From Scratch to SOTA: Training a 3B State-Space Vision Model — Krishna Prasad Srinivasan, Sarvam

AI Engineer · 21 min · 148 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.

00:16All right.
00:17Um, hi everyone.
00:18Good afternoon.
00:18I'm Krishna.
00:19Uh, and I'm a general manager at Servin.
00:23Um today I'm going to tell you how a three billion parameter model uh small enough to run on a single GPU gets state-of-the-art uh in document AI and beats
00:36models 100 times larger in size.
00:41It is also a slightly unusual model in two ways.
00:44Firstly u the language model used is not a standard transformer.
00:49Um and the entire model building from data to training to compute uh was done end to end in India and for English and for 22 official Indian languages uh which I think uh are personally
01:05uh uh one of the hardest document intelligence problems uh in the world right now.
01:09Um this is how we went from scratch to SOTA.
01:14Who are we?
01:15Uh we are serv uh a sovereign foundation model company uh in India.
01:20We work across different modalities voice, text and vision.
01:24Uh in voice we have speech to text and texttospech models.
01:29In text we have uh a 30 billion and 100 billion parameter uh uh models and in vision we have a document intelligence model which is what we are going to talk about today.
01:41Um, India is largely missing from the machine readable world.
01:46Um, according to a published language distribution study, well under 1% of the common crawl uh corpus on which frontier models are trained uh have uh um Indian language representation.
02:01Now in several forums you may have heard Frontier Labs say that um India is their um India is one of their larger uh uh and fastly growing markets and to that end
02:14um India is still missing uh in terms of data from frontier models that uh um are being trained every day here.
02:22Now why is that?
02:23Uh primarily not because there is uh a lack of data or the knowledge does not exist.
02:29It is because the data has never been digitized and we are solving that problem at server.
02:37Um why is Indic document intelligence hard?
02:41Uh firstly the goal is that of knowledge extraction not just purely text extraction.
02:47Uh extracting pure text without logical coherence is meaningless.
02:52And secondly in index scripts the shape of uh the words and the language you see are very different from what a machine sees.
03:03that is Indian languages have a complex uh set of uni codes fused together to get right the language model needs to be strong at all 22 languages.
03:14Third, majority of Indian languages can be considered low resource due to which there is a lot of uh um lack in data today to train these models.
03:26Our answer to this is server vision India's first sovereign vision language model built from scratch.
03:33um a three billion parameter model with state space architecture data compute and training all of which done in India.
03:42When we started the initial work in late 2025 most VLMs uh in the OCR space were monolith VLMs.
03:51uh they performed page level OCR and we at the time took a contrarian bet to focus on block level OCR instead adding a document hardness around the model.
04:03Um a lot of models since then uh released uh uh recently in 2026 have converged onto the same uh hardness plus small model paradigm for OCR which shows a lot of value in the direction we uh set off.
04:18Serverom vision in particular uh has two hardness modules one for layout and one for reading hardness and a state space model uh VLM for block level OCR.
04:31Again um today uh why do we use state space and not a transformer?
04:37Um most OCR models today uh like general VLMs, open source VLMs like Quen, Gemma etc. are all transformers based.
04:46We took an alternate approach uh using an SSM.
04:50Why?
04:51Both are fundamentally sequenced models transformers and SSMs but with very different underlying mechanics.
04:58A transformer has every token look at every other token where the compute grows with square of the sequence length.
05:08Uh that is the L cross L interactions and the memory goes as the sequence grows as well.
05:14Now on the other hand, SSMs have a single state.
05:18They maintain a single state through the sequence updating it token by token.
05:23Now as compute grows only linearly, the memory stays constant for SSMS.
05:29Now why is this the right uh architecture for OCR?
05:34It really depends on the uh uh trade-offs here uh for long documents in particular which can lead up to uh uh 5 to 10,000 uh visual tokens per page
05:44and quadratic complexity of you know uh compute and memory becomes very expensive for inference.
05:52On the other hand, um doing block level OCR with somewhat of a lossy recall using SSMs can be justifi justified to uh um avoid high compute costs imposed by transformers.
06:08So how do you actually train this?
06:11We built a staged curriculum four stages each uh building on the last um stage one is texton pre-training uh 13 trillion tokens across English, Indian language, text, math and code.
06:26This builds the three billion parameters language backbone and a strong language prior is what lets the model resolve a smudged or an ambiguous text from the image.
06:39the same way you can read a half blurred word because you know which word is supposed to uh u be at the right place.
06:47So we focused on building the competence of the language model first uh before the model ever even sees a pixel.
06:55Stage two is where we performed continual pre-training uh on 300 million image text pairs.
07:02This teaches the language model general vision capabilities and teaches how to see how to interpret pixels and so on.
07:10Then comes stage three where we performed uh supervised fine-tuning on 100 uh uh million uh OCR uh samples.
07:18Now the um stage three is primarily focused on letting a general VM model become strong at OCR.
07:27What that includes is having diverse data across all 22 languages and English and also incorporating all kinds of uh uh document components such as tables and uh um equations and handwritten documents and so forth.
07:45Now and stage four is re reinforcement learning which helps push past the ceiling of whatever uh supervised fine-tuning is able to achieve.
07:55Here you will see this is a standard recipe.
07:58However, the moat here is the two things underneath the data layer and the eval layer.
08:06Um the first is the data engine.
08:09For most of the 22 languages, there is no label data available readily.
08:13When label data does not exist, building the data engine becomes hard.
08:19And that is something we have done extensively.
08:22We've built pipelines to create synthetic data and doc data from real documents and also helped build the uh pipeline for continuously improving the data that goes into training based on the eval performance and so on.
08:39So we are currently actively looking at the paradigm of RLMs um as uh uh we explore the agentic vision capabilities for our models in the upcoming releases.
08:49The second mode is the eval.
08:52You cannot reach state-of-the-art if uh um you cannot measure how well your model performs.
08:59We have curated extensive evals to ensure what we are measuring is truly sota and also truly meaningful in terms of usefulness for end uh uh users.
09:10So I want to take just a minute to dwell on our RL pipeline uh stage four because uh a lot of gain comes from RL.
09:20Essentially in OCR correctness is a machine readable problem right a lot of tests can be set up to uh um reward and score the models against uh uh these
09:33uh samples that you create.
09:35And in the world of deterministic OCR, these are all machine uh uh verifiable.
09:41Hence, RL gives us a major boost.
09:45Sample a group, score with unit tests, reinforce what the average uh uh value to beat is, and then repeat the process.
09:54And this makes RLVR for OCR very very scalable.
09:58Now after all the from scratch training we've done and the data effort we've made we were able to establish sot in two of uh the global English benchmarks one is
10:10CRB bench and the other is omnidoc bench at launch we were at 84.3 on CR and then 93.2 into on omnidok bench.
10:19Um the models uh that have released since have moved the needle quite a bit and we'll soon have a a stronger model in the global leaderboard as well.
10:31Secondly and more importantly um on 22 Indian languages we have an unbeaten lead even compared to all the frontier models such as Gemini and uh chat GPT and Opus and so on.
10:44And this is where we've extended the lead significantly and we remain uh uh strong compared to all of these newer models that have come in.
10:53Now server vision pars our agentic document intelligence workbench called aure where we enable human in the loop agentic digitizations uh and uh extractions and then injections for various uh downstream
11:10uh document intelligence problems.
11:12We provide confidence scores um and we uh have block level grounding and then also the ability to do uh agentic proof reading etc. Now benchmarks and soda are uh uh one thing they have their place.
11:28Today some of the largest enterprises in the world from insurance to banking to governments to historical preservation organizations are using server vision to digitize more than 35 million pages across English
11:43and uh uh 22 Indian languages.
11:46The model is available as API and on prem and then the agentic platform.
11:54So in conclusion until 4 months ago uh there was no sovereign model uh from India.
12:01Today we have server vision which was trained from scratch and has established sot um at a price point that is extremely competitive with uh all the other solutions including open and close source.
12:15Uh first we've started by uh uh solving some of the hardest problems for Indian language document intelligence.
12:22Soon uh we'll be releasing general purpose uh VLMs which are capable of much more uh uh vision capabilities and uh look forward to uh uh all of you trying out our models.
12:34Thank you.
12:36[applause]
12:39Happy to answer any questions.
12:40Yeah.
12:48Yes.
12:49Yes.
12:49Absolutely.
12:50Uh in general, the language capability of the 22 languages enhances English uh by quite a margin.
12:57And this is applicable to any low resource language, not just Indian low resource languages.
13:05Yeah.
13:27Right.
13:28Um we don't quite go in that direction with this model because this is a vision focused uh uh model wherein we are focused on extracting information or knowledge from documents.
13:38But yes, there could be uh some parallels to um helping models use general languages to speed up coding as well.
13:47But yeah, that's not the uh periphery of this uh uh work.
13:53Yeah.
14:10Right.
14:10Um so there are two things uh we uh create artificial documents synthetic documents as they are called for general post-raining uh that includes SFT and RL.
14:22Um however coming to your specific question on RL uh we don't uh you can generate uh synthetic documents there as well.
14:29However, um the best thing to do is to take real world documents complex enough that uh help you set up unit tests or different kinds of uh uh rewards, right?
14:40like say a a reward based on character error rate or a a reward based on table structure or a math equation or something uh that is pertinent to a language
14:51or like grammar rewards or so on and then help the model iteratively improve based on uh um rules that the model is able to produce at different uh u settings.
15:02Yeah.
15:03Yes.
15:21Sorry, did you say big fan of Chundra?
15:24All right.
15:24Uh, that's from a different lab.
15:26Uh, I'm also a big fan of the lab that produced Chundra.
15:29But yes, took your question uh on Indic benchmark.
15:32Yes, we will release the server indic benchmark that we have created for 22 languages and it uh spans um a a huge time period starting from 1800s until modern day and also different kinds of layouts, different kinds of documents in Indian languages.
15:50You can imagine documents with pros, poetry, literature, uh tables, finance and all of those things.
15:56We'll soon release that benchmark uh in public as well.
16:25So translation is again uh transliteration is again not directly involved here because in OCR what we care about is high fidelity extraction right uh you want even if there is a mistake in an image you want that mistake to be extracted correctly and not the model
16:41uh uh making changes at whim uh hence that is not directly applicable but yes uh what we see is a lot of data that is coming through transliteration for even OCR training now and the
16:53um jury is out on what is the quality of that data and how useful it is etc Yeah.
17:27Right.
17:28Um there are few things to consider.
17:30Um first is no other model frontier uh closed source or open model does well with complex Indian language documents right so for a size of uh for a country
17:44with 1.4 4 billion people you want to be able to address their day-to-day lives in uh uh terms of document intelligence problems because there is a lot of paperwork in India right it is a up and cominging country and it's still being digitized as we speak and it's
18:01important that representation or ability to digitize the country exists to begin with.
18:06Secondly on uh the training side what we have uh uh done in particular is to create data that can be added to a lot of training downstream.
18:19So uh we are at a very early stage in terms of AI becoming you know a regular parts of our lives and we need the data to start to be able to eventually get to a place where we can
18:32uh have personalized agents in the language you prefer in language I prefer and however I want in uh uh in some so to do all of that we need to start somewhere and the data needs to be created and if the data is of good quality then that helps the model become sota
18:48and If a model is sot then uh from uh insurance agencies to government organizations to others who care about sovereignity in AI right uh as a government body I cannot have a model deployed elsewhere and I don't know where the data is going for
19:06uh uh the transcriptions so I need to be able to control where the data is sent and when it is used how much it is used etc. So sovereignity becomes very crucial and hence this model really is now
19:18uh uh like it's been what four months since we launched and already we are we are digitizing 35 million pages.
19:24So that is that shows that the market was waiting for something sovereign in this space that can really kickstart uh the AI digitization wave in India.
19:33So yeah primarily like three aspects one on uh sovereignity the other is the model uh capability itself and the third is data uh that is required to train these models.
19:54Yeah, it is it is uh um 40% uh indic and then uh the rest is English and comprises of math and code etc.
20:08All right.
20:09Just a curious question.
20:24Um currently the uh the deployments we've made uh with this model are in different states uh uh trying to digitize regional languages and along with all the English and
20:37mixed language uh uh documents and hence we'll soon uh be able to train with the um uh feedback we get from uh the current deployment.
20:49So yeah, we have started that uh pipeline as well.
21:07[music]