Video yükleniyor...
Video Yüklenemedi
OCRia-1 is the best document parsing model for complex tables in PDFs: nested headers, merged cells, no cell borders, and extreme density within a single page. It tops multiple parsing benchmarks on table accuracy. 91.5% on ParseBench, 92.9% on olmOCR-bench, 84.1% on PulseBench-Tab, 83.2% on fr-bench-pdf2md. We published a... show more
16,283 görüntüleme • 6 gün önce •via X (Twitter)
31 Yorum
From personal testing, it also works super well with assembly/disassembly diagrams. Honestly, great work!
❤️
Tables with merged cells and no borders are where so many document pipelines struggle, so this is really welcome. The attention map visualization as it predicts each cell is a lovely way to teach how a model reads a table. Thanks for sharing it!
Report says a version of the weights is planned. The served one mixes NVFP4 and FP8, and NVFP4 runs natively only on Blackwell, so RTX 40 and Mac users would want BF16 or a GGUF
does it keep merged cells as merged in the output or flatten them? that's where every parser I've fed financial PDFs to falls apart
Borderless tables with merged cells are where most parsers quietly give up, so 92.9 on tables is real work. How does it hold up on old scanned typewriter tables, like 1940s archive reports?
does it construct explicit layout ASTs, or are you relying entirely on visual token attention for spatial relations?
Tendría que ver la doc superan al modelo de glm-turbo o 5.2?
Can’t wait to try it 🫡
For agents that act on parsed documents, the number that matters is how often one wrong cell flows into an action. With Indian bank statements and invoices, mixed scripts and stamps included, I'd want per-field confidence that can stop the agent, not just a page score.
83.5 overall against 92.9 on tables is a wide spread. What is the asterisk on olmOCR-bench marking in the launch post?
Merged cells on financial statements are exactly where I don't want a parser guessing
how do multi-page tables that span pages hold up? structure still intact across the break?
The attention map viz is so cool! Thanks for sharing!
91.5 on ParseBench but the interesting part is the nested-header/merged-cell case. what's the sample size there, and is ParseBench against the same font-rendered PDFs or scans?
Table parsing breaks where the visual grid and the logical grid disagree. Attach a schema that names the header hierarchy and column semantics so the parser can verify each cell against its role instead of inferring structure from spacing.
Tables are where most parsers quietly fail, so a 4B model at 92.9% on olmOCR-bench tables is a big deal. Small enough to run locally for doc pipelines is the real win.
This is exactly the gap we feel on finance docs. Filings with nested tables and merged cells still trip most models before the math even starts. Better table OCR is a free accuracy bump for any finance agent. Congrats on OCRia-1.
does anything work really well for hand-circled selections of text ?
merged cells with no borders is usually where table parsing falls apart. does it hold up on scans too or mostly born-digital pdfs?
Tables are where retrieval fails quietly. One misread merged cell and the answer comes back confident, cited, and off by a row. Showing the source page next to the number is the cheapest check a reader gets.
great for tables. the harder part in prod is checking the extracted number against the system of record before an agent acts on it. Curious if you'd have thoughts on that?
Table accuracy is half of it for us. When an agent pulls a sum insured out of a property schedule, we need to point back to the exact page and cell. Does OCRia-1 return cell coordinates with the values?
说实话，嵌套表头、合并单元格、无边框、高密度全占了还能解析得这么准，表格解析这块是真被卷出新高度了。
How does it handle tables split across page breaks?
Correct text under the wrong merged header is the failure that bites in production pipelines. Does ParseBench score header-to-cell mapping separately from text accuracy? That's the number I'd check before swapping parsers.
Tables are where a lot of enterprise RAG quietly breaks: one merged-cell error upstream becomes a confident wrong number downstream. Would love to see cell-level accuracy on numeric columns reported separately, since that's what finance and ops users query.
merged cells with no borders is where most parsers quietly shift a column over and nothing errors, the numbers are just wrong downstream. how does it do on a table that breaks across two pages
Borderless tables and merged cells are the usual failure points, so leading with table benchmarks makes sense. How does it handle a table that runs across a page break with the header only on the first page? That's where extracted rows tend to lose their column names.
Strong table numbers. Question from the retrieval side: do you expose per-cell parse confidence downstream? In my pipelines one misread merged cell quietly poisons retrieval, so a confidence flag per cell would be worth more to me than another point of accuracy.
the attention map visualization shifting across cells is the key breakthrough: 1D autoregressive decoders always collapse on borderless tables the 3 architectural reasons standard parsers fail on financial tables and how vision-grounded models fix them: 1. the 1D reading-order collapse: standard OCR engines (Tesseract, PDF text-streams) project tokens into a linear 1D sequence (left-to-right, top-to-bottom). the second a table has borderless multi-line cells or nested hierarchy, tokens from column 2 interleave into column 1. vision-language models with 2D spatial embeddings preserve geometric coordinate grids directly in cross-attention layers 2. the markdown table serialization trap: everyone outputs markdown for LLMs, but markdown tables physically cannot represent 'colspan' or 'rowspan' without duplicating or dropping cells. for complex SEC 10-K balance sheets, converting to markdown strips the relational hierarchy. parsers must emit semantic HTML ('<td colspan="3">') or structured JSON coordinate matrices, otherwise downstream RAG chunkers compute false row offsets 3. whitespace valley attention snapping: without visual gridlines, generic ViT attention heads disperse across blank whitespace. specialized document VLMs train sharp localized attention kernels that anchor to typographic baselines and whitespace gutters, keeping column boundaries rigid even at extreme token density seeing 92.9% on olmOCR-bench tables proves that spatial vision-grounding is the only real fix for document extraction
