benchmark data, so really it's just hard to say).

Some upcoming models in the next few weeks may be more accurate than this, and have less restrictive licenses. But this is a really good option nonetheless.

> - 3B to 65B released or in progress

Seems they want to do 3B to 175B, although 175B is not in progress yet.

It's not efficient to do 175B. Training a smaller model (65B) on more data gives better performance for the same compute.
If you want it to just regurgitate training data, sure. But more parameters will always be better for more complex tasks.
> But more parameters will always be better for more complex tasks.

I think you should checkout this paper which discusses the relationship of performance and the ratio of training tokens to parameter count.

https://arxiv.org/abs/2203.15556

StableLM already has an optimal parameter number to tokens ratio, so what's your point? They should train the 65B model on even more tokens?

> StableLM is trained on a new experimental dataset built on The Pile, but three times larger with 1.5 trillion tokens of content

If I understand correctly, based on their prediction in Table 3 on page 8, they do have enough tokens, but they also need over a magnitude more compute time.

> It's not efficient to do 175B. Training a smaller model (65B) on more data gives better performance for the same compute.

This is OP's comment you replied to - so I was responding under OP's context that the amount of compute time would be the same, which I apologize I didn't make clear, and my response was very poorly worded.

My intent was to link the paper because I think it supports OP's statement that for the same amount of compute time and a token ratio, the performance of a smaller model will be better then a larger one (assuming they haven't converged yet which they haven't at this size).

> If you want it to just regurgitate training data, sure.

This paper was about showing Chinchilla performing with models many times larger then itself, showing you don't need to have a 175B size model for more performance then "regurgitating training data"

> you don't need to have a 175B size model…

Sure, that’s true.

…but, a fully trained larger model is going to be better.

There only reasonable reason to prefer a smaller model is because it’s cheaper and less intensive to train.

It sounds a lot like you’re saying “small models are just as good” … which is false. No one believes that.

For a given compute budget an under trained large model and a well trained small mode may be comparable, right?

…but surely, the laws of diminishing returns applies here?

There’s an upper bound to how good your smaller model can ever be, right?

Over time, someone can take a larger model which is under trained and refine that model right?

The “small model is just as good” narrative only holds up for a fixed once only training of a model for a fixed compute budget at the moment of release.

Over all of time that compute budget is not fixed.

> It sounds a lot like you’re saying “small models are just as good” … which is false. No one believes that. … a fully trained larger model is going to be better.

You're absolutely right, a fully trained larger model _will_ be better. This is meant to be under the context of OP of a "limited compute", the statement I'm trying to make is “fully trained small models are just as good as a undertrained large model”.

> …but surely, the laws of diminishing returns applies here?

They do but it's diminishing in that the performance gains of larger models becomes less and less, while the training time required changes a lot. If I'm reading the first chart of figure 2, page 5 correctly, you a 5B vs 10B, the 10B needs almost 10x the training time for a 10% loss gain. and its a similar jump from 1B to 5B. My understanding is at this also starts flattening out, and that loss gain from each 10x becomes gradually lower and lower.

> Over all of time that compute budget is not fixed.

Realistically there is an upper bound to your compute budget. If you needed 1000GPUS for 30 days for a small model, you need 1000GPUS for 300 days for that ~10% at these smaller sizes, or 10,000GPUS for 30 days... You're going to become limited very quickly by time and/or money. There's a reason openai said they aren't training a model larger then GPT 4 at the moment - I don't think they can scale it from what I think is a ~1~2T model.

The optimal training tokens for 65B parameters is like 80T.

Emad tweeted "Goin to train a 3B model on 3T tokens" last month. These 800B checkpoints are just early alpha training checkpoints.

The full training set is 1.5T currently and will likely grow.

and also easy to deploy
>Small models only trained on 800B tokens

"These models will be trained on up to 1.5 trillion tokens." on the Github repo.

https://github.com/stability-AI/stableLM/#stablelm-alpha

Devs confirmed that the small ones use 800B, 1.5T is for the large ones
@thunderbird120 asked a Stability employee and say that the plan is going to keep training the models up to 1.5T. So I don't know where do you read this.
That may be, but the weights you can download today were trained on 800B
yes of course that's why they use "will be trained" on the GH repo.
I think they are “checkpoint” models in this case.

Will be fun to compare when completed!

Are not all models checkpoints? I think you may be interpreting it too colloquially.
That's great news, but one would think that since they're behind Stable Diffusion, that they'd use the insights behind it and scale data even more than that to result in better quality at a smaller scale model that can run on most people's machines.

Like... try 10 trillion or 100 trillion tokens (although that may be absurd, I never did the calculation), and a long context on a 7B parameter model then see if that gets you better results than a 30 or 65B parameter on 1.5 trillion tokens.

A lot of these open source projects just seem to be trying to follow and (poorly) reproduce OpenAI's breakthroughs instead of trying to surpass them.

> Like... try 10 trillion or 100 trillion tokens (although that may be absurd, I never did the calculation)

But where’s the corpus supposed ro come from?

>try 10 trillion or 100 trillion tokens

Computation is not free and data is not infinite.

You could've said the same to OpenAI when they were scaling GPT from 1 billion to 175 billion parameters. We're all grateful they didn't follow that line of thought.

But Stability does have access to a pretty big cluster, so it's not paying cloud compute (I assume), so cost will be less, and data of course is not infinite...never stated that.

But considering 3.7 million videos are uploaded to youtube everyday, 2 million scientific articles published every year, yada yada...that argument falls apart.

At the very least implement spiral development... 1 trillion... 3 trillion... (oh it seems to be getting WAY better! There seems to be a STEP CHANGE!)... 5 trillion... (holy shit this really works, lets keep going)

The training corpus is the problem. An extra trillion tokens is (ballpark) an extra million KJV bibles worth of text formatted for ingestion. And you probably picked all of the low hanging fruit in terms of quality prior vetting and being in a standard format for ingestion in your first trillion tokens of training data.
There’s a difference between telling someone they’re wasting their time with their current project, and asking them why they didn’t spend 6x - 60x as much budget on an already expensive project.
They're loaded, and we know scaling works, they'd massively benefit... both in marketing and profit.

Although it is open source to be fair.

Nobody knows where to find 10 trillion tokens of good data. Publicly available / data without a license seems to cap at around 1.5 trillion tokens total. The internet isn't as big as you thought! (Or at least, all the good stuff is behind a walled garden, which I think we did know)
They mention 1.5T training tokens, perhaps for the largest model only ?
It's unclear which models will be trained to 1.5T tokens. The details of how many tokens each model saw in training are on Github - https://github.com/stability-AI/stableLM/ . But only for the ones that have been released.
I just asked a stability employee and they said the the current models ran into an overfitting issue probably due to some duplicated data somewhere in their dataset, which consists of 1.5T tokens. The 800B tokens is the number of tokens they've been trained on so far. The plan is to keep going and train on the rest of the data once the issue is resolved.
I've asked this question in a few places, and never been able to get an answer, maybe you know...

Q: Why are these LLMs trained on a single epoch, and perform worse if the dataset is repeated ?

This seems maybe related to suspecting data duplication as a cause of overfitting.

Why don't LLMs need multi-epoch training at a low learning rate to generalize? If they are managing to learn from a single epoch, that sounds more like they may be memorizing!

Never repeating your training data is what you'd ideally like to do for training basically any ML model. If you do that you don't really need to worry about overfitting since the model is constantly trying to fit a stream of new data. To reduce its training error it actually has to model the structure of the data rather than just memorizing it since each training step will involve data it has never seen before. Larger models are more prone to overfitting but also learn several orders of magnitude faster. If you can use larger models without being concerned about overfitting it's generally desirable to do so. It's just that most tasks don't actually have enough data to support doing that. Thankfully, text modeling does have enough data.
Thanks.

So when, for example, we train an ImageNet model over multiple epochs using rotation/scaling/etc augmentation, it's really better to think of this as one epoch over a unique set of images than multi-epoch per se ? I was really thinking of augmentation as a way to get coverage over the input space rather than ensuring the training data doesn't repeat, but I guess it serves both purposes.

It does still seem that many LLMs are overfitting / memorizing to a fair degree though - maybe just because they are still too big for the amount of data they are trained on ? It seems like a bit of a balancing act - wanting an LLM to generalize, but yet also to serve as somewhat of a knowledge store for rare data it has only seen once.

> Small models only trained on 800B tokens, compared to 1T for llama-7B

LLaMA is trained far beyond chinchilla optimality, so this is not as surprising to me.

But Chinchilla optimality, while an interesting result, is a strange target for most practical purposes. Training is one time, inference is many times; not training past the point where its cheaper to training a larger model for the same (proxy for) quality discounts to zero the import of the cost of inference.
Yep, but if stability has the goal of training the best possible model then that would explain the choices they made.
I mean 800B tokens on a 3B model and 7B model is still way beyond the Chinchilla scale.
They're going to 1.5T and possibly 3T. The 800B is just for the "Alpha" checkpoints released today. New checkpoints will be released later.
I'm wondering what the sweet spot for parameters will be. Right now it feels like the Mhz race we had back in the CPU days, but 20 years later I am still using a 2-3GHz CPU.
I think "sweet spot" is going to depend on your task, but here's a good recent paper that may give you some more context on thinking about training and model sizes: https://www.harmdevries.com/post/model-size-vs-compute-overh...

There have also been quite a few developments on sparsity lately. Here's a technique SparseGPT which suggests that you can prune 50% of parameters with almost no loss in performance for example: https://arxiv.org/abs/2301.00774

I was wondering if the longer training thing was a similar phenomenon to the double-descent we see in other deep learning models. Training for a really long time can improve generalization (as can adding more parameters) - but I don't know enough about LLM architecture to know if that's relevant here. My skim of the blog post led me to think it's proposing a different mechanism (scaling laws).
Well, based on all the data we have available now it seems like you don't get much benefit yet from going above 200 billion.
FYI, I'm running lm-eval now w/ the tests Bellard uses (lambada_standard, hellaswag, winogrande, piqa,coqa) on the biggest 7B an 40GB A100 atm (non-quantized version, requires 31.4GB) so will be directly comparable to what various LLaMAs look like: https://bellard.org/ts_server/

(UPDATE: run took 1:36 to complete run, but failed at the end with a TypeError, so will need to poke and rerun).

I'll place results in my spreadsheet (which also has my text-davinci-003 results): https://docs.google.com/spreadsheets/d/1kT4or6b0Fedd-W_jMwYp...

Do you also have results of GPT4 somewhere? or text-davinci-003-turbo
I'm still on the waitlist for GPT-4 API access. Note, that text-davinci-003 cost about $90 to benchmark at $0.02/1K tokens, so if you're able to use a GPT-4 model (for completion and not just instruction) that'll probably be $270-$540 in credits to benchmark...
Just think about benchmarking 32K GPT4 haha
I have GPT-4 8k access and am willing to run the evals if someone wants to pay. Email in my acc info (the character is h)

Just a note, I get errors semi-frequently when running queries against GPT-4 often (timeouts mostly…) so any code would need to handle that well.

You should benchmark GPT-3 Curie (7B) for comparison since it is the same size as llama-7B and StableLM-7B.

That will give us some indication of how much better these models are than GPT-3 at the same size.

How possible is it that every other model suffers from dataset contamination and this model is being unfairly penalized for having properly sanitized training data?
Note that this is StableLM ALPHA (only 0.52 epochs into training).

The fully trained version will surely be much better.

Also, you should benchmark GPT-3 Babbage for a fair comparison since that is the same size as 7B.

Looks like my edit window closed, but my results ended up being very low so there must be something wrong (I've reached out to StabilityAI just in case). It does however seem to roughly match another user's 3B testing: https://twitter.com/abacaj/status/1648881680835387392

The current scores I have place it between gpt2_774M_q8 and pythia_deduped_410M (yikes!). Based on training and specs you'd expect it to outperform Pythia 6.9B at least... this is running on a HEAD checkout of https://github.com/EleutherAI/lm-evaluation-harness (releases don't support hf-casual) for those looking to replicate/debug.

Note, another LLM currently being trained, GeoV 9B, already far outperforms this model at just 80B tokens trained: https://github.com/geov-ai/geov/blob/master/results.080B.md

> 128 head dim, so can use flash attention (unlike GPT-J)

mind explaining why this is so attractive/what the hurdle is for the laypeople in the audience? (me)

Standard attention has memory quadratic in sequence length, whereas FlashAttention has memory linear in sequence length. Also FalshAttention is faster.
So there must be a downside to FlashAttention. What is it?
FlashAttention is mathematically identical to standard attention, so in theory there's no downside. In practice, numerical inaccuracies of floating point mean that the results differ slightly. I don't know of any papers going in depth to analyze what impact those variances have in a range of real models, but generally speaking deep models handle slightly variances well. I've not noticed any difference in my applications training models. And tons of people use FlashAttention as a drop-in replacement on models trained on standard attention (e.g. using xformers in StableDiffusion).

Also in practice FlashAttention is still relatively new so it isn't well supported in libraries yet. Until PyTorch 2.0 you had to either implement it yourself, or use something like xformers which comes with a bag of caveats. PyTorch 2.0 now has it built-in, and it's easy to use, but the implementation is incomplete so you can't, for example, use it with an attention mask (which is needed in LLMs, for example).

tl;dr: Basically none, but it just isn't well supported yet.

installing it is a nightmare
According to the paper Flash Attention also needs quadratic memory:

Let 𝑁 be the sequence length, 𝑑 be the head dimension, and 𝑀 be size of SRAM with 𝑑 <= 𝑀 <= 𝑁𝑑. Standard attention (Algorithm 0) requires Θ(𝑁𝑑+𝑁²) HBM accesses, while FlashAttention (Algorithm 1) requires Θ(𝑁²𝑑²M⁻¹) HBM accesses.

https://github.com/HazyResearch/flash-attention#memory

"standard attention has memory quadratic in sequence length, whereas FlashAttention has memory linear in sequence length."

I guess you have just reported how many times the layer will need to access the memory, not how much memory usage scales with sequence length.

Were you able to figure out if the RL models are going to be jailed? A 65B parameter model could be a bit frightening. That's 1/3rd the size of GPT3.
I'm sure there will be a bunch of different RL tuned versions of them, RLHF isn't that expensive. IIRC Microsoft has software that will do it for a few thousand dollars for a model that size. I'm sure someone will release a non-lobotomized version, maybe OpenAssistant.
its not alway about the size, but yeah its really good!
(comment deleted)
I feel "The Pile" needs a lot more to match what OpenAI has trained their models on. The mysterious "Books2" and so on.
I think OpenAI has a few hidden advantages that are not obvious at this point. It could be additional training data, filtering/preprocessing that data, some changes to the architecture, who knows? None of the open source models are even close to GPT 3.5, what to speak about GPT 4? I've tried everything and the 60G llama variants so i'm not sure it's about number of parameters. They definitely have some hidden sauce.
Great to see Stability release this with a great license as well. Any idea on the usecases for the 3B model? Will a model that small suffer heavily from a 4bit quantization?
Generating either a 0 or 1 in response to something? Maybe useful in some testing scenarios? Honestly don't know either. New lorem ipsum?
You can use it as the assistant model to a large model, it's called speculative sampling. You generate text with the small model and validate with the large one, ensuring no deviation occurs. Speedup of 2.5x

https://arxiv.org/abs/2302.01318 (DeepMind)

That would be 1bit quantization. In reality quantization under 8bits is done in smart ways which result in higher effective output quantization and lower effective memory size quantization.

For example, bucketing identical groups of weights and then reusing one bucket for all the identical groups lowers the effective bit quantization at the memory level while retaining the uncompressed quantization quality.

There is literature on effective quantization levels below 1 ("So called 0bit quantization). But even then the actual weights are typically 2-4 bits and there is just a lot of reuse of weights going on.

Another neat trick is to line bins of weights up in order of lowest to highest weights and compute a function to produce a zero offset throughout the bins of weights; such that 0010 in one bin and and 0010 in another bin are unsampled to 16bit and then have different offsets added, maintaining uncompressed 16bit performance without the memory overhead.

There are many more tricks like this and many still to be found!

> Will a model that small suffer heavily from a 4bit quantization?

From my experience with quantized 7B llama models, avoid 3B if you can. Without benchmarks, I think this is a decent rule of thumb.

Absolutely a giant fan of Stability staying to actually open source licenses and not licenses that impose restrictions on what you can use it for. This is the future of AI! Beware of any org that uses "ethical" licenses - they are not open source. Stability is one of the few organizations that actually cares about free software, you love to see it.
> These fine-tuned models are intended for research use only and are released under a noncommercial CC BY-NC-SA 4.0 license, in-line with Stanford’s Alpaca license.

This is a no-commercial-use-allowed license; it is neither considered free software nor open source, the definitions of which disallow restrictions on what you can use the work for.

The two sentences prior are important:

> We are also releasing a set of research models that are instruction fine-tuned. Initially, these fine-tuned models will use a combination of five recent open-source datasets for conversational agents: Alpaca, GPT4All, Dolly, ShareGPT, and HH. These fine-tuned models are intended for research use only and are released under a noncommercial CC BY-NC-SA 4.0 license, in-line with Stanford’s Alpaca license.

The snippet you quoted is not talking about the main model in the announcement. It's talking about fine-tuned models based on other models. Stability has to respect the license of the originals. They cannot change it.

The main model is described higher up in the post and is permissible for commercial:

> Developers can freely inspect, use, and adapt our StableLM base models for commercial or research purposes, subject to the terms of the CC BY-SA-4.0 license

It also appears that CC BY-SA-4.0 is GPL-compatible. Not a lawyer and this is not legal advice, but it certainly seems like one could operate their own StableLM server/service and allow proprietary code to use it over a network interface, much like one could use a GPL-licensed database system.

https://creativecommons.org/faq/#can-i-apply-a-creative-comm...

Not their fault, the instruct-tuned models depend on non-open data.... Which should be open however. Scraping chatGpt is legal
Agreed. Scraping ChatGPT is against OpenAI terms of use and OpenAI is entitled to terminate your access immediately upon notice, but since ChatGPT output is not copyrighted (and copyrightable), output you acquired before termination should be freely redistributable. I am not sure why Stanford Alpaca authors think otherwise but they are wrong.

https://openai.com/policies/terms-of-use

Thank you for using OpenAI! These Terms of Use apply when you use the services of OpenAI, L.L.C. (snip) By using our Services, you agree to these Terms. (snip) You may not (iii) use output from the Services to develop models that compete with OpenAI. (snip) We may terminate these Terms immediately upon notice to you if you materially breach Sections 2 (Usage Requirements).

Interesting. A non-LLM hallucinating. And to think we used to believe that was only a property of LLMs.
Huh? Everything they said is true, isn’t it?
StableLM, the model that's theirs is CC-BY-SA 4.0.
The instruction tuned versions are under a non-commercial license though because some of the data they used is
(comment deleted)
"Ethics" will only ever be an excuse to lock this technology behind one companies paywall. The only ethical AI is actually free and open AI, how its trained is irrelevant imho as long as we can all benefit. The negatives of the work of individuals being used to train it outweigh the negatives of one company just doing that and holding the power within their walls.
Yeah I wish there was more real investigation / analysis into who is behind various "ethical AI" pushes and what they stand to gain from it. From what I can see, many of the people involved either are invested in companies that will somehow certify your AI is ethical, or just want to stifle competition so they can catch up. Of course there's also a sprinkling of "current thing" supporters.
I have to disagree. Especially in the case of LLMs where new API services are popping up all over the place, an "ethical" license like agpl that requires the source be shared for web services would would accelerate development of the space as a whole immensely.
I think when they say ethical they're talking about RAIL which has clauses about not doing bad guy stuff with their models, not AGPL.
I am definitely not in favor of OSS license for models >= GPT3...

Not unless they're aligned well.

There are all sorts of horrible use cases that these could be used for.

it is true that there are concerns relating to open source and ai, but surely the having them be closed off, manipulated and controlled untrustworthy corporations is worse.
they like the particular techbro that runs the corporation currently. people will find all sorts of ways to lick the boot of private ownership.
Both these options don't feel good to me. Hard to really tell what is ultimately worse, when I can imagine similar outcomes when irresponsible or malicious agents have access to sufficiently powerful AI.

Main positive point for open models is that we will start seeing the abuse sooner and at smaller scales. That might give us more time to build an immune system up against exploits by encouraging us to prioritize development of comprehensive AI safety practices.

For the worst use cases, like scams, why would they respect the license? Though that probably applies for some legal but immoral use cases
The ideal alignment for a computer program is to be aligned completely with the user, prioritizing their needs to the exclusion of all others.
"Alignment" is just a euphemism for "agrees with me", though. Humans aren't even aligned with each other. Demanding that AI models be "aligned" is essentially a demand that AI only be produced which agrees with your priors.
What your essentially saying is "alignment is very hard", which is what those researching alignment say. And they often use the example of how inter-human alignment is hard as evidence for why it is a hard problem. But saying it is hard is not an argument for why it is essential or not.

While humans are not perfectly aligned, especially if you just look at individuals, we are collectively aligned enough that many people can live together in communities of various scales. That imperfect alignment has been good enough that we have scaled from small tribal groups to an international network of nations. We need AI alignment to be good enough if we hope to continue advancing.

How do you feel about AI which is aligned with Iranian or Saudi cultural norms?
This presumes a lot of breakthroughs in model interpretability, corrigibility and of inner alignment. Since those are a prerequisite for AGI that we can live along side, I'd have some amount of relief that we found at least a temporary solution (but will those solutions scale to ASI?).

Now, if Iran created an AGI that poorly aligned with the global community before other nations had similar AGI, then then I suspect that would result in a future world I wouldn't be happy with. But it could be much better than a world with AGI that is unaligned with any human values, regardless of who created it.

My best case scenario could be AGI being created by a broad international coalition that is able agree with some combination of capabilities and alignment. I'm not very confident that this is our future, though. If anyone is going to do it, I think it is more likely that the USA would be the first to create a culturally aligned AGI. Which of course would still be considered a disaster for incongruent cultures.

There are all sorts of stuff people use www for, but here we are.
How is this sort of thing audited? I imagine there are all sorts of lifestyle AI businesses that won't give two shits about a license where people can't easily see or audit what is being used.
I am very happy to see them use a true FLOSS licence. However, it's a surprise to me, given Stable Diffusion is proprietary, using one of those "ethical" licences.
Indeed thats why I pay for credits on their official site/dream studio even though I want to run things locally. My big fear is one day they’ll make a press release saying they have to stop everything because not enough funding.
“also fine-tuned the model with Stanford Alpaca's procedure using a combination of five recent datasets for conversational agents: Stanford's Alpaca, Nomic-AI's gpt4all, RyokoAI's ShareGPT52K datasets, Databricks labs' Dolly, and Anthropic's HH. We will be releasing these models as StableLM-Tuned-Alpha.”

They have released the 3B and 7B of both the base and instruction tuned models. 30B and 65B in training and released later.

How can one get started adding RLHF to something like this?
Pretty soon there are going to be desktop apps for that
Demo is overloaded. Here is Colab link if you wanna check the quality: https://colab.research.google.com/github/Stability-AI/Stable...
Your session crashed after using all available RAM. If you are interested in access to high-RAM runtimes, you may want to check out Colab PRO
You can try 8bit
Using 8-bit still runs out of RAM for both the 3B and 7B models. It's unclear if it's because it still uses more than the available RAM, or if it's just quietly not using 8-bit since it's not implemented.
This notebook will need high ram and a premium GPU. It blew past the 16gb GPU that comes with Colab Pro
(comment deleted)
[flagged]