How to train your own Large Multimodal Model — with Hugo Laurençon & Leo Tronchon of HuggingFace M4

31:15and the pile is documents that were taken from good quality sources like Wikipedia, Archive and so on. It's not something that you can really scale. And however, we also noted that we obtained better perplexity scores than the ones from CIFAR, the BitDataset or OSCAR. Based on your own measurements, right? Because obviously the multimodal CIFAR people would not share it. Yeah, just based on the text. So yeah, I think this is how we computed perplexity and how we assess the quality of the dataset. But you could also run a multimodal model on this and get the perplexity from it. It would not be measuring the quality of the text, but also the alignment would come into account.

32:09Alignment of? Of image and text, because it would be easier for the model if the text is very heavily related to the image to get the next token. And then one more question, just about the whole process. How long does it take to make Obelix? So we spend a good time at the very beginning of the project just simply to iterate on the pipeline like how we collect HTML codes, how we clean them. We had to go through all of the HTML tags, which they were important. So this is an engineering part you have to be really it sounds very boring but very important it is boring and important but that's how you get the good data yeah but this is also why people don't do it the industry has not converged on a shared set of tools that everybody uses for this you're just parsing raw tags yourself yeah we did that because we found it was better we parsed raw HTML codes.

33:14So we had to clean the DOM tree, select the good HTML nodes, correctly extract the text, the images, clean, they duplicate. So as I said, there was just a good amount of time at the very beginning of the project just finding the pipeline. So maybe one month, but we were like one or two on this and it was really exploratory. and then for actually making the dataset, download all the images, do all the processing scripts and so on, I think it took us like up to two months. Yeah, but then there's also like iterations through the project where we're like, we think we should do filtering on this on top of what we were already doing.

34:01So we improve on the dataset. Something I would think makes sense for the industry is kind of an open-source set of deduplication rules because everyone seems to be reinventing this from scratch every time. For deduplication, you have to do it all the time from scratch because it depends on your original set of documents. Everyone draws from Common Crawl. Yes, but they are not from the same dumps. But that's true. Someone should take all the Common Crawl documents. It's been done recently. I don't even need the same exact rules. I just need to be like, oh, these smart guys thought about that. I should include that, right?

34:43That's very simple. Like, you know, if you have 100 rules, someone else has like 80 rules, maybe they have some that you don't have. Exactly, exactly. I think we did a little bit of this, because there's a bit of literature right around. Even the ones that we designed before for the dataset of Blue, we took a lot of trade. I'm sure you used a bunch of that. Obelix is big, but the datasets that come for NLP right now are also huge. Did you see the together one from yesterday? Yes. Impressive. 30 trillion. So they said they had a raw dataset of 100 trillion, and they got a clean, high-quality dataset of 30 trillion, which means they kept 30 % of Common Crawl, which is still too high.

35:28Yeah. so I think the idea of the project and I think I agree with this idea is that everyone can set the thresholds for the filters as they so for each document they computed they computed threshold filters, or no actually filter values for a set of rules and then you can decide whether you want to keep the document or that and then so you define your own goals so i think they remove like as you said the 70 percent of the of the data set that is really like you can't you can't do anything with it and then for the remaining part they let people decide but of course if you actually want to train something on it it will be much smaller i guess because you will remove a lot of things but data is super important and yeah my point is also that it's still very early in multimodal we're seeing now in NLP those 30 trillion data sets.

36:27It's really important. Mistral, the first thing they said when they raised their goals is how impressed they were that they were capable of doing the data pipeline in three months. They didn't talk about training the model or not. It's just like they knew what they were doing for this. It's straightforward. But creating the whole data pipeline, this is what took them a lot of time. So I think that was one thing that struck me when they made the R &D. Is it confirmed that they have 8 trillion tokens? I don't know. They won't say. Could be this. It could be more. Could be more, I think. Given we have now an open source data set of 30 trillion, I wouldn't be surprised that they have more.

37:08I just keep coming up with questions. One more thing on data sets. Did you read the GPT-4 Vision system card that they put out? They put out this paper describing a little bit of their process. Something like 95 % of the labels for GPT-4 vision was augmented by GPT-4 itself. And I was just curious, how much room is there for open source augmented datasets? I think there's a lot of room. I think there's a lot of room. And I think synthetic data works a lot, like works great, particularly in multimodality. Recently, the recent papers have been using, for example, Leon Coco instead of Leon. This is just captions on Leon created from Blip.

37:53So it's not even super extraordinary, but it does bring more performance with a lot less example because the alignment is more straightforward, I guess. And we've been observing this recently because we've been using it in our recent experiments. I think the potential for synthetic data in multimodality is very big and very underutilized right now. Even DALI 3, they said that they used heavily synthetic captions to train their model. Also over multimodal, foundation multimodal models, like Reap, they train on synthetic captions. Yeah, actually, I think I was referencing the DALI 3 paper, not the GPT-4 vision paper.

38:36Yeah, because they didn't actually put out a paper for GPT-4 vision. Cool. And then, so you created Oblix and then you trained IDFX. on top of it. In addition, so we train IDFX on ObedX, but also on other datasets, like Lion and Public Multimodal datasets, which is just sets of datasets that were open-sourced at the moment. Like conceptual captions. And yeah, could you take us through just the IDFX process? You created a smaller version, and then you scaled up to the full Flamingo 80 billion size. Actually, it was the other way around. Really? Well, we tested that it worked at smaller scale, of course.

39:23But we did not train fully our small model before doing the big one. We fully trained our big model. That's unusual. Yes. Yes, they were training pretty much at the same time. But we needed to launch the big model before also in terms of timing because it took a long time to train. So it was more like managing computer resources. But through the whole journey, it was a bit longer than just this moment when we trained to read a fixed model because we started with the objective of matching Flamingo's performance. But the open source models that are out there, they're just not good enough. So we have OPT, we have GPT Neo, but it's just so below Chinchilla that it's almost impossible to reach the performance.

40:15Suddenly when LAMA came out, that it started making sense and that we started matching the performance. And from there, we were able to train the big IDFX models. A long journey, I think because we had a lot of things to learn because we had quite a lot of instabilities. We shared a blog post about... The checkpointing every 250 steps. Oh yeah, we were checkpointing every 250 steps, but we had to restart a few times. Was that the instability you're talking about, or this is just something else? That was the final training, where we still had the instabilities, but a lot less. Before this, we were struggling to train the model at all.

40:52What saved us at that moment was the query key layer norms. I don't know if you've heard about this, but basically, there's a paper from Google where they scaled up the vision transformer to 22 billion parameters, and they needed this trick to keep the stability of the training. Without this, if I remember well the mechanism, you would get something along a hard attention. And once you get this, the model would get very stable. So you needed to normalize the queries and keys to basically avoid this. Anyway, so when we did this, we were able to train further. And that was really useful, for sure.

41:32So query key layer norms, if you want stability. It sounds like a trick that is repeatedly applied whenever you have instabilities. You just throw a Dear Norm or Softmax. Yeah, you can. Because usually it's parameters that explode, become too big, so you need some sort of regularization on them. At first you have to inspect which parameter explodes first, and then you put regularization on them. but it's really tricky to see. When you go in the gradients and in the activations and you try to see where it blows up, when it blows up and why, everything is interlinked. It's really hard to pinpoint one particular layer.

42:24So it was a tough one to crack, but very interesting. It's also hard because you never know exactly. Once, when you're in the process, you don't know where the instability comes from. It could come from bad data. It could come from a bug. It could come from the size of the model, the learning rate that's too high, the warm-up that's not. There's a lot of hyperparameters and potential bugs that can come into account, and the debugging is very, very tough. So do you have a checklist of what do you look at when you see loss explode or whatever? You see the loss explode, So you look in the activations, possibly the gradients, to see where it blows up.

43:07Across all your parameters. Yeah, you find a way to aggregate this. Otherwise it's tough. If all previous solutions fail, because this is the hard part. This is the hard part, but it tells you a little bit where it's happening. So that gives you where it's happening-ish on the model. And so what's to blame? And then you have a wide range of things to blame. but less than before. So you can look for a bug. You can look for, for example, normalizing some layers and you can look into the data to see if you have like very bad data that's an impact. That's what I would look at first, right? Exactly.

43:48I would look at that too. Is this what weights and biases would do for you or is there one integrated solution that kind of... You would wish? Yeah. I don't know. to see the activations and the go for activations to then like inspect data sets and then look at you know weights and biases would get you like the parameters it's just logging it's just logging like you have tons of them you can't really do much with it so we had a tool where we would like log them periodically we had a script that would aggregate them and display them for us in a nice way so that like interpretable way so that we could try out and see what mechanisms were and could be impacting the instabilities.

44:37But yeah, it was a very, very interesting journey for sure. Yeah, and you published knowledge sharing documents and a memo. Yes. There's some interesting detail there but obviously not everything. Anything you want to highlight just first for listeners on that one? Obviously I can send them the link. It's very high. Just any other big discoveries on learning? You talked about the query key layer norm. Query key layer norm was the big aha moment. But this and there's also the... This is one we fixed afterwards. But there's like in the mask, in the image mask, there was like a little information leak in that instead of attending to all the images, instead of attending to none of the images, sorry, for a few tokens, very few of them, it would attend to all of the images.

45:35Like basically you tell them, you go in the attention and you have this mask, this mask that's like, don't attend to anything. But in effect, it's like attend to one over N, right? And it's not like, it doesn't prevent you from training, but it doesn't help. and it's better to fix it for training for sure. Yeah, sometimes you really have to go through all your code base. Because you're doing gradient descent or something that has no information. Yeah, no, it's tricky because this, like for example, this, it would not have an impact if you only train on web documents because the documents are long.

46:16But it would have an impact if you train on image text pairs and you pack them together because you're attending to images in a document that has nothing to do with it. But you can still train with it. You can still get a very good model out of it. It's just a lot, like it's more painful and you don't get, like I think we can get better performance definitely without the bugger. Yeah, interesting. You mentioned, you know, just in terms of like the baseline foundation models that you had, LAMO has been breakthrough on the language model side. But then you also, you didn't talk as much about the vision quarter side of things.

46:50I actually had a question from Joseph from the RoboFlow episode where he talks about, where you mentioned that the larger the clip, the better results. But in your final memo, you actually went for a smaller version of clip. So Eva clip versus Lion2B. Does this ring a bell? Like basically it was kind of like unintuitive, like the clip choice there. Yeah, so Lion2B is the data set on which our clip-based model was trained on. And our clip model was indeed like 400 or 600 million parameters. And it's true that the EvaClip one is of, the biggest EvaClip one is of 5 billion parameters. So definitely at the beginning of the training, we saw a big boost using EvaClip.

47:38However, at that time, we still had instabilities to train this big Eva clip. We are not sure exactly why. However, we fixed it. And now that we, like for the next, for the V2 version of Edefix, now that we can train longer, we actually saw a boost by using Eva clip instead of the previous clip, small clip that we had. However, we now think that Evaclip is under-trained. So it means that even if it's really big, you can obtain the same performance with smaller models. So there is this Ciglip model that I mentioned just earlier, by Google, that is much smaller, 400 million parameters. And that is more efficient.

48:34You're choosing that as your base? Yeah, we did an ablation, EvacLip versus SigLip. And actually, SigLip was a bit slightly better. But the thing is that it's much faster for the inference and also for the training because there's fewer parameters. One thing is also that SigLip is a higher resolution. So that has an impact for CR tasks. 380, 84 instead of 2.24. Yes. That's great. Yeah, maybe we should talk about EFX2. so we're going to time this podcast release with whatever you guys are releasing it I actually had no idea you were working on a v2 I just came in wanting to talk about your old work but obviously you're still doing active research for edfix v2 or v1.5 whatever it would be called the major axis we wanted to improve on was the image resolution the base model We wanted a better base model and a smaller one.

49:35Sorry, pre-trained language model. So we've been using the Mistral one. We wanted to iterate a little bit also on the data to have better filters on Obedix, better synthetic data for the image taxpayers. So essentially iteration on the data by replacing original lion pairs by synthetic captions to have a stronger alignment, cleaning a bit obelics on the perplexity. So it's not removing too much, but like potential bad data. Also, yeah, using just better pre-trained models. So for the bad bones, we use like a better clip, we use also Mistral that is better than Lama 1. we are right now changing also the modeling so moving away from the flamingo architecture to something that has fewer parameters instead of incorporating the vision components directly into your llm by breaking and adding cross attentions at each layer or every end layer you can instead take your vision encoder, take the embeddings out of it, make them through, feed them to linear layers and feed them directly to the language model.

51:05And this works quite well. And this contains fewer parameters and it's much easier to train. So we are currently trained to do this. But what we can say is that without this new modeling, just by iterating on the better data and better pre-trained models. We are now matching the Flamingo 8cb performance with 9b model. So this is already a big improvement compared to our first version, without even touching the modeling part. I think also one of the big improvements with the new edafix is the licensing. The issue with edafix was it's based on LAMA, and the license is not commercial. So now with a model that's based on Mistral and Siglib, most probably, this is a lot better for anyone that wants to use the model commercially.

52:04The model will also be smaller, so a lot better for inference. And hopefully we can beat the performance of VDFIX ATB. That would be really good. So you're only producing a 9B? Not exactly 9B because we're taking off parameters. It should be about 7.5B. Right now, the focus was really better data, smaller open source models, but better. And a resolution, improving on this as well. That was the focus. You mentioned in our prep as well that there are some topics that you're paying particular attention to, like hallucinations. Maybe you could talk about the topics that you are finding are particular areas of concern with multimodal models.

52:53So we've been using hallucination at the beginning as a broad term. And we realized that it was better to categorize it a little bit more specifically to some categories that were more targeted. so for example like there's the object attributes where you would have like a small attribute like let's say the hand of a person that is that's a certain color and the model would be like it's it's yellow when it's red so that would be that would be one there's like objects that are not there but the model thinks are there like when the model is trying to reason with different elements in the picture but it gets it wrong so you get like comparisons uh kind of hallucination counting oh my god yeah you have the environment so it would talk about like the object or the person in the picture but it would get the whole environment behind wrong and a few others that I don't have in mind right now but basically categorizing those I think is important because it helps you target the type of data that's missing or the type of fine-tuning that you should do after what's that's missing.

54:03And so this is going to be very useful for us in building future datasets. I'm not sure if we'll be able to incorporate all those datasets we've been thinking about for the v1.5, but we'll definitely do this for the iteration after. And the goal here is really to get to something that's on par or better than what is currently done in closed source. like GPT-4v doesn't hallucinate as much on those topics. And you want the open source models to at least match this. But it's still an open research topic because those models still do this. If you push them a little bit, if you ask some specific questions, they will hallucinate things in the picture.

54:47Yeah, you mean including GPT-4v? Yeah, yeah. I have tried this, by the way. I tried to use it to interpret like a menu and it would just make up menu items and make up prices. Something really interesting is that GPT-4, even much better and bigger models like GPT-4, shows the same failure cases as our model. It means that the way we are training the models now is not ideal. And especially, for example, counting, you take this task. even if you train on web documents or image tech spares you will never really find this task in your training data you will always have a picture and maybe a caption of one apol or two computers that you will never get to like 10 15 20 it doesn't doesn't really happen so you you don't learn this ability just by training on on image tech spares or web documents so what you have to do is to create your own data sets that target specific tasks.

55:58Like for example, create one specifically for OCR, create one specifically for counting, create one specifically, I don't know, to challenge the model on hallucination, some types of hallucination. A sort of flan, but for multimodal. And I think in multimodality, those hallucinations are even more unforgiving than what they are for NLP in that when the language model is making up facts, you can think that it got a little bit wrong or it's making up a story. But when a multimodal model tells you there is a teapot in this picture and there is none, it's very obvious. It's very obvious very quickly.

56:39If you're the user of this model and you receive this, you are having a hard time trusting it for anything. So this is one of the big, big things to tackle. And for hallucinations in NLP, a fine tuning that has helped a lot is the reinforcement learning with human feedback or AI feedback. And this is still very early in multimodal. We're not sure exactly how much we can improve with those types of datasets, but it's likely that it would help the model understand uncertainty to a more fine-grained level. And so improve on all those types of hallucinations, ultimately. And you said, you know, Prep as well, that you're measuring this against other closed-source models like BARD.

57:22Like, basically, do you have your own internal benchmarks that you're running? No, no, this is mostly qualitative. It's more like, we know it happens in ours because we've played with it. Does it happen with theirs? And yes, it does. The thing is for ours, the evaluation data sets that we use or that we've used for EDFIX so far, they don't measure hallucinations that much. It's not targeted for this. New data sets that are coming for evaluations are exciting. And I think we'll go in this direction and will help us evaluate hallucinations and different types of hallucinations more accurately. but right now at least for edfix even with those evaluation benchmarks and even if this helps like obviously if you get very bad scores your model is is hallucinating like could be hallucinating things it's not targeted to this and we were flying blind a little bit on this uh on this topic but there are definitely uh benchmarks who evaluate hallucinations currently so i think about sugar crepe for example so you just have an image and two captions of it and the model has to select the the best one or vinogran i think so the so actually the captions are made such such that it's it's tricky we know grants like the the common sense nlpe one yes this one is really hard but you can have one with uh with uh images yeah i think this this can be a great way to to evaluate hallucinations of the models.

58:55However, it's not really commonly used. So right now on the recent models, like PANI, for example, they report their numbers on classic evaluation benchmarks. So I think it needs to be more widely adopted, this kind of new benchmarks. I think the last topic that we prepped was just overall, why is it important for there to be OSS multimodal models? What can people use them for? Where can they be useful? My understanding of this is that ultimately you want foundational models that understand the world similarly to the way we do. And multimodal models understand visual data a lot better than language model, obviously.

59:44But it means they provide a better foundational backbone for other tasks in general. I think in the future, when you want to pre-train a foundational model, you will want to have it multi-model. So this is why it's important now and it will be important in the future. And this is where everybody's going because training on text data, it gets you really far, as we've seen, but it only gets you so far. And it can only adapt on a subset of tasks that we do every day. And you need models to be able to understand vision if you want them to be helpful for, I don't know, robotics, for example, and a bunch of other use cases.

1:00:29I think, for example, even for medicine, you can have tons of applications with a vision input. So you just take an image of whatever, ask... Answer. Yeah, yeah, no, honestly you can ask for help with your model or even in your everyday's life how to build this table. You just take a picture of what you have and you ask the model. You just take a picture of even something handwritten, ask things about it, like solve the exercise shown in this picture. A lot of things that you can't do with text-only models that you can do with multimodal models. So definitely it's a big step forward. It's still a bit immature.

1:01:22And we have seen that because of all the hallucinations we mentioned and because it's still early. We're still trying to unlock some abilities for these models. Yeah, we believe in less than two years, we will only have a multimodal models in the future. Like they will overtake everything. Yes. Also, also right now it's a lot of image text models because it does most important things that you want without requiring too much compute. I think it's more computer fusion than if you would put video in it. Ultimately, I think we'll need video to be incorporated in the pre-training as well. I'm quite excited about what it could look like.

1:02:06probably the ones that are more advanced on like the ones that are the most advanced on it are probably the self-driving cars they they're doing like very very interesting work there but it's all closed source oh no i think was it coma ai that released uh gaia or something like this but yeah basically uh like video generation so i think that's very interesting is video isn't just that like a series of frames of still images what is so different about that it's hard because you need to you need to encode each frame or at least select the say how many frames you want to integrate per video also one big change is like the length of the video that can be really arbitrary while if you have an image you have always the same number of tokens associated to it so all of this plus the fact that having a big video data set is really challenging because you can't, for example, scrap YouTube like that.

1:03:11It's harder. So yeah, all of this make it hard to build a model with videos. It's also extremely heavy. Already when you go from a text data set to an image text data set, like how much it weighs is so different. It means you have to get the data pipeline. It means the dataset preprocessing takes more time. If you go with video, you do this 10x, if not more. For a transformer, it takes tokens as inputs, sequence of tokens. And for text tokens, you get one token for a subset of text or word-ish. for an image depending on the resolution you want you can get from like you can get a lot of tokens for a single image for a 224 image you'd get I think 256 tokens per image if you want to increase the resolution it goes up fast for Siglip I think it's upwards 700 tokens for 384 resolution so imagine this and then you make it a video so you have like it gets really tricky and you have to create a bottleneck, right?

1:04:28You have to pack those images together and put a thumbrockets. That's what's possible with Flamingo. It also means you have this bottleneck there. It's a lot of work and right now, I think the trade-off is not yet worth it because we have a lot more to do with just image text, but it will be at some point, I'm pretty sure. That's all the questions I had. Any final call to action? You want people to go somewhere, check out the models, check out the papers? Yeah, essentially. So we are excited for the second version, which will be released about the time of NERIPS. And yeah, we think that Obelix can be useful in the next years, even if we continue updating the model.

1:05:14I think the dataset will remain pretty much the same. And we will keep iterating on better modeling and better data. and if you want to follow our work we're at the Hugging Face M4 organization let's talk about naming what is M4 and then maybe we should mention Obelix and I've heard it fix I think so at the beginning of the project it's like massive multi-model, multilingual multitask model but then we drop the multilingual and but it's still massive it's still, well not that massive anymore since we are trained to go smaller. Go smaller. Multitasks for sure. I mean, the data sets are still massive, so that's good.

1:06:00It should be still multilingual, If you train it on Common Core? Train it on Common Core. It's just not targeted for a lot of different languages. Like, we didn't take care of having multiple languages, but there's probably, like, there's definitely other languages in there. Yeah, the naming, like, it changed through the project, but we kept it. The idea was also to have four modalities, so it was fitting at the beginning. At the beginning, we wanted audio, video, text, image. We thought that audio and video, video is just a lot of compute for not that much results right now. And from the moment we decided to reproduce Flamingo, we dropped video and audio, which was not immediately.

1:06:42This started as a different project. What's wrong with audio? It's still very, very heavy and not necessary right away because you can take the text tokens, plug it into an audio model after, and it will just spit out the audio. Ultimately, I think it will be good because it's different data as well. Like when people talk, it's not the same as when people write. So I feel there's a lot of interesting data to have with audio. Again, not worth the compute right now, I think. And also you would want this model to be integrating everything with a given size and then be applicable to whatever. You give it to your robot and you're like, what?

1:07:25Yeah. And to close the loop there, Oblix was an asterisk reference that you made into a backonym or whatever. Into an acronym. Yeah, it was really hard to find the acronym. I want to stress this out. I thought about it for days. We were brainstorming it for a few days. We had to cheat at the end a little bit. we took the S of cross-attentions and added it to... Yeah, cross-attentions. The problem is the F of Flamingo, right? For AD6. But we are moving away from the Flamingo architecture. For the next iteration, it wouldn't be really accurate, but whatever. But yeah, the acronym works only for the first version because if I remember well, it's image-aware decoder enhanced a la Flamingo with interleaved cross-attentions.

1:08:20Yes. Yeah, it's a very impressive project. And we were talking about it even before the GPT-4 vision rollout that this is the most impressive sort of open source reproduction. And it's amazing that you're still continuing to work on it. I think a lot of room left in Oblix to keep mining those rocks. and there's a lot of learnings as well in multimodality. I think it's a very important area of research. So thank you. Yeah, thank you.

1:08:55Hello, hello. This is Swix coming in from the editing room in 2024. If you're listening in this far, you're definitely a true fan. Thanks so much. And I hope you enjoyed that conversation as much as I enjoyed recording it. We recorded that conversation on Halloween in 2023 in the hopes that we would be able to release it at NeurIPS or around NeurIPS with IDFX V2. V2 was supposed to be updated with a new base model, which is going to be Mistral, and a bunch of other dataset updates. And a couple of things have happened, but you know what? Hey, let's just get Leo back on to talk about it. Hello from 2024.

1:09:34It's Leo. Just wanted to add a few things since the recording was done a while ago now. We haven't trained the model yet. because we're trading a lot more on data. So we're making progress on OCR, on image-to-code capabilities, and we want to be a lot more thorough in the image text pairs data set that we use. I don't know if you've heard, but the Lay-in data set has had an issue with CSAM images. And so we want to get ahead of this and fix the problem before we start the training. But we will start training soon. In the meantime, we released a website. It's an image to HTML dataset. The idea behind this dataset was to show that we could create a very useful synthetic dataset at scale with open source models.

1:10:25Hugo spearheaded the effort, and so he used Mistral and the DeepSeq coder model to generate the pairs of screenshots and HTML code. and then we fine-tune an early version of IDFX 2.0 on the dataset to have a demo. You can find the model on the hub, you can find the demo on the hub, and you can find the dataset on the hub. So definitely go there and check it out. I think some people have already started training their own model on it. But it's been very well received, and so we think we're going to do a lot more small releases like websites in the future. Basically switching from releasing everything in one package with the model trained to releasing the data sets, architecture and training insights that we get along the way and then releasing the model.

1:11:21So stay tuned. I hope you will enjoy the podcast. And thanks, Sean, for having us.

1:11:39Thank you.