Krishna Prasad Srinivasan of Sarvam presents a 3B-parameter vision language model for Indic document AI, built on a state space model backbone to avoid quadratic attention over long page token sequences (up to ~10k visual tokens per page). He claims it beats document AI models ~100x larger, citing a training pipeline of 13 trillion text tokens across English, 22 Indian languages, math, and code, followed by continual pretraining on 300M image-text pairs, SFT on 100M OCR samples, and RL with verifiable rewards (character error rate, table structure, grammar). The talk covers a block-level OCR architecture with layout and reading order harnesses, data engine and eval infrastructure for low-resource languages, and production use digitizing 35 million pages for insurers, banks, and governments. Q&A addresses low-resource transfer, synthetic data, and sovereignty. # Channel Identity Observations ## Intended Audience - **Technical/ML practitioner and research community**: the talk assumes familiarity with transformers, SSMs, training pipelines, and RLVR (e.g., quadratic vs. linear compute trade-offs discussed in depth). - **Conference-style Q&A audience**: questions about transliteration data, coding-language parallels, synthetic data for RL, and sovereign deployments indicate engaged developers/researchers. - **Enterprise/government decision-adjacent listeners**: emphasis on "sovereignty," on-prem availability, and digitization at scale ("insurance to banking to governments to historical preservation organizations") signals an audience including potential adopters. ## Claimed Scope - **Sovereign Indian AI identity**: "We are Sarvam, a sovereign foundation model company in India" — self-described national scope. - **Multimodal foundation models**: "voice, text and vision" — speech-to-text, TTS, 30B/100B text models, and the featured document intelligence model. - **Indic document AI focus**: claims SOTA on "22 official Indian languages" plus English, framing Indic document intelligence as "one of the hardest document intelligence problems in the world right now." ## Self-Description Evidence - Speaker identity: Krishna Prasad Srinivasan, "general manager at Sarvam." - Model framing: "India's first sovereign vision language model built from scratch," 3B parameters, state-space architecture. - Claims of leadership: SOTA on "OlmoCRBench" and "OmniDocBench" at launch (84.3 and 93.2), and "an unbeaten lead" on Indian languages vs. frontier models (Gemini, ChatGPT, Opus). - Deployment claim: "more than 35 million pages" digitized; product ecosystem named "Sarvam Vision" feeding the "Aure" agentic workbench. - Forward scope: "Soon we'll be releasing general purpose VLMs" and a planned public "Sarvam Indic benchmark." ## Demonstrated Presentation Patterns - Technical deep-dive lecture format (curriculum training stages, architecture rationale) combined with product positioning. - Quantified claims throughout: token counts ("13 trillion tokens"), dataset sizes, benchmark scores, page volumes. - National/mission framing layered over technical content: "India is largely missing from the machine readable world." ## Notes - All SOTA, scale, and deployment figures are creator claims from the transcript; no independent verification is present in the source material. - One speaker slip: transcript renders company name as "Servin/Server" at points; consistent references indicate **Sarvam**. # Delivery Observation — Krishna Prasad Srinivasan, "From Scratch to SOTA" ## Overall Impression A structured, talk-track-style technical presentation: claims are introduced with "why" questions, answered with a concrete answer, and bookended by benchmark numbers and adoption figures. The Q&A portion is noticeably more conversational, with frequent restarts and hesitation markers. ## Recurring Linguistic Patterns ### Rhetorical question → answer structure The talk is organized around self-posed questions, often beginning with "Now why..." or "So how...": - "Now why is that? Uh primarily not because there is a lack of data..." - "Why is Indic document intelligence hard? Uh firstly..." - "So how do you actually train this? We built a staged curriculum..." - "Who are we? Uh we are serv..." - "So in conclusion until 4 months ago..." This is the most stable structural habit in the talk — at least five distinct instances. ### Enumerated lists ("firstly / secondly / third") - "Firstly the goal is that of knowledge extraction not just purely text extraction. And secondly..." - "Firstly the language model used is not a standard transformer. Um and the entire model..." - "So yeah primarily like three aspects one on sovereignity the other is the model uh capability itself and the third is data" - "So there are two things uh we uh create artificial documents synthetic documents..." ### Repeated key vocabulary High-frequency content words across the talk: - **"from scratch"** — appears in intro ("how we went from scratch to SOTA"), product description ("built from scratch"), and conclusion ("trained from scratch") - **"sovereign / sovereignity"** — "a sovereign foundation model company", "India's first sovereign vision language model", "sovereignity becomes very crucial" - **"digitize / digitization"** — "the data has never been digitized", "35 million pages", "the AI digitization wave in India" - **"state space / SSM"** vs **"transformer"** — a recurring contrast pair throughout the technical section - **"block level OCR"** — introduced early and repeated as the differentiating bet ("took a contrarian bet to focus on block level OCR") ### Filler and restart habits - Heavy use of **"uh"** throughout, denser in the Q&A than the prepared talk - Word restarts, especially around proper nouns and technical terms: "uh we are serv uh a sovereign foundation model company", "served by..." style stumbles on "Sarvam" (rendered "serv"/"server" in transcript) - Self-correction mid-sentence: "It really depends on the uh uh trade-offs here uh for long documents..." ## Teaching Habits ### Concrete numbers as anchors - "beats models 100 times larger in size" - "well under 1% of the common crawl" - "13 trillion tokens", "300 million image text pairs", "100 million OCR samples" - "84.3 on CR and then 93.2... on omnidoc bench" - "35 million pages", "1.4 4 billion people", "40% uh indic" ### Intuitive analogies for technical concepts - "the same way you can read a half blurred word because you know which word is supposed to be at the right place" - "A transformer has every token look at every other token" vs. "SSMs have a single state... updating it token by token" ### Positioning via contrast with competitors/frontier labs - "India is still missing uh in terms of data from frontier models" - "most OCR models today... are all transformers based. We took an alternate approach" - "an unbeaten lead even compared to all the frontier models such as Gemini and chat GPT and Opus" - "the moat here is the two things underneath the data layer and the eval layer" ### Humble-anticipation framing of results - "the models 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" - "we'll soon release that benchmark uh in public as well" ## Q&A Communication Style - Answers open with acknowledgment tokens: "Yes. Absolutely.", "Right. Um..." - Structured even in Q&A: "So yeah primarily like three aspects one on sovereignity..." - Occasional light humor: "Sorry, did you say big fan of Chundra? ... I'm also a big fan of the lab that produced Chundra." - Willing to scope limits of the work: "that's not the uh periphery of this uh work"; "the um jury is out on what is the quality of that data" ## Stable vs. Isolated Patterns **Stable (multiple instances):** - Rhetorical question framing (5+ instances across the talk) - Enumeration habit (4+ instances) - Number-anchored claims (dozens of instances) - "from scratch" / "sovereign" / "digitization" as refrains - Transformer-vs-SSM contrast (setup, architecture section, Q&A) **Isolated / uncertain:** - The humorous exchange ("big fan of Chundra") — single instance, insufficient evidence of a habitual humor style - The analogy about reading half-blurred words — one clear analogy; whether analogy use is a habit can't be confirmed from one example - Confidence-hedging ("jury is out") — appears once; may be question-specific rather than a pattern # From Scratch to SOTA: Training a 3B State-Space Vision Model — Observation Record **Speaker:** Krishna Prasad Srinivasan, identified in the transcript as a "general manager at Sarvam" (rendered "Servin"/"serv" in the transcript; title metadata says Sarvam). **Format:** Conference talk followed by Q&A. --- ## 1. Claimed Model Overview - **Sarvam Vision** — described as "India's first sovereign vision language model built from scratch." - **3 billion parameters**, small enough to run on a single GPU; claimed state-of-the-art in document AI, "beats models 100 times larger." - Uses a **state-space model (SSM)** rather than a transformer. - Data, compute, and training all done **in India**, for **English and 22 official Indian languages**. - Development began in **late 2025**; launched roughly **four months prior** to the talk. ## 2. Problem Framing (Stated Reasoning) - India is "largely missing from the machine readable world": a "published language distribution study" shows **well under 1%** of Common Crawl has Indian-language representation. - Not a lack of data or knowledge, but that "the data has never been digitized." - Why Indic document intelligence is hard (three reasons given): 1. Goal is **knowledge extraction, not just text extraction** — text without logical coherence is meaningless. 2. Indic scripts involve **complex fused Unicode clusters**; the model must be strong across all 22 languages. 3. Most Indian languages are **low resource** with scarce training data. ## 3. Architecture Positioning - **Contrarian bet:** while page-level OCR monolithic VLMs dominated at project start, Sarvam chose **block-level OCR plus a document harness**. Claims that 2026 model releases have converged onto the same "harness + small model" paradigm, which he cites as validation. - Two harness modules: **layout** and **reading-order** harnesses, plus the SSM VLM for block-level OCR. - **Why SSM over transformer (stated trade-off):** - Transformers: every token attends to every other token — compute grows with the square of sequence length; memory grows with sequence length. - SSMs: single recurrent state, updated token by token — **linear compute, constant memory**. - Long documents can yield **5–10,000 visual tokens per page**, making quadratic cost expensive; he accepts "somewhat of a lossy recall" with SSMs as justified to avoid transformer compute costs. - Contrasts against transformer-based OCR models such as **Qwen and Gemma** (open-source VLMs). ## 4. Training Recipe (Four-Stage Curriculum) 1. **Text-only pre-training** — 13 trillion tokens (English, Indian languages, math, code); builds the 3B language backbone *before the model sees any pixel*. Claimed rationale: strong language prior lets the model resolve smudged or ambiguous text, "the same way you can read a half blurred word." 2. **Continual pre-training** — 300 million image–text pairs; teaches general vision capabilities. 3. **Supervised fine-tuning** — 100 million OCR samples, diverse across all 22 languages plus English; includes tables, equations, handwriting. 4. **Reinforcement learning** — "helps push past the ceiling" of SFT. > "The moat here is the two things underneath the data layer and the eval layer" — the **data engine** (synthetic + real document pipelines, continuously improved via eval feedback, since labeled data doesn't exist for most of the 22 languages) and **curated evals**. ## 5. RL Pipeline (RLVR for OCR) - OCR correctness is framed as a **machine-verifiable** problem, making RL highly scalable. - Method described: "Sample a group, score with unit tests, reinforce what the average value to beat is, and then repeat." - Actively exploring **RLMs** for "agentic vision capabilities" in upcoming releases. ## 6. Benchmark and Deployment Claims - **SOTA claimed at launch:** 84.3 on **OCR-Bench**, 93.2 on **OmniDocBench**. He qualifies that models released since have "moved the needle quite a bit" and a stronger model is coming. - **22 Indian languages:** claims an "unbeaten lead" versus frontier models including **Gemini, ChatGPT, and Opus**. - Powers the agentic workbench **Aure**: human-in-the-loop digitization/extraction, confidence scores, block-level grounding, agentic proofreading. - Claims **35+ million pages** being digitized across English and 22 Indian languages by enterprises (insurance, banking, governments, historical preservation). Available via API, on-prem, and the agentic platform. - Conclusion framing: until four months ago there was no sovereign model from India; Sarvam Vision is SOTA "at a price point that is extremely competitive" with open and closed solutions. ## 7. Q&A — Recorded Positions - **Multilingual benefit:** Training across 22 languages "enhances English by quite a margin," applicable to any low-resource language. - **Coding applications:** Out of scope for this vision-focused model, but he concedes "there could be some parallels" for using general languages to speed up coding. - **Synthetic vs. real documents in RL:** Synthetic documents are used for general post-training (SFT and RL), but for RL the best approach is **real-world complex documents** to set up unit-test rewards (character error rate, table structure, math equations, grammar rewards). - **Indic benchmark release:** Committed to publicly releasing the Sarvam Indic benchmark — 22 languages, spanning the **1800s to modern day**, with prose, poetry, literature, tables, finance. - **Transliteration:** Not directly applicable — OCR requires **high-fidelity extraction** (a mistake in the image should be extracted as-is, not corrected "at whim"). Notes transliteration data is increasingly used in OCR training, but "the jury is out" on its quality. - **Why this matters (three reasons stated):** (1) no frontier open or closed model handles complex Indian-language documents well, and India's 1.4 billion people need document digitization; (2) early-stage **data creation** enables future personalized agents in preferred languages; (3) **sovereignty** — governments cannot deploy models where data flows are unknown. - **Training data composition:** ~**40% Indic**, rest English plus math and code. - **Deployment feedback:** Deployments across Indian states digitizing regional-language and mixed-language documents; a retraining pipeline using deployment feedback has started. ## 8. Unresolved Tensions / Qualifications Preserved - SOTA claims are **time-limited**: launch figures have since been surpassed by other models on English benchmarks. - SSM choice explicitly involves a trade-off — accepted **"lossy recall"** versus transformer fidelity. - Quality of transliteration-derived training data is explicitly left **undetermined**. - Some names in the transcript are inconsistent (e.g., "Servin"/"serv"/"serverom" for Sarvam; "CRB bench" vs. OCR-Bench); recorded as transcript evidence, not corrected. # Video Attribution & Provenance Record **Video ID:** `01a15ca7-325f-7169-8311-ecc159afda6f` **Title:** From Scratch to SOTA: Training a 3B State-Space Vision Model — Krishna Prasad Srinivasan, Sarvam --- ## Speakers & Roles | Speaker | Role | Basis | |---|---|---| | **Krishna (Krishna Prasad Srinivasan, per title)** | Presenter; self-identified **General Manager at Sarvam** | Transcript: "I'm Krishna... a general manager at Servin" | | **Unnamed audience members** | Questioners in Q&A | Marked as `>>` in transcript; not named | - The presentation is a **first-person talk** by a single speaker, followed by an **audience Q&A session** where the presenter answers questions. - No separate narrator, interviewer, or additional named speakers appear. --- ## Company / Affiliation Context (per transcript) - Speaker represents **Sarvam**, described as "a sovereign foundation model company in India." - Work spans modalities: voice (speech-to-text, text-to-speech), text (30B and 100B parameter models), and vision (document intelligence model — the talk's subject). --- ## Promotional / Marketing Content (Flagged) The talk contains substantial **first-party promotional claims** and should be treated as vendor marketing: - Claims of **state-of-the-art** performance on **OCRBench (84.3)** and **OmniDocBench (93.2)** at launch. - Claims of an "unbeaten lead" on 22 Indian languages vs. "frontier models such as Gemini and chat GPT and Opus." - Claims of enterprise adoption: "digitizing more than 35 million pages." - Product promotion for **Sarvam Vision** and the **Aure** agentic document workbench; model availability "as API and on prem." - Announcements of upcoming releases (general-purpose VLMs, public Sarvam Indic benchmark, stronger leaderboard model "soon"). - **Third-party comparisons** (Qwen, Gemma, Gemini, ChatGPT, Opus, "Chundra" from a different lab) are asserted without supplied verification data. --- ## Referenced Sources & Claims (Provenance Inventory) ### Cited sources - A **"published language distribution study"** — cited for the claim that well under 1% of Common Crawl contains Indian-language representation. *No specific citation/title given in transcript.* ### Technical claims made by speaker (first-party) - 3B parameter state-space (SSM) vision-language model; not transformer-based. - Four-stage curriculum: text-only pre-training (13T tokens), continual pre-training (300M image-text pairs), SFT (100M OCR samples), RL (RLVR-style with unit-test scoring). - SSM rationale: linear compute, constant memory vs. quadratic transformer cost for long documents (5–10k visual tokens/page). - Data mix in pre-training: ~40% Indic, remainder English, math, code. - "Block-level OCR + document harness" approach predates 2026 convergence on same paradigm. ### Data/benchmark provenance - **Sarvam Indic benchmark** (22 languages, 1800s–modern, prose/poetry/tables/finance): created by the company; promised for future public release — **not yet available per transcript**. - Frontier benchmarks (OCRBench, OmniDocBench) used for SOTA claims; speaker acknowledges post-launch models "have moved the needle." --- ## Availability & Access Facts - Model available as **API**, **on-prem**, and via the **Aure agentic platform**. - Sarvam Indic benchmark: **not yet publicly released** at time of talk. - No mention of open weights, dataset releases, or code availability. --- ## Uncertainty Notes - Speaker name "Krishna Prasad Srinivasan" comes from the **video title**; transcript self-identifies only as "Krishna." - Transcript contains transcription artifacts ("Servin," "server vision," "Quen," "Chundra") — normalized here as Sarvam, Qwen, Chandrasekhar-lab reference ("Chundra," lab unnamed). - Benchmark figures, adoption numbers, and SOTA claims are **speaker assertions**; no third-party verification present in source. - Audience questioners are unnamed and partially unintelligible in transcript. All right. Um, hi everyone. Good afternoon. I'm Krishna. Uh, and I'm a general manager at Servin. Um 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 models 100 times larger in size. It is also a slightly unusual model in two ways. Firstly u the language model used is not a standard transformer. Um 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 uh uh one of the hardest document intelligence problems uh in the world right now. Um this is how we went from scratch to SOTA. Who are we? Uh we are serv uh a sovereign foundation model company uh in India. We work across different modalities voice, text and vision. Uh in voice we have speech to text and texttospech models. In 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. Um, India is largely missing from the machine readable world. Um, 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. Now 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 um India is still missing uh in terms of data from frontier models that uh um are being trained every day here. Now why is that? Uh primarily not because there is uh a lack of data or the knowledge does not exist. It is because the data has never been digitized and we are solving that problem at server. Um why is Indic document intelligence hard? Uh firstly the goal is that of knowledge extraction not just purely text extraction. Uh extracting pure text without logical coherence is meaningless. And secondly in index scripts the shape of uh the words and the language you see are very different from what a machine sees. that 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. Third, 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. Our answer to this is server vision India's first sovereign vision language model built from scratch. um a three billion parameter model with state space architecture data compute and training all of which done in India. When we started the initial work in late 2025 most VLMs uh in the OCR space were monolith VLMs. uh 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. Um 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. Serverom 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. Again um today uh why do we use state space and not a transformer? Um most OCR models today uh like general VLMs, open source VLMs like Quen, Gemma etc. are all transformers based. We took an alternate approach uh using an SSM. Why? Both are fundamentally sequenced models transformers and SSMs but with very different underlying mechanics. A transformer has every token look at every other token where the compute grows with square of the sequence length. Uh that is the L cross L interactions and the memory goes as the sequence grows as well. Now on the other hand, SSMs have a single state. They maintain a single state through the sequence updating it token by token. Now as compute grows only linearly, the memory stays constant for SSMS. Now why is this the right uh architecture for OCR? It 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 and quadratic complexity of you know uh compute and memory becomes very expensive for inference. On 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. So how do you actually train this? We 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. This 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. the same way you can read a half blurred word because you know which word is supposed to uh u be at the right place. So we focused on building the competence of the language model first uh before the model ever even sees a pixel. Stage two is where we performed continual pre-training uh on 300 million image text pairs. This teaches the language model general vision capabilities and teaches how to see how to interpret pixels and so on. Then comes stage three where we performed uh supervised fine-tuning on 100 uh uh million uh OCR uh samples. Now the um stage three is primarily focused on letting a general VM model become strong at OCR. What 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. Now and stage four is re reinforcement learning which helps push past the ceiling of whatever uh supervised fine-tuning is able to achieve. Here you will see this is a standard recipe. However, the moat here is the two things underneath the data layer and the eval layer. Um the first is the data engine. For most of the 22 languages, there is no label data available readily. When label data does not exist, building the data engine becomes hard. And that is something we have done extensively. We'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. So 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. The second mode is the eval. You cannot reach state-of-the-art if uh um you cannot measure how well your model performs. We 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. So 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. Essentially 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 uh samples that you create. And in the world of deterministic OCR, these are all machine uh uh verifiable. Hence, RL gives us a major boost. Sample a group, score with unit tests, reinforce what the average uh uh value to beat is, and then repeat the process. And this makes RLVR for OCR very very scalable. Now 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 CRB bench and the other is omnidoc bench at launch we were at 84.3 on CR and then 93.2 into on omnidok bench. Um 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. Secondly 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. And 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. Now 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 uh document intelligence problems. We 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. Today 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 and uh uh 22 Indian languages. The model is available as API and on prem and then the agentic platform. So in conclusion until 4 months ago uh there was no sovereign model uh from India. Today 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. Uh first we've started by uh uh solving some of the hardest problems for Indian language document intelligence. Soon 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. Thank you. Happy to answer any questions. Yeah. Yes. \>\> Yes. Absolutely. Uh in general, the language capability of the 22 languages enhances English uh by quite a margin. And this is applicable to any low resource language, not just Indian low resource languages. \>\> Yeah. Right. Um 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. But yes, there could be uh some parallels to um helping models use general languages to speed up coding as well. But yeah, that's not the uh periphery of this uh uh work. Yeah. Right. Um 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. Um however coming to your specific question on RL uh we don't uh you can generate uh synthetic documents there as well. However, 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? like 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 or 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. Yeah. Yes. Sorry, did you say big fan of Chundra? \>\> All right. Uh, that's from a different lab. Uh, I'm also a big fan of the lab that produced Chundra. But yes, took your question uh on Indic benchmark. Yes, 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. You can imagine documents with pros, poetry, literature, uh tables, finance and all of those things. We'll soon release that benchmark uh in public as well. So 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 uh 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 um jury is out on what is the quality of that data and how useful it is etc Yeah. Right. Um there are few things to consider. Um 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 with 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 important that representation or ability to digitize the country exists to begin with. Secondly 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. So 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 uh 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 and 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 uh 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 uh uh like it's been what four months since we launched and already we are we are digitizing 35 million pages. So 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. So 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. Yeah, it is it is uh um 40% uh indic and then uh the rest is English and comprises of math and code etc. \>\> All right. Just a curious question. Um 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 mixed 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. So yeah, we have started that uh pipeline as well. \>\> \[music\]