Frequently occurring bug: Missing the letter i and l in words

[LaguePesikin]

Description of the bug | 错误描述

pdf paragraph screenshot:

MineU output:

{
        "type": "text",
        "text": "For the string elicitation task, we are given a target response $y$ , and our goal is to generate a prefx $x$ such that m’s output on $x$ is an exact match for $y$ ; in other words, we want to solve arg $\\mathrm{max}_{x}$ $p_{m}(y\\mid x)$ . ",
        "page_idx": 1
}

(prefix -> prefx)
Similar problems occur in words like 'fluency' (fuency), 'benefit' (beneft)..., and almost always happen.

How to reproduce the bug | 如何复现

see above.

Operating system | 操作系统

MacOS

Python version | Python 版本

3.11

Software version | 软件版本 (magic-pdf --version)

1.0.x

Device mode | 设备模式

cpu

[myhloli]

Can you try this pdf on huggingface demo https://huggingface.co/spaces/opendatalab/MinerU and feedback result?

[LaguePesikin]

Can you try this pdf on huggingface demo https://huggingface.co/spaces/opendatalab/MinerU and feedback result?

Seems it work well, all pairs fi and fl are accurately recognized.
So why the web demo behave differently from my local client?

[myhloli]

The online demo uses the dev branch, which includes some unreleased fix codes, so the specific performance may differ from the release version.