data extraction project ideas

[rom1504]

Ideas:

• extract text+image+audio link+video link from pages ; could be valuable for future multimodal gpt3-like models
• extract text content+images from pages : much more similar to image+text dataset but that would already provide some more context for images
• extract embeddings+metadata from videos from torrents
• extract embeddings+metadata from audio from torrents

Each of these ideas could be fleshed out in these ways:

• build a POC of data extraction, estimate the amount of data available
• imagine in more details models that could benefit
• estimate the infra needed to do the extraction
• describe how the distribution of the dataset could be done
• check any legal issues

[rom1504]

audio ongoing

[rom1504]

New idea: extract image popularity score

• Page rank like for images
• Twitter popularity
• reddit popularity
• Instagram popularity Etc

Can be used to train appreciation models