Datasets:

Modalities:
Image
Text
Formats:
parquet
Languages:
English
ArXiv:
Tags:
License:

How to do preprocess for obelics

#8
by Awiny - opened

Thanks for this great dataset!!!
To preprocess interlevel data, open-flamingo input first 256 tokens of each document for MMC4 data.
Could you please share some details of how to preprocess obelics document. For example, how to select the 1024 tokens from each document.

Thanks :)

Hi @Awiny
That's a great question.

We pack examples (a single web document) into samples of sequence length N tokens (let's say N = 1024 for instance).
How do we do that packing?
If a given web document is is shorter than N, then we just pack it.
If that document is longer than N tokens, we will extract (by sampling) a subsequence of N tokens of that long web document. That extraction/sampling is not uniform: we bias the starting index (choosing a starting point is equivalent to choosing the subsequence) so that this subsequence contains text+images. If we don't do that, with our data, there is a risk that the subsequence is text only which might not be the best use of the compute for multimodal adaptation.

HugoLaurencon changed discussion status to closed

Sign up or log in to comment