SoMark 开源 1.3B 文档解析模型 OXR-1.0
SoMark open-sourced OXR-1.0, a 1.3B-parameter document parsing model that turns PDFs and page images into Markdown or JSON,
SoMark 开源了 1.3B 参数的文档解析模型 OXR-1.0，可将 PDF 和页面图像转成 Markdown 或 JSON，权重已发布在 Hugging Face，轻量测试可在单张 24GB GPU 上运行。该模型在 olmOCR-Bench、OmniDocBench 和 DocParsingBench 上排名第一，在单张 NVIDIA RTX PRO 6000 GPU 上解码速度 558 tokens/s，端到端解析 100 页约 101 秒。其训练后阶段采用递归自我改进（RSI）循环，由通用 VLM 判定解析结果并生成新标注页面，无需人工标注。
推荐理由
AI 评分 65
原文给出了 OXR-1.0 的架构拆分方式与 RSI 训练循环，读者可据此判断文档解析模型的工程取舍。
主题
正文
SoMark open-sourced OXR-1.0, a 1.3B-parameter document parsing model that turns PDFs and page images into Markdown or JSON,
> It ranks first on olmOCR-Bench, OmniDocBench and DocParsingBench while being the fastest of 21 models tested.
> runs on a single 24GB GPU for light testing
Its post-training uses a recursive self-improvement (RSI) loop that turns the model’s own parsing mistakes into new training data, with no human labeling, a first in document parsing.
Parsing a page means finding every block (title, paragraph, table, formula), reading each one, and putting them in human reading order.
Most parsers tie layout detection to recognition or to reading order, which makes layout hard to retune for a new kind of document.
OXR splits it out: a small layout model draws a box around each block, and the 1.3B OXR model reads every box and sets the reading order, so a team can swap in its own layout model and keep the rest.
It outputs 13 element types with page coordinates, including tables, math as LaTeX (the text format scientists use for equations), code blocks, stamps and chemical structures.
In the RSI loop, OXR parses each training page several times, and a general vision language model (VLM) marks the result correct, wrong, or uncertain (meaning OXR’s own runs disagreed).
The VLM then writes a reason for each failure, and those reasons are grouped into clusters of the same weakness.
For each cluster, the VLM generates new labeled pages styled after a different real document each time, and OXR trains on them before the next round.
> On a single NVIDIA RTX PRO 6000 GPU, it decodes 558 tokens per second and parses 100 pages end to end in about 101 seconds, which is 3.05x GLM-OCR’s page throughput and 33.5% ahead of the next-fastest model, OvisOCR2.
> The weights are on Hugging Face, it runs on a single 24GB GPU for light testing, and a free API works without an account or key.
The accuracy lead is thin on olmOCR-Bench (0.24 points) and OmniDocBench (0.27 points), but OXR gets there while parsing pages about 64x faster than Infinity-Parser2-Pro, the olmOCR-Bench runner-up.
The RSI loop works as automatic hunting for the model’s hardest cases, with one general VLM acting as both judge and label writer, so the loop can only fix errors that VLM can spot.