text stringlengths 0 2.46k |
|---|
[4890.50 --> 4895.14] It was the largest model of its type in the world by a sizable margin. |
[4895.22 --> 4897.44] It was about three times the size of the four. |
[4897.92 --> 4901.76] Four times the size of the largest open source model of its type in the world. |
[4902.44 --> 4902.84] Yeah. |
[4902.84 --> 4915.48] And the Pythia models, we trained on, like, 120 A100 GPUs for a couple weeks, which is certainly a lot of computing resources, but it's not, like, mind-boggling amounts of compute. |
[4915.60 --> 4919.02] There are lots and lots and lots of companies that have that that could... |
[4919.02 --> 4926.28] You know, it's less about it actually being too expensive and more about kind of having the political will to actually go do it. |
[4926.62 --> 4926.82] Yeah. |
[4927.04 --> 4929.68] Are you focused on training open source models? |
[4929.80 --> 4930.38] Is that your focus? |
[4930.38 --> 4934.68] So our focus is on open source AI research in general. |
[4935.16 --> 4945.66] Our kind of area of expertise is large-scale AI, and most of what we do is language models, but we've also worked on training and releasing other kinds of large-scale AI models. |
[4945.78 --> 4947.74] So we are part of the OpenFold project. |
[4948.90 --> 4956.18] So DeepMind created an algorithm for modeling protein interactions called AlphaFold. |
[4956.24 --> 4957.34] That was a really big deal. |
[4957.34 --> 4964.30] And we helped some academics scale up their research and get that and replicate that and release it open source. |
[4965.00 --> 4974.08] We've done some stuff in the text-to-image space, both on our own, and some of our staff have kind of gone on and worked at Stability AI on some of their language... |
[4974.08 --> 4975.22] Sorry, image models. |
[4975.22 --> 4979.48] ...and we are a big proponent of open source research in general. |
[4979.48 --> 4993.82] So our kind of...the reason we decided to start training these large language models was back in the summer of, like, 2020, we thought, you know, this G2B3 thing is going to be a major player in the future of AI. |
[4993.82 --> 5003.94] And it's going to be really essential if you want to be...if you want to be doing something meaningful in AI, you probably want to know how these things work. |
[5004.00 --> 5006.12] You want to be able to experiment with them and want to have access to them. |
[5006.40 --> 5009.54] And back then, you couldn't even pay OpenAI to let you use the model. |
[5009.92 --> 5010.00] Yeah. |
[5010.10 --> 5011.56] They announced that they had it, and that was it. |
[5011.56 --> 5015.22] And so we said, well, what the...let's try to train a model like that. |
[5015.30 --> 5016.78] We'll learn something along the way. |
[5017.18 --> 5021.98] And so we started building, like, an open source infrastructure for training large language models. |
[5022.20 --> 5028.14] We created a data set called the Pile, which is now kind of the de facto standard for training large language models. |
[5028.50 --> 5040.34] We created a evaluation suite for consistently evaluating language models, because everyone runs their evaluations a little differently, and there's huge reproducibility issues. |
[5040.34 --> 5050.02] So we built a framework that we could release open source and run on our own models, run on other people's models, and actually have kind of meaningful apples-to-apples comparisons. |
[5050.68 --> 5052.28] And we started training large language models. |
[5052.38 --> 5058.72] We trained a 2.7 billion parameter model, which was, like, a little bit bigger than G2B2 was at the time. |
[5059.02 --> 5060.62] And then we started training larger models. |
[5060.74 --> 5066.06] 6 billion parameters was the largest open source G2B3 style language model in the world. |
[5066.06 --> 5072.84] 20 billion parameters was the largest language model of any sort to be released open source in the world. |
[5073.54 --> 5079.52] You know, since then, there's been a lot more investment and willingness to train and release models. |
[5079.66 --> 5081.40] There's several companies that are now doing it. |
[5081.96 --> 5087.36] So Mosaic is a company that released a 9, I want to say, something. |
[5087.70 --> 5088.60] A large language model. |
[5088.60 --> 5092.38] That seems really excellent, like last week. |
[5093.04 --> 5097.40] There is Meta, which has been training and releasing sort of models. |
[5097.92 --> 5102.54] They'll tell you that they're open source releasing models, but that's just not actually correct. |
[5103.36 --> 5108.76] They're under non-commercial licenses, and they're not open source, despite their rhetoric to the contrary. |
[5109.48 --> 5111.20] But there's a whole bunch of companies. |
[5111.44 --> 5113.42] Stability AI is training large language models. |
[5113.54 --> 5117.30] So now there's a lot more people in this space and doing it and releasing it. |
[5117.30 --> 5122.24] And honestly, from my point of view, we got into training large language models mostly because we wanted to study them. |
[5122.32 --> 5127.44] We wanted to enable people to do essential research on interpretability, ethics, alignment, |
[5127.60 --> 5130.66] understanding how these models work, why these models work, and what they're doing, |
[5131.02 --> 5137.92] so that we can design better models and so that we can know what appropriate and inappropriate deployment contexts for them are. |
[5138.68 --> 5143.42] And so now that there's a lot more people working in kind of this open source training space, |
[5143.42 --> 5148.66] we're moving more towards doing that kind of scientific research that we've always wanted to do. |
[5149.26 --> 5155.06] So in the past six months, we've been doing a lot of work in interpreting language models |
[5155.06 --> 5158.28] and kind of understanding why they behave the way they do. |
[5158.68 --> 5164.90] My personal kind of area of focus is tracing the behavior of language models back to their actual training data. |
[5164.90 --> 5170.62] So the models that the DALI-2 is trained on, the Pythia suite, what kind of makes that special |
[5170.62 --> 5175.94] is that most language model suites are very ad hoc constructed. |
[5176.60 --> 5180.20] I'm calling them suites because you have several models that are similar of different sizes. |
[5180.70 --> 5180.82] Right. |
[5180.98 --> 5188.24] So like the OPT suite by Meta, for example, ranges from 125 million parameters to 175 billion parameters. |
[5188.76 --> 5191.92] But they're not actually very consistent between them. |
[5191.92 --> 5194.50] Some of them even have different architectures. |
[5194.62 --> 5195.56] They have different data order. |
[5196.10 --> 5200.58] There's a lot of stuff that kind of limits your ability to understand, |
[5201.24 --> 5203.12] to do controlled experiments on these models. |
[5203.30 --> 5208.40] And so we sat down and we said, if we wanted to design from the ground up a suite of large language models |
[5208.40 --> 5211.62] that was designed to enable scientific research, what would it look like? |
[5211.92 --> 5213.62] What kinds of properties would it have? |
[5213.82 --> 5217.14] What kinds of experiments do we think people are going to want to do that we're going to need to enable? |
[5217.14 --> 5222.92] And we built this list of requirements and then created a model suite that satisfies that. |
[5223.28 --> 5225.46] So it was trained on entirely publicly available data. |
[5226.10 --> 5228.18] All of the training, it was trained on the same data. |
[5228.38 --> 5230.64] Every model in the suite was trained on the same data in the same order. |
[5231.26 --> 5233.64] And we have a whole lot of intermediate checkpoints that are safe. |
[5233.70 --> 5239.54] So if you want to know, you know, after 10 billion tokens, how each model in the suite is performing, |
[5239.68 --> 5242.36] you can go and grab those checkpoints after 10 billion tokens. |
[5242.36 --> 5247.08] And then you can say, okay, what's the next data point it saw during training after 10 billion tokens? |
[5247.16 --> 5248.84] What was the 10 billion first token? |
[5248.94 --> 5254.40] And you can actually use some stuff we've uploaded to the internet to actually load that data |
[5254.40 --> 5256.02] in the same order it's seen by the models. |
[5256.10 --> 5260.68] You can study kind of how being exposed to particular training data influences model behavior. |
[5261.14 --> 5264.76] So we've been using this right now primarily to study memorization, |
[5265.06 --> 5270.44] understanding because language models have a pre-pensity for reproducing long exact sequences |
[5270.44 --> 5271.82] from their training corpus. |
[5272.36 --> 5275.72] And we're interested in understanding what causes memorization, |
[5276.28 --> 5279.00] why certain strings get memorized and others don't. |
[5279.20 --> 5282.20] Right now I'm wrapping up our kind of first paper on that. |
[5282.26 --> 5285.16] We have some more research in the works, trying to understand, you know, |
[5285.46 --> 5288.40] looking at the actual models throughout the course of training |
[5288.40 --> 5290.58] and looking at kind of the training data points that they see |
[5290.58 --> 5295.98] and trying to reverse engineer what that actual interaction between the model and the data is. |
[5296.74 --> 5299.06] And yeah, this is something I'm personally really high on. |
[5299.06 --> 5306.20] Most interpretability research right now is kind of focused on final trained models as like pre-existing artifacts. |
[5306.20 --> 5310.40] So you have this trained model and you want to understand what behaviors it has. |
[5310.40 --> 5317.72] But, you know, my perspective as someone who trains these models is much more focused on kind of where they come from. |
[5318.04 --> 5321.58] And what especially like my overarching goal is to kind of, you know, |
[5321.68 --> 5327.06] if I as a person who trains a large language model have a particular desire for a property the model has, |
[5327.12 --> 5328.42] a property the model doesn't have, |
[5328.42 --> 5333.50] what decisions can I make to actually influence that and to make the model have the properties I want it to have. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.