跳到正文
OpenBMB· @OpenBMB · X·· 11 小时前AI 评分51

MiniCPM-OCR 1.0 对比 GLM-OCR 评测:倾斜扫描更准但整体 0–3 落败

AI 导读

博主 OCR Cup 第 9 场在 16 GB M4 上对比 OpenBMB 1.3B 的 MiniCPM-OCR 1.0 与卫冕的 GLM-OCR,结果 GLM-OCR 3–0 保留头衔:单页 6.04s 对 10.5s,批处理 0.30 对 0.14 页/s,类扫描页字符错误率 0.15% 对 0.29%,发票行 48/48 对 42/48。MiniCPM-OCR 在倾斜 20°(0.18% 对 5.23%)和旋转 90°(0.38% 对 4.53%)的歪斜页上更准,但会丢失下划线,扫描页 57 个 __names__ 只保住 41 个。OpenBMB 回应感谢对比并称会继续改进精确文本准确率。测试用两个 GGUF(1.7 GB)加 llama.cpp,离线 72 秒下载完成。

正文 · 原文

Thanks for the detailed comparison! 💙 Great to see MiniCPM-OCR performing well on tilted and rotated scans. We appreciate the feedback on exact-text accuracy too. Lots to keep improving! 🙌

引用Model Test Lab@ModelTestLab

OCR Cup, match 9: MiniCPM-OCR 1.0 (OpenBMB, 1.3B) vs the champion GLM-OCR, on a 16 GB M4.

GLM-OCR keeps the title, 3–0.
One page at a time: 10.5 s vs 6.04 s. 330 pages, 10 at once: 0.14 vs 0.30 pages/s.
Scan-like pages: 0.29% vs 0.15% character errors.
Invoice rows exactly right, as returned: 42/48 vs 48/48.

What MiniCPM-OCR does better: crooked pages. Tilted 20°: 0.18% errors vs 5.23%. Rotated 90°: 0.38% vs 4.53%.

Bug of the day: Python names lose their underscores. *__slots__* came back as "* __slots _ *". On the scan pages 41 of 57 __names__ survived; GLM-OCR kept all 57.

So you should: use MiniCPM-OCR for crooked scans of plain prose, and GLM-OCR for anything where a name or a digit has to be exact.

Setup: two GGUFs (1.7 GB) and a llama.cpp binary, no Python, 72 s to download, offline. @OpenBMB

在 X 查看被引用的帖子

来源:OpenBMB · x.com