The Messy Reality of Scale: Synthetic Data and Pre-Training — Marah Abdin & Robert McHardy, poolside

Hi everybody. Thanks for coming to our talk. Uh my name is Mara Abden. I'm I'm from Pulsar. I don't know if you're familiar with our models. Uh our data team and today my colleague Robert and I will be talking a little bit about um some of the challenges that we've seen as we scale our models over here. Uh particularly if you haven't heard, we've switched from recently from releasing our models towards enterprise to also releasing towards everybody. We've actually put out two open weight models. They're on Hugging Face a few weeks ago. We have Laguna M and Laguna XS. We also put out a tech report which has a ton of detail if you're interested. Um as you can see here by the Laguna M point one and XS point two, this uh this is this is actually is because we have switched out quite a few things between those two models and so big flavor of this talk is going to be about kind of how did we transition from one to two. And in fact, we've continued to do so and now we actually have a newer version and Robert will give a sneak peek about soon to be released model. Okay. So I will be particular talk about synthetic data part of things. So um there's three things that we did on the data side to kind of resolve some of the issues that we've seen with scale. One is that we implemented auto mixer that basically just gives us a chance to do a cheaper sweep on clusters of our sets before moving on to more expensive experiments. And then we improved we just rethought our sampling of web data for higher recall. Um and then the third one is that we relied a lot more on synthetic data in a few forms which I'm going to. Okay, so before we kind of going into what does that mean and what have we done, etc. One uh why would we kind of It's It's Sometimes it's fair at least to ask why synthetic data and the thing is that uh at least at Pulsar we don't see it as a way to replace organic data. I don't see it so in the current state of the world at least, but uh it is a way to kind of complement it. And the thing is that organic data has a lot in it that is basically kind of implicitly hidden. A lot of things that could teach the model or not very presented in the most optimal way sometimes. And so synthetic data gives us a track to extract some of these features and project them on some new planes. And this is how we get to expose um implicit rationale, implicit planning, implicit structure, and a way for us to fill gaps and regularize not only how we present the tokens, but also how we are teaching the model. Um for access point two in particular, we settled on 13% of the mix. This is only pre-training stages before post-training. Um and since then we've just been continuously generating more data in a bunch of directions. Now we have a six trillion token uh corpus that's continuously growing. Okay. Um so yeah, so uh what kind of what I I I I just said is that we saw some some I guess limitations switching from Luganem to uh point one to our point two models. And so one of those things is that we basically we started on data and this is really not a crazy kind of problem. We very intuitively started from a place on a smaller scale where we were um we were basically focusing on quality versus quantity um maybe a little too much because eventually when we started scaling our models, we had to scale our training budget and with that uh came some limitations because we started hitting repetition uh like non-optimal repetition on some of our high-quality data which saturated the model a little too early. So um we One of the ways we've particularly for this like a token uniqueness problem uh, we relied on which is a very common form of synthetic data rephrasing which you know you just may heard in like Beyond Web for example. It's become pretty trendy these days. Um, and you can see here that uh you know, this is a take this ablation this is an ablation result so take the numbers with a grain of salt but what persists pretty consistently is the diff between using the orange you would be just the seeds with repetition and then the green would be replacing some of those repeated tokens with higher like with the multi-mode reverse or at least yeah, all of them or at least reducing the repetition. Um, and so uh for rephrasing particular we did to the you know, what everyone's doing with the you know, generic kind of multi-mode very scalable pipeline but we also took it a step far and we did two other specialized pipelines one from to go from Morocco to code and text and one to go specifically for scan data. Just because this is a very cheap scalable pipeline so it kind of you kind of have to rely very heavily on the seed and we push a little further on that for the stem documents. Okay. So okay, so if you kind of think of everything is a kind of an modular way you can think of every synthetic data pipeline is composed of the same six components and so you have your seeds, your primary inputs, your metadata, your secondary inputs, your generator and function which can be an agent with tools or one you know, with some prompt with some prompt templates and then some supplementary functions like filters and validators and so on. And really you can compose just about all pipelines from like very simple to very to very expensive pipelines like this. And on on that note and kind of we've we have covered quite a bit of wide scope on the the axis complexity and you kind of can think of it if one end you have like the cheap scalable pipelines that have used smaller models and can get get with it because they're seed heavy examples of phrasing. And then on the other end you have more complex pipelines with a little more orchestration in the workflows. This is reserved when the we're building on something that's worth it. Educational data. But really it's this is how we're not blocked or limited by whatever teacher model can do. And this is how we can be ambitious in our synthetic data. Because the rule of thumb is if task is too hard for your model, then your model will start to fall on its face. Lose correctness, lose diversity. So break down the task, make it simpler. And yeah, I will give some examples of kind of shapes rather than just like something more concrete about how do we use this this modularity. And one shape is the formula writing is just rephrasing. We already talked about this. Multi-stage pipelines and multi-stage workflows. Basically this is what I also just said. You take a step and you break it down into multiple steps. You can aggregate the processing. Slowly build up the the generation. Example of this were if you wanted to generate a novel for example, it's you could generate one chapter at a time, but you could also, you know, take it a little slowly. One first like generate, you know, the setting, the character names, the character styles, the plot, you know, some twists. And then and then from there go into generate the chapters one by one. You will absolutely get a better novel. Um Okay, third is cross-domain porting which really is just like like moving from one mode to another. Example would be changing like translating code. Another example would be something we did which is take our math problems and convert them to code. The last one is multi-turn role. What I mean by that all I mean is that instead of having kind of a very one like very singular or or linear or even non-linear kind of view of things you have more of an iteration. This encapsulates pretty much everything. And and that like an example of that would be multi-turn chats when you have two agents talking to each other or a task evolution pipeline where, you know, you have a judge and an evolver going back and forth for some k amount of time. Um, and so on. Okay. So, lastly, uh, forewarning off to Robert, I do want to kind of just mention that we because of the modularity of the way we think about this, we can implement an infrastructure that's pretty configurable. Um, so this is how we present Hive. Hive is basically a way for us to easily, um, that uh, to easily construct uh, generations where now you have uh, a queue of agents that you define. Each one has, you know, its prompt, its parameters, its model, etc., inputs, outputs. But it also has when you can configure when enters the queue, when it exits, how many have frequency to come in. And then we have orchestration in the middle between agents. And with this orchestrators are really useful because they give you more flexibility and kind of presenting a hierarchy between LLMs that are generating is very uh, it's this is how you police them, basically. But also give them some some form of creativity and dynamically change instructions for the next agent or choose which agent goes next, which agent skipped, and so on. Um, and lastly, you have the supervisor, which basically someone who polices the orchestrator and has more of a global view. Um, cool. Okay, that's uh, that's it for me and Zendaya. I hope you learned something interesting. Handing over to Robert for uh, pre-training stuff. >> All right, thank you, Mara. Um, cuz I will talk a little bit more on the um, actual pre-training side rather than just data. I liked in the previous talk, um, the the speaker made a point that we should treat different data mixes holistically, different training stages. I want to make the same point that we should treat uh, data and implementation of your training code base, uh, correctness of it, and so on, also holistically. If you've got data that sucks, you can't train a good model. If you've got a training code base that sucks, you also can't. Um, so I specifically focus on architecture work on and distributed training and so on. Um and the way we we look at things in my team is uh we don't trust anything. There's so many things that can go wrong when you scale models to billions of parameters to hundreds of billions of parameters um training on thousands of GPUs and so on. And I want to show you some of the learnings that we got from training Laguna m.1. Um And yeah, some of the surprising things that happen at scale. Um so one thing we do is we've got these um model replica hashtags. So essentially when we train a model we've got multiple replicas of the same model, right? Um distributed data parallel. And uh we know there's an invariant, the weight should always be the same across all of these replicas. That's something you can verify, right? We you can calculate a hash over the weights and you know that should always be the same across all replicas. So we do that in training and periodically compare them. Um if all of these hashes are identical, then we know we can continue training. If they're not identical, we know something has gone seriously wrong uh because that should never happen. And we crash the training. And I'll give you some examples now of things that we've not shared before publicly like this. Um so I hope they're interesting. Um so the first example here of uh that happens at scale are broken GPUs. On the left-hand side we've got two loss curves uh and on the right-hand side the corresponding uh gradient norms um that we observed during training. And you can see that these loss curves look quite different, right? Like the purple one has got quite some bumps, looks a bit spiky. The gradient norms are huge for that run. And there's actually no difference uh in terms of model configuration, training data, training implementation um between these runs. They're exactly the same run. Just in one of them we were got unlucky and we had a broken GPU included. That broken GPU caused silent data corruption and um therefore made the training behave the way it did. Uh and that is one of those cases that you can catch with these hashtags because you know this computation should be the same across all replicas, but it wasn't. Um Which brings me to the next uh instance of that happens at scale. Uh in this case exploding gradients. Um again, we're looking at two different loss curves and the corresponding gradient norm curves. Um, the purple run is our initial training run for Laguna M 1. We're a bit further into training here around 50,000 steps or so. And you can see it stops converging, right? Like it just flattens out. And the reason here was that during training the uh activations grew and grew um right before the um LM head, the unembedding. And um we have to perform some sort of accumulation here uh because we use tensor parallel for the uh unembedding. And that accumulation um was performed in BF16 by default. And because of the growing scale that we observed in the activations, um there wasn't enough um numerical precision available anymore um to do this accurately. And hence the model just couldn't learn anymore. And this is also very uh dramatic point for this to happen because it from there on it really like back propagates into the full model trunk. Um, the orange curve is essentially just adding a fix on that. So, we took the checkpoint from the purple curve. We moved that accumulation into FP32 and from there on the model started converging again. Um, the gradient norm, as you can see, actually started decreasing. Um, before then we had an increasing trend. And um this is also something you can only observe at scale. And that will break your model if you if you're not careful about it. So, as Mara said, we took all of these insights on data, on um numerics, and so on, and we turned them from M 1 into XGen-2. That's why we say it's a new generation model. Um, this included uh increasing diversity, reducing data repetitions, um all of these numerical things I just mentioned, and just adding more observability and checks on that side, as well as generally optimizing the training in the architecture. And XGen-2, if you look at it, it's open weights, right? So, you can download uh you can download and use it for free. Um, it's one of the most competitive models for its size and for coding specifically. That's what we focus on Agentic coding. Um, so we're pretty happy with it. However, you can say that um the model with 33 billion parameters is pretty small. Like you wouldn't probably observe any issues anyways training it. So, what was important to us was to scale this, right? And this is where Laguna S comes in. This model is not public yet, so this is a preview. Um as I said, we treated XS as a test bed in a sense, and uh with Laguna S we scaled this to a model that's 118 billion uh total parameters and 8B active parameters. Again, we trained it on 30 trillion tokens on 4,000 GPUs. So, the scale was sufficient to not only test whether uh all the improvements we made on the data and architectural side had hold, but also if any of these numerical issues come up again. And um of course, something happened. Um in this case, it doesn't actually have anything to do with scale, so it was just unfortunate. Um in this case, we had a race condition because we added FP8 training uh based on Deep Chem FP8 kernels that are also like open source. Um we noticed these because we hit illegal memory accesses as well as NaNs in the gradients, uh which after a while of debugging we traced back to those kernels. Um there's also an unobservable unobservable effect um that you wouldn't know about if you don't know that there's an issue. Uh in our case, we noticed about 0.5% of the gradient gets silently corrupted, essentially replaced by random values. Um we do have a fix available that's in a PR right now. Uh it's not been merged to Deep Chem yet, but it's public on that QR code if anyone is interested. Um and it's also an interesting point because it's a blind check and a blind spot in the hash checks. In real training runs, you don't have any redundancy where you have the same model weights and the same data, so you can never check if forward and backward actually behave the same across different model replicas. So, you can also never check if there's a race condition in that. Uh that's something that we're working on right now uh to essentially have a hash checker um that can also do that as a at least as a dry run. Um and I want to end on um some early results from uh this new model um and demonstrating how it performs against some open weight models and also against our previous models. So, first I want to caveat this with these are base model evals, right? They are partly indicative of how the final model will look, but also not perfectly, right? Um there's still post training happening. Uh not all of these will translate one-to-one to the final model. Um but if we look at them uh specifically on the coding part of of the E valves, so for instance multiple E love code bench, big code bench, uh Laguna S is not only stronger than access to two, which is our previous uh smaller model that performed very well, but also than the much larger M dot one. Um and it's also much better than GLM 4.5 air, which is admittedly a bit older, and then we turn 360, which is quite recent, and then deep seek V4 flash max, which is quite recent and a fair bit larger. Um we can see it's competitive on big bench hard for instance. It doesn't achieve the the top E valve results compared to these models, but it's quite close. Um we also see it's quite close on E valve plus, and uh quite importantly for us, it does very well on sweep bench agent less multilingual, um which we use to sort of proxy agentic performance during pre-training. Um and in that case it performs much better than all the other all other models we tested here. Um I also want to point out that of course there are like it's not the strongest model in the world, right? Like for instance MMLU pro knowledge benchmark is something we don't care about that much compared to coding because we want to build the strongest agentic coding models. Um so here like compared to Nema tron and deep seek uh we have to say that they perform much better, and this mainly comes down due to data, right? It's a data gap um that we could plug if we wanted to. Um but I think the the point is all of the things we we found before were included in the recipe. The recipe held, it scaled, and we will continue scaling it from here. So this model will also be available sometime in the future relatively soon. Um again open weights, so all of you can download it and use it for free. And um with that I want to thank everyone for attending our talk. Um I added also a link to our careers page and our Twitter page if you want to uh check it out. And yeah, thank you very much. >> [music] >> Hey.