{
  "video": {
    "id": "T72nqdC92PM",
    "title": "From Scratch to SOTA: Training a 3B State-Space Vision Model — Krishna Prasad Srinivasan, Sarvam",
    "duration": 1269,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "T72nqdC92PM",
    "title": "From Scratch to SOTA: Training a 3B State-Space Vision Model — Krishna Prasad Srinivasan, Sarvam",
    "one_liner": "Sarvam Vision is a 3B-parameter OCR/document-intelligence VLM built on a state-space (SSM) language backbone rather than a transformer. It was trained from scratch in India through a four-stage curriculum ending in RL with verifiable rewards, set state of the art on English document benchmarks at launch, and leads frontier models on 22 Indian languages.",
    "summary": "Krishna Prasad Srinivasan says Indian languages are largely missing from the machine-readable world: well under 1% of Common Crawl. The cause is not missing knowledge but documents that were never digitized. Sarvam's answer was to bet on block-level OCR: a small SSM-based VLM wrapped in a document 'harness' with a layout module and a reading module, instead of the monolithic page-level VLMs of late 2025. They chose SSMs because compute grows only linearly and memory stays constant, and pages can run to 5–10k visual tokens. The model went through four training stages (text pretraining, vision continual pretraining, OCR SFT, and RLVR). He argues the real moat is the data engine and the evals, not the recipe.",
    "key_points": [
      "The model has 3B parameters, fits on a single GPU, and is claimed to beat models 100 times larger on document AI. It covers English plus 22 official Indian languages, and data, compute and training were all done in India.",
      "Architecture: two harness modules, one for layout and one for reading, around an SSM VLM that does block-level OCR. This was a contrarian bet in late 2025. He says many OCR models released in 2026 have since converged on the same 'harness plus small model' approach.",
      "Why an SSM over a transformer: transformer compute grows with the square of sequence length (L×L interactions) and memory grows with the sequence. SSMs keep a single state updated token by token, so compute is linear and memory is constant. With 5–10k visual tokens per page, he argues the 'somewhat lossy recall' of SSMs is a justified price for block-level OCR.",
      "Four-stage curriculum: (1) text-only pretraining on 13T tokens of English, Indian-language text, math and code, about 40% Indic; (2) continual pretraining on 300M image-text pairs; (3) SFT on 100M OCR samples covering all 22 languages plus tables, equations and handwriting; (4) reinforcement learning to push past the SFT ceiling.",
      "A strong language prior comes first so the model can resolve smudged or ambiguous text, the same way people read a half-blurred word from context.",
      "RL works well for OCR because correctness is machine-verifiable. The loop is: sample a group, score it with unit tests, reinforce what beats the average, and repeat. That makes RLVR for OCR 'very very scalable'. Rewards include character error rate, table structure, math equations and grammar. The speaker says the best RL data is complex real-world documents, not synthetic ones.",
      "Results at launch: 84.3 on 'CRB bench' (as transcribed) and 93.2 on OmniDocBench, both state of the art at the time; newer models have since moved ahead and a stronger Sarvam model is coming. On 22 Indian languages it claims an unbeaten lead over Gemini, ChatGPT and Opus.",
      "Deployment: it powers Sarvam's agentic document intelligence workbench (human-in-the-loop digitization, confidence scores, block-level grounding, agentic proofreading). It is available as an API, on-prem, and through that platform. Enterprises in insurance, banking, government and historical preservation have digitized more than 35M pages in about four months since launch."
    ],
    "takeaways": [
      "For long, token-heavy inputs like document pages, consider an SSM backbone and a block-level harness instead of a monolithic page-level transformer VLM. Linear compute and constant memory can beat a model 100× larger on cost.",
      "Build the language prior before the model ever sees a pixel: a strong text-only backbone is what lets an OCR model fix blurred or ambiguous glyphs.",
      "Where correctness can be checked by a machine, add an RLVR stage with unit-test rewards (CER, table structure, equations, grammar) on complex real-world documents to get past the SFT ceiling.",
      "Invest in the data engine (synthetic plus real-document pipelines, retraining driven by eval results) and in curated, meaningful evals. The speaker calls these the moat, not the standard training recipe.",
      "In OCR, prioritize high-fidelity extraction: reproduce mistakes that are in the source image rather than letting the model 'correct' them. Be cautious with transliteration-derived training data, whose quality is still unproven."
    ],
    "topics": [
      "ocr",
      "document-ai",
      "state-space-models",
      "vision-language-models",
      "reinforcement-learning",
      "low-resource-languages",
      "sovereign-ai",
      "evals"
    ],
    "tools": [
      "Sarvam",
      "Sarvam Vision",
      "Common Crawl",
      "Qwen",
      "Gemma",
      "Gemini",
      "ChatGPT",
      "Opus",
      "OmniDocBench"
    ],
    "quotes": [
      {
        "text": "we focused on building the competence of the language model first uh before the model ever even sees a pixel.",
        "at": "06:49",
        "url": "https://www.youtube.com/watch?v=T72nqdC92PM&t=409s"
      },
      {
        "text": "the moat here is the two things underneath the data layer and the eval layer.",
        "at": "07:58",
        "url": "https://www.youtube.com/watch?v=T72nqdC92PM&t=478s"
      },
      {
        "text": "You cannot reach state-of-the-art if uh um you cannot measure how well your model performs.",
        "at": "08:51",
        "url": "https://www.youtube.com/watch?v=T72nqdC92PM&t=531s"
      },
      {
        "text": "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",
        "at": "16:37",
        "url": "https://www.youtube.com/watch?v=T72nqdC92PM&t=997s"
      }
    ],
    "words": 3077
  },
  "summary_url": "/#T72nqdC92PM",
  "transcript": {
    "html": "/transcripts/T72nqdC92PM.html",
    "txt": "/transcripts/T72nqdC92PM.txt",
    "vtt": "/transcripts/T72nqdC92PM.vtt"
  }
}