Add ParseBench evaluation results
#4
by boyang-runllama - opened
This PR ensures your model shows up at https://huggingface.co/datasets/llamaindex/ParseBench.
This is based on the new evaluation results feature: https://huggingface.co/docs/hub/eval-results.
Note: this includes per-dimension performance across all 5 ParseBench dimensions (text_content, text_formatting, layout, chart, table) along with the overall mean score.
Worth a caveat next to these numbers. On clean, structured pages it scores well, which is most of what ParseBench measures. On faint or low quality scans it hallucinates when it can't read the text it invents plausible content instead of leaving it blank.
In a side by side which I performed, it was the LEAST faithful of several OCR models -
olmOCR-2 > Qianfan-OCR > PaddleOCR-VL > HunyuanOCR > Unlimited-OCR
can you share your training data