Add overall OlmOCRBench results

#31
by nielsr HF Staff - opened

OlmOCRBench was recently updated to display "Overall" results by default, this PR ensures your model shows its score on the leaderboard.

It will show up here: https://huggingface.co/datasets/allenai/olmOCR-bench.

LightOn AI org

Hi Niels,
Thanks for adding this!
There is one important distinction for the "Overall" results: since the "Headers & Footers" category rewards ignoring/not outputting visible text, we choose to exclude it from the Overall average, in fact, in the RLVR setup we try to minimize the H&F score so that the model does full page transcription, including page headers, footers and page numbers.
I think we should have an Overall without this metric since it's a bit misleading from first sight.

Ok, thanks for clarifying. Note that the evaluation feature includes a "notes" field, where you can specify additional information. Have updated this PR to reflect that.

For now I'd use the "notes", I'll discuss with AllenAI to potentially create a separate leaderboard/task for it.

Btw, would you be up for helping us add GLM-OCR to the leaderboard as well? Happy to set up a Slack channel with you

LightOn AI org

Great!
Happy to help for benching GLM-OCR, I have been willing to do so, just didn't have time before.

staghado changed pull request status to merged
LightOn AI org

@nielsr
I tried benchmarking GLM-OCR on OlmoOCR-bench today. It proved quite challenging : GLM-OCR is a two-stage pipeline (layout analysis + region recognition) rather than an end-to-end model. There are no official standalone inference scripts; the intended workflow relies on their SDK which integrates PP-DocLayoutV3 for layout detection and routes each region to the appropriate task prompt (text, formula, or table).

As a first pass, I ran the model directly with just the "Text Recognition:" prompt on all images using this script as reference for vLLM inference. Here are the results:

Category Score
headers_footers 92.3%
long_tiny_text 87.6%
arxiv_math 80.4%
multi_column 79.9%
old_scans_math 74.9%
table_tests 42.5%
old_scans 39.9%
Overall 71.1% +-1.1
Overall (wo h/f) 67.5%

For better results, we will need to include the layout detector but since they don't provide it in a standalone model it's kind of a hassle to use.

Ok, would it be easier to just use their API? I.e. :

from zai import ZaiClient

# Initialize client
client = ZaiClient(api_key="your-api-key")

image_url = "https://cdn.bigmodel.cn/static/logo/introduction.png"

# Call layout parsing API
response = client.layout_parsing.create(
    model="glm-ocr",
    file=image_url
)

# Output result
print(response)
LightOn AI org

yeah it might be best to just use the API directly, I do not have a GLM API key though.

We'd be happy to give you one. Is there any email address I can reach you on to invite you to Slack? Couldn't find it on your X/Github profiles

LightOn AI org

shared via dm on twitter/x.

Sign up or log in to comment