Datasets:
MiniCPM-OCR results and leaderboard submission process
Hi Real5-OmniDocBench team,
We recently evaluated the MiniCPM-OCR 1.0 SDK document parsing pipeline on all five Real5 scenarios, with 1,355 images per scenario. Our preliminary Overall results are:
| Scenario | Overall |
|---|---|
| Illumination | 93.72 |
| Scanning | 94.43 |
| Screen-Photography | 93.64 |
| Skew | 93.24 |
| Warping | 93.39 |
| Five-scenario average | 93.68 |
The pipeline uses MiniCPM-OCR 1.0 with PP-DocLayoutV3 for layout detection:
- GitHub: https://github.com/OpenBMB/MiniCPM-OCR
- Model: https://huggingface.co/openbmb/MiniCPM-OCR-1.0
We would like to confirm the evaluation protocol and ask about submitting the result to the official leaderboard:
- Do you accept results evaluated by model teams, or do you conduct an independent evaluation?
- What materials would you need, such as prediction Markdown files, scoring logs, evaluator configuration, and reproduction commands?
- What is the process for inclusion in the dataset README and the Space leaderboard?
We would be happy to prepare the predictions and reproduction materials as needed. Thank you!
Hi MiniCPM-OCR team,@Longkun111
Thank you very much for sharing your evaluation results and for your interest in Real5-OmniDocBench! It is great to see the performance of MiniCPM-OCR 1.0 combined with PP-DocLayoutV3 across the five physical distortion scenarios (averaging 93.68%).
Regarding your questions about the evaluation protocol and leaderboard inclusion, here are the details:
Evaluation & Verification:
To ensure the rigor and fairness of the benchmark, we conduct independent verification for leaderboard submissions. We will be verifying your model's accuracy on our end in the near future.
Required Materials:
To help us expedite the independent verification and reproduction process, we would appreciate it if you could share the prediction Markdown files, scoring logs, and any specific configuration files or reproduction scripts you used. You can share them via a public link (e.g., GitHub Gist, Google Drive, or Hugging Face repository).
Inclusion Process:
Once our team independently cross-checks and confirms the reproduction results based on your materials, we will be more than happy to update the official leaderboard on our Hugging Face Space and include your results in the dataset repository README.
Thank you again for your contribution and support of Real5-OmniDocBench. We look forward to reviewing your materials and completing the verification soon!
Best regards,
Real5-OmniDocBench Team
Hi @ChengCui ,
Thank you for clarifying the verification and leaderboard inclusion process.
We have uploaded our evaluation materials here:
Download MiniCPM-OCR Real5 evaluation materials
The archive includes:
- Prediction Markdown files for all five scenarios: 1,355 per scenario, 6,775 in total.
- Per-scenario scoring logs and metric summaries.
- Inference configuration and reproduction scripts.
We look forward to your independent verification and would be happy to help investigate any differences during reproduction.
Thank you!