Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAI

AI Engineer · 19 min · 217 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:13Good morning everybody.
00:15My name is Ari Marcos.
00:16I'm the CEO and co-founder of Datlogy AI.
00:19Uh and really excited to kick off the data quality track uh today.
00:23Uh data quality is is what we live and breathe uh at Datlogy.
00:26It's all we think about.
00:27In fact, the company's name literally means the science or study of data.
00:31Um, so very excited to see the increasing excitement and interest in this area and a amazing lineup of talks uh today.
00:38So today I'm going to tell you about why data quality is a compute multiplier that we're all overlooking and where we can make massive gains by just working on better data.
00:47Um, we've seen that compute availability over the last six months has become extremely scarce and is only getting worse.
00:55We saw H100 prices reverse their several year-long drop which is normal for hardware and all of a sudden come up where now they're about 40% up from their lows
01:03um at the end of last year.
01:05Um as test time compute has become a critical part of models and as we have put more and more thinking tokens in we're seeing that the number the token usage is absolutely
01:15skyrocketing.
01:16Um reasoning models use eight times as many tokens um as non-reasoning models and that's projected to 5x again in the next year or so.
01:23Um, so the number of tokens we're pushing through goes higher and higher and that constrains compute even further.
01:28And this has led actually to a world where it's not implausible today that we might see access to some of the frontier APIs actually get limited or go away.
01:36Um, as one example, Google just capped uh Meta's Gemini usage because of inference constraints.
01:41OpenAI has effectively started selling token futures or you can guarantee token capacity some amount of time into the future.
01:47This is only necessary because they are legitimately wondering there might be a world where access to frontier API tokens is limited not as a business decision but because there's just simply not enough inference and first party products will be prioritized.
01:59So in a world where compute is increasingly scarce and you need to make models better.
02:04Well, what do you do?
02:05Well, we work on data.
02:06Um I you're going to hear me say this over and over again.
02:08Data quality is a compute multiplier because what it does is it it makes the learning curve steeper.
02:13So this is just a very simple schematic of performance on the y-axis as a function of data on the x-axis.
02:18Note that this x-axis, you could swap it out for data, for compute, for time, for dollars.
02:23They're all the same x-axis fundamentally.
02:26Um, and if you can make data quality better, you can turn this gray curve um into this blue curve.
02:32And that means that now you can get dramatically better performance for the same compute budget as if you had trained with far more compute.
02:39Um and similarly you can get the same performance for a much smaller compute budget.
02:44Um which is exactly showing you how you can get performance as if you had spent 10 times this 10 times as much on compute and really shows this compute multiplier
02:53point.
02:55Um so how do you actually do this?
02:56Well fundamentally the idea is we want to make it so that we get the maximum signal per token and per batch.
03:02Um for those that are a little bit more technically what we want to want to do here is maximize the marginal information gain per data point when we show it to the model.
03:09what data is going to teach the model the most.
03:11Um, and that's all about finding data that's relevant to the use cases that you want.
03:16One thing that you know is very important is that there's no one golden data set to rule them all that's good for everything no matter what you want to do.
03:23A data set's only going to be optimal with respect to a particular set of output tasks that you want the model to do.
03:28So if you want a great legal model, you're going to want legal data more than healthcare data and vice versa.
03:33It needs to be diverse.
03:34A lot of the issues we see with model robustness and brittleleness comes from try training on data that's not diverse enough.
03:40So the model can answer a question correctly if it's presented just so but if it's presented a little bit differently now everything breaks um needs to be information and you have to mix the data correctly.
03:49This is a hugely difficult part of you now have many different sources.
03:52How do you combine them to actually drive the largest improvement in performance?
03:56So um this is a high level of what we do at here.
04:00Um you can think of us as the oil refinery for data.
04:02We don't source new tokens like many data providers.
04:04Rather we take existing tokens coming from public data sets, proprietary data sets and licensed data sets and make them way better.
04:10And how do we do that?
04:11Well, we do that through these four C's.
04:13Um, clean, curate, create, and compose.
04:15Um, so cleaning is fairly straightforward.
04:17This is doing things like heruristic filters, all a gopher and things like that.
04:21Removing documents that have, you know, only 10 characters in them or all winging.
04:25That's kind of basic table stakes.
04:27Um, benchmark decontamination is incredibly important.
04:30As I'm sure you all know, benchmaxing has become a real problem and makes it very difficult to interpret model results.
04:35So, we rigorously decontaminate all of our training data with respect to all downstream benchmarks with a pretty low engram to ensure that that's the case.
04:43That gets you to a point where now you can feed the model into the data, but it's still sorry, feed the data into the model, but it's still not very good.
04:49So then how do you make it better?
04:50Well, it's a combination of many things ranging from quality classifiers and tonomy across different topics and balancing that redundancy reduction.
04:57So removing data points that are not the same that are semantically similar but convey very similar information even if they're not the same pixels themselves say upsampling and downsampling data points based off the quality and the relevance and then task distribution matching identifying what data do you actually need in order to solve this given task.
05:14That now gives you a data set that is very high quality but is typically still too small.
05:19And that's where synthetic data comes in.
05:20Now we can go and rephrase that data as effectively a very fancy form of data augmentation to produce dramatically more data in many different formats and this helps a lot both with kind of uh data size and with diversity
05:33um because we can really inject a lot of diversity in through this and then finally you know have these data sets how do you combine them and how do you com and how do you sequence them across different training stages it's now become table stakes that any large model is generally trained for at least three phases of
05:46data um how do you do that um and can you actually even do continuous uh curricula and things like that which is a lot of what we work on at Dtology.
05:54Um and that ultimately gets you a much better data set out.
05:56All right, so that's a high level of kind of what we need to do.
05:59What can you actually get out of this?
06:01Can this actually really make a massive difference?
06:03Um so we've about half of our team at Datlogy are just researchers.
06:07Um and you know we uh do all of our own research on how we do data curation effectively.
06:12um because this is such a critical part of the uh model building pipeline uh there's very little published here um because there's a very strong disincentive not to share how you do this
06:22um the kind of foundational paper for daty was one I wrote when I was at meta called beyond scaling laws um which was fortunate to get a best paper at nurips a couple years ago which showed that if you choose your data correctly you can actually bend the scaling laws itself you can change the exponent
06:37um and that's because you're now not wasting your time looking at redundant or unnecessary data That was very much the proof of principle for all of Dtology.
06:44And we've since expanded this into many public research releases.
06:47We've shared of various ways to improve models just through data curation.
06:50I'm going to go through a couple of those results now um and show you what we've been able to to to achieve.
06:55Um so first let's talk about vision language models.
06:58How can we improve um VLMs just through data curation alone?
07:03Um so in this case what we did is we took the mammoth data set.
07:06This is a fairly small data set about 25 billion um tokens uh that we use for the purposes of training the fusion adapter layer between your um text uh your text your text model and your vision model.
07:20Um and what you can see this is a scaling plot where we have error on the y-axis as a function of log flops um on the x-axis.
07:27Um so there's about a scale of a thousand from the left uh most part of this plot to the rightmost part.
07:32Um and what you can see is that if you look at the parto frontier defined by many of the best uh public VLMs like the Quen 3 series and 3.5
07:39intern VL etc you can see that a model trained on daties data are able to go well well beyond uh that frontier and I'll note this is actually without any post- training
07:48as well.
07:49Uh so you can get very strong performance across many different benchmarks.
07:52Um so just looking at kind of taking the input data set that gray diamond there that's the input data set we use to do our curation.
07:59Um you can see that just through curation you're able to get around a 14 absolute percentage point improvement um holding everything else constant just through better data alone.
08:08Um and not only that you can also see that we can roughly match the performance of quen 3.54b come with about a percentage of it um while using 145x
08:16less training compute in a world with no with less compute.
08:20How do you do more?
08:21You make data better and now it's as if you had 100 times um the compute.
08:25Um, interestingly I mentioned that reasoning models are be are using tokens at a very high rate as well.
08:30Well, another thing that we found is that data curation can also lead uh to more concise answers um depending on how you represent the data.
08:37So what's plotted here is the mean number of tokens per response um across all the same set of models for the large most part um that we just showed.
08:45Um and you can see that that models trained on data those three blue lines right at the top are all extremely concise.
08:51Um and if we do kind of the same sort of plot um but now on the x-axis instead of log training flops this is now um log flops per response.
08:58Um so this is inference efficiency.
09:00Um you can still see that we go well on the ex by beating that router frontier um and you know roughly get uh similar performance to quen 35 with 35 fewer
09:09uh times fewer flops uh per correct answer.
09:12So data curation can make a huge impact in VLMs.
09:14What about text models?
09:16Um, one of the most uh challenging things about many models is that they work very well on English data, but they don't work well um for non-western use cases in general.
09:24The internet is an extremely biased view of the world that does not represent the world uniformly at all.
09:29And this has major implications for fairness and for the usability of these models across the world.
09:34I don't want to live in a future where only developed countries can access this very effectively.
09:38Um, so how do you do this?
09:40Well, curation again can be a massive lever here.
09:42Um so what I'm plotting here now is a similar plot error on the y- axis as a function of log flops about 100x going from left to right here.
09:50Um this is highlighting multilingual MMLU performance.
09:53You can see we have a parto frontier here defined by many models.
09:56The quen models um the liquid some of the liquid models.
09:59Um the green square is tiny coher's best multilingual model.
10:03You can see again that we're well off the predto frontier with a couple things I really want to highlight.
10:07First off we only use 8% of the data here as multilingual tokens.
10:11Um, so most languages actually only had um at max 6 billion tokens here.
10:15So these are not massive amounts of data uh in the non-English languages that are going in here.
10:20Um you can again see we get the same sort of compute multiplier effect.
10:24We're a little better than Quen 3 um while while having roughly 8x less compute budget um here.
10:29So you can make a huge improvement.
10:31One last thing I want to show here is that if you look at the two blue points on the upper left here um those are both dense llama style models trained for a trillion tokens
10:39um on curated data.
10:40The point on the lower right here I'll come back to but is a model trained by one of our customers rci trendy large that was trained on 17 trillion tokens and is a hypersparse.
10:49Um and what you can see is that if you take the line defined by the two smaller models, um it goes mostly right through that blue star uh which is training large despite it being trained
10:58um with 50x more training compute.
11:01So if you use your data correctly and you simulate token scarcity appropriately, you can also get um very predictable scaling to much larger models and you can derisk a run with 50 or 100 times less compute effectively
11:13um before you actually go and scale up the hero run um and find that maybe it doesn't end up where you want it to be.
11:19Um an interesting scientific result I want to share here is that we also see very strong cross-lingual benefits um from curation.
11:26Um so what's plotted here is the non-English accuracy um where in the the the left bar is showing not curating anything at all and then the right bar just curating the English.
11:36Um we also curate all the non-English data and that leads to much better performance.
11:39But I just want to show this because I think it's quite interesting that curating English data benefits non-English performance.
11:46Um and that's because we see this cross-lingual transfer where the model understands how English relates to say Spanish and and so therefore making it better at English would also make it better at Spanish to some extent and interestingly we see that the the um the
11:58size of magnitude of that transfer um is strongly correlated uh with the similarity between English and that language um and we also see it go the other way although the effects a bit smaller um where curating the non-English data also helps to benefit
12:11English data um performance um okay And let me talk a little bit about synthetic data.
12:17Um we take an approach to synthetic data that we call rephrasing.
12:20This is something that our team pioneered several years ago and has now become um basically table stakes for building a very strong model.
12:26Um in anyway I think we'll hear a lot about the synthetic data in various forms uh throughout the day.
12:30Um but fundamentally with beyond web our goals is how can we define a synthetic data platform that works extremely well um and can be applied to anyone's proprietary doc data and documents.
12:40Fundamentally we want to help folks build models that wouldn't be able to do so otherwise.
12:44And that's what where data quality can make an absolute difference.
12:47Um so to give an example, we might take a document like this about a corporate takeover and we might convert that into one of hundreds of templates.
12:54One of which might be a series of true false questions.
12:57Um by doing this, there's a couple things that are really great.
13:00Number one, because all the information is coming from the document on the left, you don't have any issue with model collapse.
13:05Um and you can actually train models um that are much better than the rephrasing model because the rephrasing model doesn't actually have to teach and understand all the concepts.
13:13All it needs to do is transform the left document into a true false questions accurately which is a much easier task.
13:20And then you do this into many many different formats um throughout data.
13:23This effectively increases diversity and it makes it so you learn a lot more from the highest quality data points.
13:29One thing that's really critical here, what do you rephrase?
13:32All documents are not created equal for rephrasing.
13:34If you just pick random sets of documents to rephrase, you will not get a great result.
13:37Um but if you find the high quality documents and rephrase them, it can make a big difference.
13:41Um and we've seen if you compare this to lots of other public synthetic corpora, we can get much better performance much faster.
13:48Um and critically this can be applied to any any proprietary data um in one of our customers own environments.
13:56All right.
13:56Uh let me spend the last few minutes just quickly talking about a couple um of what we've seen uh the things we've seen with our customers where this can actually drive really gains.
14:03Um so first one of our customers Thompson Reuters um has really focused on post training quite a bit.
14:10um and they have a very sophisticated post- trainining infrastructure with the goal of uh building better legal models on their proprietary high-quality legal data that they have.
14:19So we partnered with them to mid-train a model first on a combination of their data um and uh public data um to then make much better legal reasoning models.
14:28So what do we see?
14:28Well, first off, look at the left here.
14:30Um in this case, we took a an open source model um and then uh just did continued pre-trainer and mid-training on a 100 billion tokens.
14:37What you'll see here is that we see that legal capabilities go up about five percentage points um as measured by uh legal bench after you do this 100 billion mid-training which was less than 1% of the pre-training budget.
14:47Um but you don't get catastrophic forgetting.
14:49You also see the general capabilities go up as well.
14:51I'm sure that many of you have seen or experienced um when you try to adapt a model to a particular domain, you lose general performance.
14:59Not if you use the data correctly.
15:00Um, the key here is actually the majority of the data we showed the model was actually data that was representative of the pre-training distribution.
15:06There was only some of the domain specific data in and that's necessary to prevent the model from losing the capabilities that it had before.
15:14Um, and you can solve this entirely through better data.
15:17Um, but this actually isn't the most exciting part here.
15:19Um, as I mentioned, the TR team had done a lot of work on post- training.
15:23Um, and had a very sophisticated post-training harness.
15:25Well, they then applied that to the mid-train model versus just the default instruction tune model.
15:30And what they found was that the gain deriving from post-training.
15:33So the y- axis here is a delta um as a result of post-training um almost tripled uh when you applied it to the mid-train model versus to the um just default instruction tune model.
15:42And that's because its policy um when it starts is now much more accurate um and it can make much better inference.
15:48Um, so you can even even if you don't change the post- training data at all, showing your model better domain specific data can actually make post- training two to three times more effective out of the box.
15:58Um, which I think really goes to show not only how important data can be in these uh factors, but it also actually goes to show how we really should be thinking about all these stages synergistically
16:07rather than as uh three completely independent stages of pre-training and then I hand it off to somebody else who mid-trains who then I hand off to somebody else um who post trains.
16:17Um okay uh in the last minute or so I just want to quickly talk about um one other one which is rci who I mentioned that large model trendy large um you'll actually hear from verun who was the pre-training lead for this model later today um so look forward to that talk um
16:31but in this case they trained a model this is a fully open um US-made model um uh on 17 trillion tokens that we curated from public uh data sets no proprietary data involved here um and no closed model usage so no asking Claude uh to do this for you Um,
16:47and with that, RC was able to train a model that is competitive with the Open Frontier.
16:51Um, matches uh, GLM5 and and Kimmy on many tasks.
16:55Um, and even outperforms Claude on a couple tasks.
16:57Um, but I think what's most exciting about this is that the RC team um, had not trained a model prior to uh, the middle of last year when they started working with us.
17:05Um and critically in total across salaries, across compute, across R&D, um across everything for this and several other models, they were able to get to a model that's competitive with the open frontier,
17:17um for less than $20 million total.
17:20That includes all the repetitions, that includes compute, that includes everything.
17:23So if you hear this story over and over again, oh, if I want to customize a model, it's going to cost hundreds of millions of dollars.
17:28That's just not true.
17:29you can train an immensely powerful model especially in a narrow domain for high six figures million dollars.
17:35It's very doable to get a model that's extremely performant.
17:37Um this is for a general purpose.
17:39So this is kind of the upper bound of that.
17:41Um and data quality is how you can do that.
17:43All right.
17:44So um last slide here just kind of say um summarizing um focus on kind of what's going to give you the most signal per token.
17:51That's the thing that matters a lot more than more tokens.
17:53that it's almost always better to repeat highquality data than it is to show lowquality data at a certain point um up to a threshold.
18:01Um but focus on kind of how can you get that?
18:03This is data quality remains the single most underleveraged compute multiplier.
18:07If you're sitting in a world where you want to build a model or customize a model and you're limited on compute, how do you get past that?
18:12Invest in data and that's something that can do a tremendous amount of effort.
18:16Um and then finally, this is a frontier research engineering problem.
18:19You need to be able to score and understand data across many different axes.
18:22because that's a frontier research problem.
18:23And then have that scale up to pabytes of data, massive scale, um, and can be very critical and it can lead to tremendous leverage.
18:31Um, and with that, I'll say thank you.
18:33Um, I will note that we're hiring for a bunch of different roles on the left here.
18:36Um, if you're interested in building or customizing your own model and um, would like to to get much better data out of the box or apply it to your own data, um, would love to chat with you.
18:44Um, and uh, thank you very much.