Hugging Face's logo

Datasets:
Grozkal
/
PaperSeek-OpenAlex-Embeddings

Tasks:
Question Answering
Feature Extraction
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
10M - 100M
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files
xet
Community
3

Code / Minor inconsistencies

#2
by phi1 - opened
Discussion
phi1

Hi @Grozkal ,

thanks a lot for creating and publishing this dataset!

While troubleshooting some issues with trying to produce compatible embeddings, I was surprised to find 3 different embeddings for the 11 records which all have title = 'Abstract withdrawn' and abstract = 'Abstract' - which I would expect to be identical (based on the prompt described in the ReadMe: Title: {title}\n[SEP] Abstract: {abstract}). The differences are pretty minor, so might be explained by rounding errors, e.g. resulting from using different dtypes.

However, trying to reproduce the embeddings for the oddball 'Abstract withdrawn'/'Abstract' entries using the prompt
Title: Abstract withdrawn\n[SEP] Abstract: Abstract, I get wildly different results (also when using a variant with literal '{/}' included and/or without [SEP]).

Any hints? Is the code used for generating this data set available anywhere (and/or maybe even open source)?

Cheers,
Pascal
Grozkal
 Owner

I am glad you found a use of the dataset.

The full project that used this dataset is available on github and opensource https://github.com/Mohammadsaknini/PaperSeek/. The data was generated using the scripts that can be found in Folder.

Regarding the oddball this is actually expected, given that we do not include abstracts that are contain less than 300 characters because they often contain irrelevant content.
phi1

Thanks a lot for your reply and for making the code available as well! (Side note: Surprised I hadn't found your GitHub project before, I tried both Google and GitHub search again out of curiosity, but once again without success...)

Regarding the oddball this is actually expected, given that we do not include abstracts that are contain less than 300 characters because they often contain irrelevant content.

I'm confused: How come the dataset still includes the mentioned oddball entries then (along with the varying embeddings as described)? (Skimming the code, I don't see any corresponding filter criterion which would remove them, but I might have overlooked it.)

Sign up or log in to comment