Jev 与文本生成模型的取舍
Jev 不生成自由文本。Diogo 用一段对比视频说明并行决策模型与逐字生成模型在任务形态上的差别。

帖子指出,即使 OCR 结果看起来干净,仍可能含有不易察觉的错误;MiniCPM-OCR Jev 会将输出与原始图像对照并评分准确率,帮助决定哪些结果需要进一步检查。
A clean OCR result can still contain mistakes you won’t notice. What caught my attention here is MiniCPM-OCR Jev checking the output against the original image and scoring its accuracy. That gives you a way to decide what needs a closer look.