Does MonkeyOCR provide confidence scores for OCR recognition results?

[wuaman]

I’m currently integrating MonkeyOCR into a downstream pipeline where the reliability of OCR outputs is critical. To filter out low-quality predictions, I would like to use a confidence score (or similar reliability metric) for each recognized text segment.
However, I couldn’t find any documentation or examples indicating whether MonkeyOCR currently outputs confidence scores along with the text predictions.
Could you please clarify:

1. Does MonkeyOCR currently support outputting confidence scores for OCR results?
2. If not, is this feature planned in future releases?
3. If it's already computed internally, is there a way to expose it (e.g., via a flag or modifying the output structure)?

If needed, I’m happy to contribute or test any experimental features related to this.
Looking forward to your response — thanks again for the great work!

[zenosai]

Hello, since MonkeyOCR uses an MLLM for recognition, it cannot provide a confidence score.

[wuaman]

Hello, since MonkeyOCR uses an LLM for recognition, it cannot provide a confidence score.

thanks for your reply, I want to design a set of criteria to evaluate the usability of text blocks or formula blocks in a layout by extracting the logprob information of each output token. Do you think this approach is feasible?

[zenosai]

I think it's feasible