Datasets:

Modalities:
Image
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
Dask
License:

Concerns About Dataset Quality

#11
by miroslavjaros - opened

Hello,

I am quite new to AI, but I believe this could be an issue with the training data. As you can see in the picture, I took one image from the dataset, and it states that the dominant colors are blue and green. Is there something I am missing? Look at the picture; it's obviously not true.

Thanks for your explanation.
Miroslav

Screenshot from 2024-05-19 12-37-40.png

Hi @miroslavjaros
That is a great observation.
The underlying problem you are pointing at here is the limitations of the code instruct model.
More precisely: to create these webpages, we prompted an LLM to generate "ideas" of website (the 3rd column in the screenshot), and use these ideas to prompt a code llm to generate HTML code (the 2nd column in the screenshot) which renders into the screenshot-ed webpage (1st column).
This is a failure from the code llm to follow the instructions given in the "idea".

Oh, I was wondering why the dataset was so poor. Any plans on a v3 which utilizes actual html?

HuggingFaceM4 org

We are currently not planning to do a v3. This might change if there is significant interest from the community!

Sign up or log in to comment