Does correcting in the online environment help train the model?

[gitAlwin]

Hello there,

I am using the online interface of MinerU (after pressing the 'Try Online' button on the main page) and I am wondering if the corrections I make there are used to train the model further. Does it help if I make corrections there?

I'm curious because It would be an easy way for me to give back.

Thank you in advance,

Alwin

[dosubot[bot]]

Hi @gitAlwin! I'm Dosu and I’m helping the MinerU team.

Corrections you make in the online MinerU interface are not currently collected or used to train or improve the model. The online demo is set up for interactive use only and does not send your corrections or feedback back to the developers or into a training pipeline (source).

If you're interested in contributing, the most impactful way is by preparing annotated datasets or joining community discussions about model improvement. MinerU's core model can be fine-tuned with custom data, but this requires manual preparation and isn't connected to the online demo corrections (source).

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[AndyBodnar]

Great question! Based on how most ML services work, the online demo is typically running inference on a frozen model. Your corrections in the UI are likely just for your current session.

However, there are ways your feedback could help:

1. If they have a feedback mechanism - Look for a report issue or submit correction button

2. Contributing to training data - The best way to help is:

   • Report issues on GitHub with example PDFs
   • Contribute to their dataset if they have a public collection effort
   • Open issues describing failure cases

To confirm, I would ask directly in their Discord/GitHub or check if the online demo has any explicit submit feedback feature. Some projects do collect corrections for fine-tuning but it requires explicit opt-in for privacy reasons.

Appreciate you wanting to give back - thats the open source spirit!